235 comments

[ 3.1 ms ] story [ 257 ms ] thread
Model Cores/Threads Boost10 / Base Frequency Total Cache PCIe TDP SEP (USD)

AMD Ryzen 9 7950X 16C/32T Up to 5.7 / 4.5 GHZ 80MB Gen 5 170W $699

AMD Ryzen 9 7900X 12C/24T Up to 5.6 / 4.7 GHZ 76MB Gen 5 170W $549

AMD Ryzen 7 7700X 8C/16T Up to 5.4 / 4.5 GHZ 40MB Gen 5 105W $399

AMD Ryzen 5 7600X 6C/12T Up to 5.3 / 4.7 GHZ 38MB Gen 5 105W $299

Interesting that TDP increased from 65W on Ryzen 5 5600X to 105W on Ryzen 5 7600X, even though 7600X uses a newer process.
Partially due to PCIe5 and DDR5.
The entire chip is probably just one giant fully utilized vector unit by now
Connect all the data points:

* Zen 4 is more efficient (up to 74% more efficient at 65W)

* Zen 4 will be used for desktop, server and mobile chips

* Ryzen 7000 is clocked a good bit higher than Ryzen 5000

* On the desktop, higher maximum clock speeds for single and multi-core workloads are desired

* The higher TDP is to allow for higher clock speeds for single and multi-core workloads on the desktop

> Zen 4 is more efficient (up to 74% more efficient at 65W)

So the TDP doesn't scale linearly?

TDP isn't something that scales.

> TDP stands for Thermal Design Power, in watts, and refers to the power consumption under the maximum theoretical load. [0]

It's basically a ceiling.

[0] https://www.intel.com/content/www/us/en/support/articles/000...

Yeah very poorly phrased on my part, sorry.

Since you mentioned "up to 74% more efficient at 65W" what I meant is if the perf per watt (and hence the temp) doesn't scale linearly with the amount of work.

The 7700X has about double the TDP of the 3700X but about a 50% increase in perf.

That is correct. I think it's always been the case that as clock speeds increase, power draw and heat generation accelerate (diminishing returns).

When you're comparing 7700X to 3700X, are you talking about performance at the same TDP limits, the same clock speed, or maximum performance?

I'm comparing max perf and TDP.
Not when it comes to clock speed.

Note that the total power a CPU draws is proportional to its capacitance (fixed) * its frequency * the square of its voltage + some other stuff that's generally the same across CPUs of the same generation.

But that doesn't quite tell the whole story, either; as you increase the frequency you also have to feed more voltage to the CPU to keep it stable for signal-processing reasons (a higher voltage means more difference between 0 and 1, and when you're right on the edge of it not working that will make the difference between a successful compute and a crash).

As such, when you're trying to make a CPU and want to chase clocks, you get bit twice: first by the increased frequency due to physics, and twice by the (square of the) increased voltage requirements for stability. So running at 6 GHz and requiring 1.2V to get there is significantly more expensive power-wise than running that same processor at 4 GHz and only needing 0.8V.

The only real way to fix that is microarchitecture, of course; the trick with Intel in particular is that they've been milking the same architecture they came up with in 2007 and aside from adding some Atom cores to its latest models have done basically nothing but shrink the die so it's natural that they're falling behind now.

AMD could cap the 7600X at 65 watts, would that make you happier?

Perf/watt has substantially improved, as have capability (like double the memory speed, double the memory channels, and double the PCIe bandwidth).

Sure if power and cooling are available it clocks up higher, but you don't have to. Most BIOS will let you control the clock speed if you want.

> AMD could cap the 7600X at 65 watts, would that make you happier?

Even more, you can cap the 7600X at 65W. All AMD cpus have configurable TDP, the advertised number is the default. Going above is overclocking, but dropping it is within spec. If you do it, your cpu will just use less power under load and be a bit slower.

Indeed, the nice thing is power use grows quickly with clock, so even something like 5GHz vs 5.5GHz can allow a chip to run a fair bit cooler, and since there's various bottlenecks like memory latency that the performance difference will be less than 10%.
It seems we are stuck at 16C/32T for desktop CPUs for a bit longer. I guess most things that scale beyond run even better on GPUs. But not all! I wonder if we will get non-"Pro" Threadrippers this generation?
tbf, a lot of things still haven't been optimized for when AMD introduced >4 cores with first gen Ryzen. I'm running a 5950x and unless I'm doing something like 3D rendering I rarely notice more than ~8 of my cores ramping up at a time.
Sounds like you need to install Gentoo. Excellent replacement for space heaters in winter time.
Unfortunately most packages spend more time during the configure step or other ~single threaded parts than actually compiling stuff in parallel when you have that many cores. You can compile multiple packages in parallel but doing that without running out of memory when you do get to a couple of big packages is tricky.

Ideally, portage would take memory (including ramdisk) requirements of packages into account when scheduling builds (some packages do check for free memory) and ask as a make jobserver [0] (with patches for non-make build systems to support jobservers) to always fully use the available resources.

[0] https://www.gnu.org/software/make/manual/html_node/Job-Slots...

There's really no point in going beyond 16 core CPUs. Games only use 4 cores, maybe 6 cores max. Which means for games, you want to maximize performance-per-core and don't care about having more than 8 cores. More than 8 cores is only useful for very few use cases where the task is very parallelizable or you care about power consumption efficiency and some of those cores are efficiency cores.
I make use of all the cores on my Threadripper regularly for multi-process parallelism. Large compiles regularly saturate all my cores. Development against multiple Kubernetes clusters also makes it beneficial to have a ton of cores.

So there is definitely a use case, it’s just maybe not for you

I didn't say there were no use cases, I said there were

> very few use cases

The average person is not compiling programs or dealing with Kubernetes.

The average person is not buying a 7950 so I have no idea what you are getting at.
I get the confusion, I meant the average person buying these CPUs is not compiling programs or dealing with Kubernetes.
I don't know. Typically the 5950/7950 is not targeted towards gamers. It sacrifices single core performance for more cores. Usually it's more of a prosumer audience that actually do need those attributes.
I've been seeing these CPUs marketed as next gen gaming CPUs but maybe I'm wrong
Yes the 5900/7900 but those have 12 cores. The higher end 5950/7950, which has a higher core count, actually perform slightly worse in games and are targeted towards a prosumer.

Don't get me wrong, they still work amazingly well for games. It's just that they sacrificed some performance in those chips for the higher core count so it only makes sense they'd want to push that further.

But my guess is they don't want to compete with their threadripper line. Unfortunately those come much later in the product cycle so we'll probably be waiting some time for a Zen 4 threadripper.

7950X has more L3$ per CCX and higher single threaded clock, so I'm not sure why you think the 7900X would perform better in games.
Maybe I wasn't clear. 24 or 32 core CPUs are only useful for the select few who are running a lot of intensive and parallelizable computations. The majority of people buying these mainstream CPUs do not get any more use out of a 24 core CPU than a 16 core CPU and yet would be paying hundreds of dollars more. It would only make sense to have these kinds of CPUs for server-grade or otherwise specialized tasks which do not fall under the mainstream performance line of CPUs as seen and talked about here. That is why I think AMD will go no further. There is no benefit, only cost to the kind of mainstream consumers who are buying these CPUs.

> aren’t even the actual use cases for a high core workstation CPU.

I was talking about 16 core CPUs in general which also includes laptop CPUs. Another reason why 16 cores makes sense for a general consumer CPU is because of the power efficiency gains with efficiency cores in laptops.

Your Threadripper also has a lot more DRAM bandwidth than can be economically provided on a mainstream motherboard. 16 cores is a reasonable limit for AMD's mainstream platform, and going beyond that would require more expensive motherboards and a bigger, more power-hungry IO die—exactly the wrong direction for them to be moving with this product segment.
Umm Threadripper CPUs have historically been on their own sockets/mobo platforms. They aren’t drop-in replacements for the mainstream CPUs. The grandparent comment was wondering if we’d get Threadrippers in this generation, which has nothing to do with “this product segment,” whatever that means
Raptor Lake is rumoured to go to 24C/32T. 16 of 24 the cores are efficiency cores though, we'll see how that compares to the all performance core design of Ryzen 7000.
There is little doubt that Intel Raptor Lake will be better than AMD Raphael at single-threaded tasks, because both will have about the same clock frequency, but Raptor Lake will have a higher IPC.

There is also little doubt that AMD Raphael will be better than Intel Raptor Lake at multi-threaded tasks, even if they now have the same number of threads. The reason is that AMD is made with the much more efficient TSMC process and the consequence is that at equal power consumption the AMD CPUs will reach much higher clock frequencies.

This is the essential difference between single-threaded and multi-threaded tasks, during the former the performance is limited by the maximum turbo clock frequency, while during the latter the performance is limited by the clock frequency at a fixed power consumption, which depends more on the manufacturing process than on the microarchitecture of the CPU.

I wish they had an 8 core option that maximized the performance-per-core possible. I don't have much use for a 16 core CPU.
I'm confused. Why wouldn't the 7700X 8-core CPU work for what you're suggesting?
Lower boost clock speed than the 7950X
The 7700X is one 8-core CCD, which won't clock all cores as high as two 8-core CCDs with half the cores disabled.

Personally I think the trade-off in cross-CCD computation is worse than the slightly lower all-core clock for the kind of workloads that benefit most from 16 threads, but I could see the benefits of higher all-core clocks for desktop use cases like gaming.

There will almost certainly be a 7800X3D before spring, that will probably be pretty close to what you want.
Soon I will relocate my PC outdoors, power consumption has skyrocketed over the last couple years and its now both CPUs and GPUs increasing YoY.

The era of smaller form factors and increased efficiency is over for now.

>> Soon I will relocate my PC outdoors, power consumption has skyrocketed over the last couple years and its now both CPUs and GPUs increasing YoY.

I don't understand. Why will you relocate your PC outdoors? To help dissipate heat somehow?

Because it makes his room too hot I guess.
(comment deleted)
They still sell low powered CPUs, you know?
This is just obviously untrue. You can run these chips with more power, but they automatically scale, so if you limit them, they'll still outclass current low TDP options. You might pay more for the privilege and leave potential performance on the table, but these are actively more efficient than past options, not less, you just can drive them for more performance rather than more efficiency if you want.

Of course, having higher performance can actually be more efficient in some workloads, as there are cases where running for a short time at higher power draw and then reverting to a very low power idle state is better than running for longer periods of time at more limited power draw to do the same work.

They explicitly list a 25% performance-per-watt gain. They push more power to get even more performance, but you can choose if you want to do that. They claim at 65w TDP Zen 4 is 74% more performant than Zen 3.

> The switch to 5 nm gives "Zen 4" 62 percent lower power for the same performance,

I know this is a desktop launch and all but I wonder how much does this update catches upto Apple M1/M2s. With 6000U series CPUs the gap was already closer - https://www.youtube.com/watch?v=YOSQIUGGdYE but it looks like this will get it pretty close. Obviously getting Laptops on new 7000 series will take time. AMD's achilles heel has always been availability when it comes to Laptops.Many paper launched SKUs of 6000 series Laptops are still not available in US market. Heck, you will be hard pressed to find a 99W laptop with AMD only CPU (basically Apple's flagship 16inch configuration).

For Zen 4 a good part of the power reduction is that the i/o die went from the Global Foundries 12nm process to TSMC's 6nm process. That probably accounts at least partly for the large drop in power for the same performance, making it unlikely that Apple will see the same drop.
Yeah the M1/M2 chips are amazing. In this one test the 5950X is about 50% slower than the M1Max so perhaps the 7950X will close the gap even more: https://forum.blackmagicdesign.com/viewtopic.php?f=21&t=1498...
Encoding is usually done by hardware accelerators (located on the GPU) and is generally unrelated to the performance of a processor. Additionally in the video they're comparing hardware accelerated video encoding to software encoding; it's honestly surprising it's only half as fast. Here's a comparison showing the rtx3090 ahead of the Mac Studio in encoding: https://youtu.be/f-mbtSRz2nk?t=550
Sure but the 3090 is going to use a ton more power and release a lot more heat than say a M1Max studio.

Then again, at the same time, the 3090 machine can do CUDA and other useful things.

> I wonder how much does this update catches upto Apple M1/M2s.

>The greatest performance gains are actually at the lowest TDPs, where the 7950X saw a 74% increase in Cinebench R23 MT performance. These advantages actually decreased as TDPs went up, dropping to 37% at 105W, and finally 35% at 170W.

When AMD isn’t loading CPU clockspeeds into the stratosphere – which is always well into the diminishing returns of the voltage/frequency curve – Zen 4 is significantly more power efficient than its predecessor.

https://www.anandtech.com/show/17552/amd-details-ryzen-7000-...

The problem is that Intel is reviving their Pentium 4 strategy of performance via ridiculous frequencies and voltages, and AMD is following suit.

Contrast to Apple, where the strategy is throwing die space at more execution units and running the clocks at the efficient end of the voltage frequency curve.

The M2 clocks up to 3.5 Ghz, while AMD is targeting up to 5.7 Ghz.

> The greatest performance gains are actually at the lowest TDPs, where the 7950X saw a 74% increase in Cinebench R23 MT performance. These advantages actually decreased as TDPs went up, dropping to 37% at 105W, and finally 35% at 170W.

My money is on the new I/O die being the source of most of the improvement in the low TDP range. That thing was pretty power hungry on 12nm GloFo; 6nm TSMC is a pretty sizeable leap.

Obviously there are improvements everywhere in the V/F curve, implying that the CCDs have also had solid improvements, but since I/O die power draw doesn't change much with CCD frequency, it's a bigger slice of the pie (so to speak) at lower clocks.

> The problem is that Intel is reviving their Pentium 4 strategy of performance via ridiculous frequencies and voltages, and AMD is following suit.

> Contrast to Apple, where the strategy is throwing die space at more execution units and running the clocks at the efficient end of the voltage frequency curve.

That's the benefit of having a single customer. Apple knows they're never going to build mobile devices with good thermals, so their chips will never clock to the moon, so they have freedom to make a much wider core that (almost certainly) has larger clock domains and would be difficult to run at higher clocks.

Otoh, AMD has many customers and those customers like high clockspeed, and their competitor will give it to them, so AMD needs to have a design that can clock to the moon, but also be power efficient at lower clocks.

> Apple knows they're never going to build mobile devices with good thermals

They seem to be doing just fine with thermals.

>Like the 2020 M1 MacBook Pro, this laptop doesn’t get overly warm. Its underside reached a maximum temperature of 85 degrees, which is ten degrees lower than what we consider to be uncomfortably hot for a laptop. Likewise, the touchpad never went above 79 degrees.

This is one of, if not, the coolest-running and quietest laptops I’ve ever used.

https://www.tomsguide.com/reviews/macbook-pro-13-inch-m2-202...

Contrast to the Dell XPS 13 Plus:

>the XPS 13 Plus' fan was really struggling here because, boy oh boy, did this thing get hot. After a few hours of regular use (which, in my case, is a dozen or so Chrome tabs with Slack running over top), this laptop was boiling. I was getting uncomfortable keeping my hands on the palm rests and typing on the keyboard. Putting it on my lap was off the table.

https://www.theverge.com/23284276/dell-xps-13-plus-intel-202...

For those who want a laptop, chasing performance via ever increasing clock speeds and voltages is problematic.

The point is that they are doing well with the limited thermal stress their CPUs put off, very intentionally.
Apple's family of chips by large focuses on performance per watt because it's a primarily mobile chip - made for laptops. The Mac Mini/Studio is more of a side effect than primary target.

If they will have something special for Mac Pro, it would resemble AMD's strategy.

> AMD is following suit

I don't think it's a "problem" per se.

TDP is a ceiling, "power consumption under the maximum theoretical load", and it allows for higher clocks, yes, and especially across more cores.

Unlike Apple, AMD and Intel have to compete with each other and care about how well games perform. So single core clock speed matters more. So while they work towards efficiency on the one end, they still try to push upwards on thermal headroom and clock speeds.

It would be a problem if Intel and AMD only chased high clock speeds, but ignored everything else (especially efficiency) entirely. But Zen 4 is a terrible example of AMD ignoring efficiency. (+13% IPC, +74% performance at 65W compared to Zen 3.)

The M2 is on Low Power Node, AMD is on High Frequency Node along with IOD.

So in theory AMD Zen 4 APU on the same Low Power N5 should be even more efficient.

But if we do Geekbench 5 with linear scaling, my guess would be the Zen 4 still only gets about 1400 point with 3.5Ghz in the best case scenario. Which is still not bad.

A lot of what Apple has shown to the world with A13 to A15's design will be coming to Zen 5 and its iteration Zen 6. I suspect only then will we see AMD catching up.

Or Nuvia / Qualcomm and Apple still has a few tricks to further increase Pref / Watt.

Where does the article say "The switch to 5 nm gives "Zen 4" 62 percent lower power for the same performance,"?

Was the article linked changed?

I think I'll stick with the 5800X3D. I was expecting these new CPUs having the same cache tech but apparently not.
I’m willing to bet that’s their ace up their sleeve for Intels 13th Gen. Wait till around Jan and AMD will pull the rug out from under Intel.
It's in their roadmap as the next release with distinct SKUs.
Fairly sure AMD has already indicated that an "extra cache" model(s?) of Zen 4 consumer chips will appear down the road.
v-cache hbm is coming to the 7000 line, but they will stagger the launch for MBA reasons.
v-cache and HBM are two entirely distinct technologies, there are currently no products that use both and there are not even any rumors of any products that would use both.

v-cache is AMD:s branding for attaching additional SRAM cache directly on top of the die using TSMC hybrid bonding, HBM is a JEDEC standard for DRAM. They have literally nothing to do with each other.

That makes v-cache that much more impressive. Do you know if the GPU cores will be able to read directly from v-cache?
In Raphael (the cpus that were just launched) the GPU is a tiny, weak one that's sitting on the IO die, while the v-cache chip is integrated on top of the main CPU die. I would suspect that it can't make any use of it, not that it would benefit much because it just doesn't have the compute power. It's really only intended to make the CPUs able to do basic desktop work without a separate GPU.

AMD currently has no published big APUs that would use v-cache, but many enthusiasts, including me, have noted that if they made one and allowed the GPU to share the last-level cache, it would potentially be a very compelling product for mobile. Who knows when/if they make one.

They do have upcoming discrete GPUs which are rumored to optionally contain v-cache (RDNA3, probably released as Radeon RX 7000).

How good is the 5800X3D!?! I've recently upgraded from a 3900XT which was permanently clocked to 4.4ghz. Whilst certain titles such as Star Citizen, DCS World and Flight Simulator are having staggering improvements, I'm also finding it very solid for desktop / programming use - e.g. Visual Studio.

The reduced power consumption doesn't hurt either; I needed constant ~1.27vcore for the 4.4ghz before.

So far sounds pretty good. I wonder how good will the 7950X be for workstations.
does anyone have any idea why amd is not making a cpu like the 3700X anymore?

the 65W TDP for the perf was something to be seen

Because Intel hasn't been able to keep the edge on architecture for a long time, so they started a race to the bottom, abusing the TDP.

Unfortunately, the market seems not to care much, so I "can't blame them" (well, I do blame them for the effect).

This is even more perplexing considering the laptop market, where power efficiency is crucial, and Intel should be dead now (Alder Lake hasn't been as efficient as hoped, and on Linux, the matter is even worse due to suboptimal support), yet, it's still the most common platform.

They do[0].

It was released much later in the lifecycle of Zen 3 than the 4 "headline SKUs" that were launched here as well. I think we can expect the same for Zen 4.

[0] https://www.amd.com/en/products/cpu/amd-ryzen-7-5700x

The 7600x and 7700x have already been announced, and they're rated 105W.

The 5x00 lineup has a few models with lower clock (5500, 5600, 5700), but they consume as much as their higher clocked counterparts (5600x, 5700x).

The hope is for a 7500 with lower TDP, which is a possibility.

You're right, I misspoke ! For Zen 3 they announced the 5800X and released the 5700X later (with the same TDP as the GP's 3700X), whereas for Zen 4 they announced the 7700X and the 7800X will probably be released later.

Indeed every CPU announced here has a higher TDP than the previous generation, which isn't the best. Hopefully lower TDP SKUs come out as you mention.

Id assume that the 7700X is basically the 1:1 successor for the 5800X and the reserved the 7800X name for the X3D equipped version that presumably comes later. They might release a 7700 without the X.
In the first generation the 1600X was a 95w TDP part, with the 1600 being 65w. They shuffle around where the lines are, the numbers aren't always 1:1.

They listed Zen 4 at 65w TDP being 74% more efficient than Zen 3 in their slides, which implies they do want to sell a part at that spec at some point.

If your concern is purely noise/heat and so on, and not budget, then you can just buy one of these chips and apply restrictions to it, given modern chips auto-scale performance. Can feel like a waste, of course, but should result in a very efficient setup.

TDP is not a measure power consumption across the range, but only at "maximum theoretical load"[0].

If you missed the presentation, look for in-depth coverage[1]. These chips are more efficient than Zen 3, consuming less power for the same performance. The higher TDP allows for higher multi-core clocks and much higher performance, but that's not the same thing as power consumption (across all levels of performance).

Sure, if you can use the power at the high end, you can consume more power, but you'll get a lot more performance out of it.

[0] https://www.intel.com/content/www/us/en/support/articles/000...

> TDP stands for Thermal Design Power, in watts, and refers to the power consumption under the maximum theoretical load.

[1] https://www.anandtech.com/show/17552/amd-details-ryzen-7000-...

> If you missed the presentation, look for in-depth coverage[1]. These chips are more efficient than Zen 3, consuming less power for the same performance. The higher TDP allows for higher multi-core clocks and much higher performance, but that's not the same thing as power consumption (across all levels of performance).

That's coverage from the slides, a.k.a. marketing. I don't doubt that this will be a significant performance/efficiency improvement, given the soft cap, but desktop is another thing. TDP has been correlated with average consumption, on Intel CPUs, and all the GPUs. It'd be surprising that this didn't apply to this specific CPU family.

I get what you're saying - if a CPU is given cart blanche, it might tend to just run at higher clock speeds and power draw because of the higher TDP ceiling. So it may come down to motherboard settings and OEM configurations.

I know from personal experience, my CPU rated for 105W TDP tends to consume 30W or less during my normal usage, though certainly more during heavy computation including gaming. Higher base clocks on the Ryzen 7590x (4.5Ghz) compared to Ryzen 5950x (3.4Ghz) could lead to overall higher power usage, if the increased efficiency at those clocks isn't enough to cover the difference.

At least there are motherboards where you can set power consumption or temperature limits.

I just built a new Zen 3 system (5600G without a discrete video card for now) and I found a place in the BIOS where I can set the max CPU temperature. It doesn't turbo boost as much then but it's fun at the least. For what I use that machine, I can drop full load power from 100 W to 70 ish by setting temp limits and only lose like 12% speed.

If you do your research you can get a board where you can set the power limit in W directly. I was in a hurry so I'll have to do with the temp cutoff.

And what fewer people know is that the power limit is often disguised as “cpu cooler type”. It’s bizarre but I guess it makes sense in a target market consumer way. Some (usually cheaper) motherboards won’t even run the processor you get at it’s stock maximums until changing the TDP (cooler type) manually.
Unlike the 3000 the new ones have an iGPU and AVX512 support, so just looking at TDP doesn't tell the full story.
This is a pattern I am glad to see fade-away: lower TDP chips. Because you can just tell a modern chip to use however much power you'd like and it'll do GREAT! This has confused a lot of people who think low power is impossible! There is some silicon lottery, there is some difference in chips - some lower power, some higher power, but it's been increasingly faint & the justification for having separate chips has mostly gone away. We don't need lower power bin chips anymore: just set the power-point lower on the chip you have! That alone is great. If you want to get fancy, you can undervolt to drop idle some, restrict your max clocks, and it'll sip even less power.

This was a fairly memory-intensive task, but Matthew Dillon of DragonflyBSD writing up the Ryzen 2700X[1] he got in 2018 keeps coming to mind:

> I Enable XFR2, then set the PPT, TDC, and EDC limits. Set TDC and EDC high enough so they don't get in the way, then limit power consumption by adjusting PPT. By using the PPT limit instead of manually setting the CPU frequency, the motherboard gets the best of both worlds... it will idle just as low as it did before, it will still run one or two cores at full speed (~4.1 GHz), and it will ratchet down the frequency when all cores are loaded. Using a PPT limit with XFR2 is far, far superior to using manual OC frequency settings for the CPU.

> With standard XFR enabled in auto mode the 2700X will pull around 180W at the wall at full load. This might be useful if I had DDR4 3000 memory in it, but I don't, so there's no real need to pull that much power at full load. I was able to reduce this all the way down to 85W at full load without really impacting a concurrent -j 32 nativekernel NO_MODULES=TRUE test compile.

I really look forward to seeing how people throttle-down & get ultra-efficient 8c 7700X's with this new generation. (Especially now that the CCX die went from 12 -> 7nm.)

[1] https://lists.dragonflybsd.org/pipermail/users/2018-Septembe...

The 7950X looks amazing and is 50 bucks cheaper than the 5950X's MSRP at launch!

This isn't even the Threadripper Zen4 line -- if the desktop line is this good the HEDT line is going to be even more amazing-er.

It looks great, although I was a bit disappointed - I was really looking forward 24 cores :)
MSRP will remain theoretical - due to supply chain disruptions I expect that average consumer will have to pay at least 50% more, if they are lucky enough to find those processors in stock.
AMD said they are not expecting to be supply constrained.
But you need new motherboard and worst of all its only DDR5 which is still not worth it for workstations.
(comment deleted)
> Support for new and evolving technologies like PCIe® Gen 5 and DDR5 memory empowers users to grow with their Socket AM5 solution, which AMD will support with platform longevity through 2025

Does it seem like sockets are getting less and less years of support? Just 3 years of "longevity" doesn't seem that long, I probably upgrade my CPU once per 3 years and that would mean if I now move to AM5, I'm gonna have to yet again get a new motherboard when AM6 gets launched in ~2026.

Intel have been doing two generations of support for absolutely ages. AMD doing four generations on AM4 was the outlier rather than the norm, and it sounds like they are giving themselves wiggle room this time to do similarly, but potentially cut it a bit shorter if they feel they need to. Given how many reviews talked about AM4 chips being a no-brainer upgrade choice if you had an existing board, they have incentive to do it. AM4 being on DDR4 and older motherboards being PCIe 3 really limited it. Given DDR5 is new and PCIe 5 is brand new, it feels like AM5 has plenty of potential to go longer, but I can understand not wanting to tie themselves too far ahead.

I think it's more a matter of target: Intel probably care most about the pre-built market where people just buy a new computer, so there isn't much value. AMD focused on taking the enthusiast crowd and getting the mindshare that way, and long socket support definitely benefited them there.

AM4 being supported for that long caused more trouble than it was worth IMO. When buying a new CPU+motherboard, you had to make sure you got a motherboard with a new BIOS, since BIOS updates were required to support new CPUs. AMD would even send out older CPU models just to help consumers update their BIOS of necessary. A lot of cheaper motherboards didn't have enough flash memory to support both the last and the first CPUs, so you had to make sure not to install the wrong BIOS update or your CPU wouldn't work anymore.

Intel's approach, which means you can always count on the BIOS supporting your CPU if the CPU physically fits, and which meant a BIOS update would never remove support for your CPU, is honestly arguably the more consumer friendly approach.

Eh, it's a trade-off for sure, but "consumer friendly" is a stretch. Intel still have the issue of requiring BIOS updates for new CPUs, and the issue with having to remove BIOS support can be solved by just telling motherboard manufacturers to spec larger for it so they have room for more.

In contrast, the upside of being able to slot a Ryzen 5000 CPU into a machine that you bought for first gen is a big deal in terms of the lifetime of those other components, and gives consumers a lot of flexibility. I get some people just won't utilise that and will buy fresh, but just because it isn't worth it for you personally doesn't mean it isn't a good feature more generally.

> and the issue with having to remove BIOS support can be solved by just telling motherboard manufacturers to spec larger for it so they have room for more.

I think I've read somewhere that the cause of that issue was that the SPI block on older AM4 CPUs can only address up to 16 megabytes of the SPI ROM chip. There's not much the motherboard manufacturers can do, when it's hardware on the CPU itself which does the reading directly from the ROM chip.

I'm not sure requiring a new motherboard vs flashing a bios is consumer friendly, unless you assume the consumer is loaded and technically illiterate. If someone paid me $150 to flash a MB, I'd probably take that deal every time.
Sure, but if I'm buying a new CPU+motherboard combo anyways, I'd prefer if that just worked and I didn't have to acquire a temporary old CPU for the purposes of flashing BIOS with the CPU I actually want to use.
Having the option > not having the option.
In isolation, yes. But that wasn't the discussion we were having, so I'm not sure why you're telling me this.
Being able to buy a new CPU and put it in your old motherboard is better than having to buy both a new CPU and a new motherboard, even if you have to spend five minutes googling which BIOS you need to flash.

You seem to have missed the point because the original discussion claimed that it was somehow "consumer friendly" to require the purchase of a motherboard as well. You've somehow morphed that into, "if I'm buying both a CPU and motherboard". The whole point was that you don't need to buy a new motherboard.

> AM4 being supported for that long caused more trouble than it was worth IMO.

Note that AM5 was deliberately designed to fix those issues. All AM5 boards are capable of doing a bios upgrade without a CPU in the socket (the spec requires a USB port connected to the chipset which has a microcontroller that does it). AM5 motherboards also have a much higher minimum BIOS flash chip size.

But the option was there.

With Intel it was cough up $$$ for a complete new system or nothing. Saying that is more consumer friendly is pretty silly - it's worse in every way.

In reality only a geek is going to upgrade their CPU, and a geek worth their salt should be able to sort out these issues.

>Intel have been doing two generations of support for absolutely ages.

It's worth noting that most of that was artificial; just enough electrical incompatibility to make you buy a new motherboard if you wanted a new processor (it's ultimately all just slightly different configurations of LGA1156 anyway- look up ASRock's P67 Transformer for more conclusive proof).

Back when they were competing against AMD (pre-2007) they only had one socket, that being LGA775 running from Prescott Pentium 4s all the way through the last Core 2 systems, and supported it for a very, very long time. Sure, it'd require your mainboard manufacturer to keep up to date with BIOS updates to support the new CPUs, but they were ultimately all compatible.

They are committing to minimum 2025. This gives them the flexibility to either continue support or make changes if new CPUs require new sockets.

Regardless it’s still better than intels 2 generation change. With AM5 you should get 4-5 generations of cou minimum.

AMDs roadmap says Zen 5 in 2024, so you’ll be more likely to get 2 or possibly 3 generations.
Hmm I thought the roadmap for zen 5 was 2023? But also we will get Zen4 3d, there will prob be Zen5 3d, and if Zen6 is AM5, that makes 4-5 generations.
In the event yesterday they "confirmed that Zen 5 was on track for 2024".
This is probably exactly what their marketing people will try to say, yes. “No no, we added some cache, new generation!” I’m waiting for them to call each chip it’s own generation, and then we can add another generation for the APUs, and now we’re probably well into the double digits.
That quote ends with "and beyond", which you cut off.
And deservedly so, since it’s just marketing fluff. Like how having 7 foo in your bar suddenly became “over 7 foo”. Take the difficult upgrade history with zen 3 and older boards, add the roadmap with Zen 5 in 2024, and what this estimate says is they’re planning on AM5 for Zen 4 and 5.
especially since mobo manufactures were not happy with this longevity either
Not really, Intel has been rolling out a new socket every 2 years (give or take) for pretty much all of the 2000s.

AM4 was very much a departure from the norm, and it’s also the case that compatibility was a pain in the ass. Whenever you wanted to upgrade you needed to read data sheets, watch reviews, wait for AMD to change their mind^, and then flash the AGESA on your board and hope.

^ https://arstechnica.com/gadgets/2022/03/amd-reverses-course-...

I know AMD had a lot of trouble with the interposer after Zen2 or so, due to being stuck on the same socket. Keeping compatibility became more and more expensive. Though Zen 1 was a much different product than the later generations (no I/O die) so maybe that was an unusually bad situation.
A good observation! That does feel pretty short. One thing to note, PCIe's regular cadence is three years (3.0 lasted way way longer.):

> The group responsible for developing and updating the PCI Express standard, the PCI-SIG, aims to update that standard roughly every three year. [1]

PCIe 6.0 is already released & uses the same transmission rate, but is PAM4 to send two bits per tick. Which I expect AM5 will be able to handle. Samsung is already making DDR6 ram. By 2026, there almost certainly will be pressure to advance the platform.

This platform feels amazing & fresh right now, & hearing it'll only have 3 years life does feel short. AM4 lasted >5 years. It'd be neat if AMD would back-release new processors on older sockets in the future, but I can also imagine the bios support for these scenarios being tangled & gross.

[1] https://arstechnica.com/gadgets/2022/06/months-after-finaliz...

I wish AMD would add more cache for faster speed and lower energy consumption

  Ryzen 5xxx L1 32/32,   L2 512,   L3 32768
  Ryzen 7xxx L1 32/32,   L2 1024,  L3 65536
  M2         L1 192/128, L2 16384, SL 8192
  M1 Ultra   L1 192/128, L2 49152, SL 98304
[edit: see comment about latency cache comparison below]
> I wish AMD would add more cache for faster speed and lower energy consumption

Ryzen 5xxxX3D: 96 MB of L3? https://www.anandtech.com/show/17337/the-amd-ryzen-7-5800x3d...

Yes, which boosted performance, and I think the X3D is not on the die, so easier to increase.
Probably also worth noting that the M1 Ultra is a 20C CPU so the cache per core ratio on 5800X3D (8C) is much higher despite similar total L3.
They did that with the 5800X3D; the results were overall positive, although uneven. The cost was significant though - around 100$ more than the 5800X (which was $350).
Where it really mattered, the results are staggering. E.g. Star Citizen, with everything depending on its rather complex entity streaming. There, the 96MB of cache which I believe is for the 5800X3D's single CCD blows all other CPUs out of the water.
They are going to, AMD confirmed that Zen4 3D is the next release after regular Zen4

Like the 5800X3D it will use 3D stacking to squeeze more cache into the same die area

(comment deleted)
Those numbers aren’t directly comparable — caches usually trade off size for performance, and Apple’s L1 sits in performance somewhere between Ryzen’s L1 and L2. Ditto for Apple’s L2 versus Ryzen’s L2 and L3.

I can’t remember the exact article I read that in, but it’s pretty obvious when looking at Anandtech’s latency charts.

M1 latency: https://www.anandtech.com/show/17024/apple-m1-max-performanc...

Ryzen latency: https://www.anandtech.com/show/16214/amd-zen-3-ryzen-deep-di...

Thanks! To anyone else looking at the graphs, beware they are logarithmic.
There are tradeoffs there. Bigger caches mean larger areas, which means longer wires and more complicated clocking. The M1 tops out a little over 3 GHz, vs. 5 GHz for Intel chips and almost 6 for these new Zen 4 parts.[1] Apple then bets that they can make it up with wider execution units, which seems to have worked out for them (it also explains why boring/unoptimized scalar logic in amateur tests aren't meeting the numbers seen from benchmarks and HPC tests).

But it's not about "cache" per se, it's a different design. You need to do a lot more than just increase cache area to duplicate Apple's design choices.

[1] A big part of that is, of course, the architecture's genesis as a phone processor. Everyone loves to talk about how efficient the M1/2 are, but the truth is the CPU is a comparatively small part of the energy budget on devices with 5-10W displays. If you were designing a priori for a laptop, you wouldn't necessarily be as power-constrained in your design as the M chips.

Raptor Lake increases L1 cache size and nearly doubles L2 cache size from Alder Lake (then again double the size of Ryzen).
> I wish AMD would add more cache for faster speed and lower energy consumption

There's a trade-off: a bigger cache means a higher latency. According to https://twitter.com/chiakokhua/status/1564413952108335105 this doubling of the L2 cache already caused an increase of 2 cycles of latency.

> Ryzen L1 32/32

> M1/M2 L1 192/128

This is a consequence of Apple using a 16kiB page size, while the x86 world is stuck with a 4kiB page size. For technical reasons (the L1 lookup runs before or in parallel with the virtual to physical address translation), the L1 cache can only be indexed by the bits which don't change between virtual and physical addresses (that is, the bits representing the offset within the page), making it hard to increase the L1 cache size while keeping its latency low (and a low latency is very important in the L1 cache).

I don't know the latency for the Apple L2 caches, but it wouldn't surprise me if it's higher than the latency for the Intel and AMD L2 caches; having a larger L1 cache would mean Apple can afford to have a higher latency to the L2.

(comment deleted)
Yikes... those are some power hungry chips. Why would a person still buy x86? Compared to ARM, it's inefficient and outdated technology. Buying computers with AMD and Intel chips is like buying a gasoline car when electric cars are the future.
Because lots of people need more flexibility or lower cost than Apple Silicon. GPU, RAM, storage configurations, etc.

Apple Silicon is great but doesn't serve every need.

Why does your beloved apple still sell Intel based systems if this is like a "gasoline engine"?

https://www.apple.com/shop/buy-mac/mac-pro

https://www.apple.com/shop/buy-mac/mac-mini

Maybe not everyone wants/needs to be locked into the apple Ecosystem?

PS. Electric cars are the past as well: https://en.wikipedia.org/wiki/History_of_the_electric_vehicl...

But seriously, why do they sell the Intel Mac Mini?

The Intel Mac Pro still exists because they haven't made a modular apple silicon machine and their remaining large scale video production customers would leave them if they were told just to to migrate to M1 Studio machines.

I don't see a similar argument for the Mini, especially as it's still on 8th gen. I'm honestly surprised Intel is still supplying Skylake CPUs for it

I have the exact same thoughts as you.. While you can sort of see the Mac pro's place, the intel mini makes zero sense?

Perhaps they have a warehouse of these somewhere and they are still trying to unload them?

I see Intel based iMac's for sale at Costco yet they are not on Apple's site?

I think you're right and that they're hoping to unload the last of the 2019s this Black November at a ridiculous discount (again).

I think I saw a decently specced 21.5" Intel iMac going for $599 last year, although an M1 Mac Mini would still be my choice.

The word out of the supply chain is that the Mini is due for an update when the 3nm M2 Pro ships.
I would if there were any decent options that weren't Macs, honestly, and if there was a standardized platform such as the PC.

The success of x86 wasn't only due to software, but also thanks to the standardization of hardware and firmware. You can release a single image and boot it on every PC-compatible computer, because stuff like the BIOS, PCI, VGA, ... are all standard. A bare x86 CPU is of little use without all the thingamajigs that make a PC a PC.

Viceversa, on ARM outside the few boards or computers with an UEFI it's a far west of options, configurations, and so on. ARM PCs will never succeed outside of Apple closed garden until they get as convenient to install and upgrade as their x86_64 counterparts are.

Because most apps still run on x86?
That’s only true for apps that are compiled and run on a PC. Most general purpose computers in the world run ARM/Android.
These are desktop processors.
They are, but cross-platform and cross-architecture software is quite common nowadays because x86 PCs are now a minority of computers. I am challenging the claim that "most apps" require x86.
For now, at least, Apple doesn't serve the upper end of the market with anything remotely affordable.

I bought an M1 Mini early on and it's been an absolutely wonderful machine. But it's got 16 GiB RAM and a few ports on the back.

On the same desk, I've got an AMD 3900X w/ 64 GiB and expansion slots and room inside the case. The point being, different needs are being met.

True for personal purchases, at least. We just bought three fully specced Studios because the team that uses them will earn the cost back within the quarter.

Apple isn't ever going to win any hyperscale awards, but they're putting on a good show for HEDT despite Threadripper and other recent high performance CPUs. They're very good at what they're good at.

And for me, anything I'd want to put Linux on these days is either embedded or lives in a rack. Desktop is a big market, but give me something cool and silent or else far away any day.

Can I build a customizable gaming PC with an off the shelf ARM CPU yet?
For me at least, an AMD/Intel desktop box plus its multi-year power bill still is likely to come out cheaper than a comparable Apple desktop or high-end laptop plus its multi-year power bill. The lower-end Apple laptops are not an option for me due to the "one external display" limitation. Competing ARM desktop hardware is not an option for me due to insufficient performance - I might as well not upgrade at all.
FWIW, I was a long-time dual monitor user who migrated down to one due to MacBook support. I still have dual monitors at the office, and find my home setup broadly comparable - I use a 34” ultrawide with my MacBook mounted on a monitor arm next to it and serving as a second monitor. (Obviously doesn’t scale to >2 monitors.)
I can't play the video games I want to play on an ARM CPU.
That's a pretty apt analogy, because there's a lot of negatives about electric cars currently.

I wouldn't buy an electric car right now, and I similarly wouldn't buy ARM desktop PC that's bundled with anal probing from Apple and unlimited limitations.

It's interesting to see this sentiment surface again, as x86 was far behind other architectures for such a long time and this argument was a basic ingredient of the platform debates in years past.

The answer of course is software compatiblity, OSes, CPU platforms, HW markets etc are not interchangeable, and x86 was and is the only practically open platform with big enough market that there's working competition between hw/system vendors, os vendors, sw vendors, etc.

Then eventually in the late 90s, x86 matched and later overtook the competitors, helped by the huge volumes leaving process investments of competitors too far behind. (Everyone had0 their own private fabs then, and process generations got expoenntially more expensive).

After the RISC camp struggled head to head with x86 for a few years the race was over in 5-ish years. Probably the 21164 and 21264 were the last chips to hands down beat x86 by a large margin. They had a Rosetta style translator back then and most (?) x86 Windows apps ran faster on the Alpha/WinNT platform than on fastest native x86 chip for a time.

It's funny, whenever they refer to "competition", they mean Intel, whereas they should worry about Apple, too, especially as far as efficiency is concerned.
Apple can’t run Windows programs and games reliably so they are not a contender.
... and not all Linux programs any more, sadly. Due to the switch to ARM I had to get a separate x86 box for some of my work (namely yocto builds and the like).
> ... and not all Linux programs any more, sadly.

Historically, it's never been a problem. One of my first distros was Yellow Dog Linux. There will be many more in the future.

Most distributions have an AArch64 build. Secondly, macOS Ventura will have Rosetta 2 for Linux as well.
Most stuff I can do natively. I am not messing with a yocto build except on their recommended platform though. Too complex to be worth the effort of trying it on something else if not needed.
... and it's slow enough even without Rosetta ...
Because apple's ~10% of market share of the desktop market is a concern?

https://www.trustedreviews.com/news/apple-takes-90-of-arm-pc...

Efficiency is critical on laptops, but not really on desktops. Electricity is incredibly cheap.

if you left a 300 watt desktop computer on 24x7 it would use ~$260 in electricity for the year.

https://energyusecalculator.com/electricity_incandescent.htm

it works out to $0.72 a day, a cup of coffee costs more

> Electricity is incredibly cheap.

Beware, this is something that might change at short notice. Starting from October I'm paying 0.28€/kWh, making the daily 7.2kWh cost 2.02€, which would indeed buy me a cardboard cup of coffee at the kiosk.

Europe?

There really isn't too many reasons why electricity should have "shock pricing" like this.

Here (Ontario) electricy prices are fairly flat over long time periods. https://www.oeb.ca/consumer-information-and-protection/elect...

probably because once the plants are built, the cost to maintain them is fairly fixed?

Off peak power is 8.2 kwh Feb 8, 2022 and was 8.7 back on Nov 1, 2016

in order for this to change, it requires government approvals so it wont be "changing at short notice".

The Russia/Ukraine conflict is causing problems with energy production throughout Europe.
Yes. My son is a nuclear physicist and found the whole "Germany shuts down its nuclear plants in favour of Russian gas" a strange choice.

They made a choice, now they see the results of the decision?

Maybe this will be a catalyst to revive nuclear?

Perhaps our shitty leaders can stop the "we hate carbon" mantra and start selling our massive natural gas supplies to Europe removing the shortages?

(comment deleted)
Your electricity is incredibly cheap. I'm paying $0.70 per kWh. your example costs me $1840 per year.
Maybe you should elect better politicians.
Maybe they did
?? They elected politicians which drove the price of electricity up and you think this is a "good thing"?
1. Politicians have limited control over the price of power. There are many extraneous factors that influence pricing as well.

2. Politicians and their constituents often have many priorities other than electricity prices.

One thing is for sure, countries with the cheapest power are definitely not countries known for their great politicians.

https://www.globalpetrolprices.com/electricity_prices/

Again, this is highly dependent on many things including how the market is structured.

If your politicians are not concerned about energy policy, you did not elect the right ones. Politicians should have the same concerns as those they represent.

Here (Ontario) we had the liberals who implemented many disastrous "green energy" programs including "fixed rate" contracts well above market rates. This resulted in substantial increases in our rates.

The price of electricity became a major election issue and when an election year came, the Liberals were decimated. They not only lost the election, but also lost party status.

Our politicians have a lot of influence over electricity rates including direct intervention:

- January 18, 2022 - Fixed Electricity Price The Ontario government has announced that electricity prices are to be set at the off-peak price of 8.2 cents per kilowatt-hour, 24 hours per day for 21 days starting January 18, 2022, until the end of day February 7, 2022, for all Regulated Price Plan customers. Read the government's news release and our FAQs.

-June 1, 2020 - Fixed Electricity Price The Government of Ontario introduced a fixed electricity price of 12.8 ¢/kWh for consumers paying time-of-use prices to support them while Ontario plans the safe and gradual re-opening of the province. Read the government’s news release and our FAQs

If your government has zero control over your markets, maybe you should ask if this is in your best interest?

Your chart proves very little. on the opposite end of the scale (expensive power) are many countries with questionable politicians as well?

The countries at the top end of the chart are mostly countries with expensive power because they have very little resources or because they are liberal democracies that prioritize environmental/health/safety/equity issues

Canada has some of the cheapest power for a free democracy, not sure why you think it is particularly worthy of criticism for high energy prices. You can thank your wealth of resources for that.

You act like one vote counts.
Wow, is that in the U.S.? We pay $0.089 which is almost 10x less in the Midwest. Either way I'll pick up solar when Biden's inflation reduction handouts come along lol.
It's a good 50% more for us in Kansas.
Netherlands. Electricity prices are linked to gas prices (EU law). Gas prices went 15x in the last year. It's becoming a issue all over Europe but sooner in countries with a more liberalized energy market (UK/NL).
Efficiency is extremely important on mobile. It is important in the datacenter. And finally it's becoming important on the desktop, too. Maybe not in the USA, but in Europe the prices went by the roof. This might have some positive consequences, with people becoming more and more aware of the fact that electricity is not practically free. I'm afraid we will face a similar story with water.
Maybe you should restructure your energy market so it is not so "profit driven" to avoid shock pricing?

Our electricity prices going back to 2006 : https://www.oeb.ca/consumer-information-and-protection/elect...

It has gone up over time, but no "shock price hikes". To change the rate, you need government approval and politicians are not fond of hiking rates and losing votes.

Basic essentials need "user fees" so there is not blatant waste, but perhaps having controls in place is a good thing?

In Ontario (and most of Canada) we regulate Natural gas, electricity and water rates to prevent "gouging". Power rates in Quebec are even lower thanks to their massive hydro electric generation which is extremely cheap and this is passed to the consumers. Ontario is mostly Nuclear, which costs more to generate vs hydro hence our higher rates.

> Maybe you should restructure your energy market so it is not so "profit driven" to avoid shock pricing?

Well, it's a long story, but in short, Germany insisted on gas from Russia for two reasons: 1) it was cheaper than from other sources, 2) politicians hoped that they can somehow civilize Putin in this way. It didn't work well, and Mrs Merkel is embarrassed by it (I don't even mention previous chancellors).

Electricity isn’t always cheap, but more importantly heat isn’t free to cool either. A hot or loud computer isn’t desirable in many situations and as we see hotter summers due to global warming, I see many more people already valuing efficiency
here is our electricity prices since 2006, show me where it was "expensive".

https://www.oeb.ca/consumer-information-and-protection/elect...

If people value efficiency, they can buy energy efficient CPU's and not the latest and greatest high-power CPU's from AMD, which is what this thread is about?

You do realize the entire world doesn’t live in Ontario right?
As great as the Apple chips are, they’re not competition for AMD. AMD sell chips, Apple doesn’t.

Apple is a package deal and as such doesn’t cut into AMD’s market for console chips, servers or gaming systems.

AMD should eat the bullet: add an on-die a gigabit ethernet controller and an HDA codec if the EM allows it. Will need AM6 socket for the new pins though.
> AMD Ryzen 7950X processor enables single-core performance improvement of up to +29%

It appears that the new micro architecture is not providing any performance increase itself since 4.5Ghz is 32% faster than the previous 3.4Ghz clock ... and AMD is measuring a 29% IPC increase.

All of the performance gains must be purely from going to 5nm (vs 7nm).

For reference, the previous chip (5950X) had a base clock of 3.4Ghz [0] and this new 7950X has a base clock of 4.5Ghz.

[0] https://www.amd.com/en/products/cpu/amd-ryzen-9-5950x

EDIT: I'm confused, why the downvotes? Why not just message below if you don't agree with something and we can have a discussion about it.

Yeah there are no major microarchitectural changes that we're aware of for these chips. Bigger L2$, support for AVX-512 (kinda ... double pumping 256bit units), possibly wider front-end but not a huge amount more; it's primarily the new process making the difference.
The support for AVX-512 is not "kinda". AVX-512 at the Ice Lake level is supported. The most widespread Intel CPU with AVX-512 support is Tiger Lake and its support is no better than that of Zen 4, it also provides one 512-bit pipeline for FMA or multiplication and two 512-bit pipelines for simple operations, the same as Zen 4.

What happens is that Zen 4 has the same execution units as Zen 3, so any program which can keep all the execution units busy is accelerated on Zen 4 only by the greater clock frequency.

However Zen 4 has a new frontend for instruction fetching and decoding and for branch prediction. Many programs will be executed more efficiently than on Zen 3, with a better utilization of the execution units, leading to the claimed IPC improvement of 13% on average.

Additionally, rewriting a program to use AVX-512 can also improve the utilization of the execution units, leading to a speed-up greater than the clock frequency ratio.

Support for a certain ISA does not imply anything about the speed of the CPU, even if sometimes the CPU vendors change in the same generation both the ISA and the microarchitecture, resulting in greater throughput.

In this case AMD has postponed the improvement of the execution units for Zen 5. Even if the support for AVX-512 does not improve the maximum possible throughput, it improves the average throughput over many programs. The same is true for most of the Intel CPUs that support AVX-512, except for the top models of server or workstation CPUs, because they have one of the 512-bit FMA units disabled, which results in the same maximum throughput as on Zen 4 or on the older CPUs, since Haswell.

You are correct in saying that ISA doesn't matter. The only difference is that it is easier to do 8 instruction parallel pre-fetcher on ARM or another fixed length instruction architecture than on x86_64, and decoding more instructions can be better for re-ordering and register renaming.
Didn't AVX-512 turned out to be a flop? e.g. https://en.wikipedia.org/wiki/AVX-512#Performance -

"On some processors AVX-512 instructions cause a frequency throttling even greater than its predecessors, causing a penalty for mixed workloads. The additional downclocking is triggered by the 512-bit width of vectors and depend on the nature of instructions being executed, and using the 128 or 256-bit part of AVX-512 (AVX-512VL) does not trigger it. As a result, gcc and clang default to prefer using the 256-bit vectors. ()"

() - https://stackoverflow.com/questions/56852812/simd-instructio...

AVX-512 was a bit of a flop initially, because of how Intel implemented it. AMD's solution doesn't provide quite as much peak throughput for highly-optimized code, but is a better way of providing the flexibility benefits of AVX-512 to the masses without the severe downclocking. There may still be plenty of situations where it would make sense to use 256-bit vectors with AVX-512 instructions, but on Zen 4 there won't be a strong reason to avoid 512-bit vectors where they are useful.
It was a flop because it was intended for a process node that Intel was delayed on for years. It had massive problems to the point of not really making sense when backported to older nodes.
I don't think it's accurate to say AVX-512 was backported. The original Skylake consumer CPUs released as the second generation of products on Intel's 14nm already had space reserved in the CPU core floorplan for the AVX-512 register file. That space didn't get used until the Skylake server CPUs shipped, still on 14nm several years later. AVX-512 support didn't arrive in the consumer desktop product line until Rocket Lake, which was backported to 14nm but was not remotely the beginning of the AVX-512 story.
The AVX-512 instruction set has never been a flop. It is much a much better instruction set than AVX.

Most AVX-512 instructions have 3 variants, with 512-bit registers, with 256-bit registers or with 128-bit registers.

When using the 256-bit or the 128-bit AVX-512 instructions, there has never been any disadvantage versus using AVX.

The only problems have been when using the 512-bit AVX-512 instructions, especially on the CPUs derived from Skylake Server, due to the way how Intel has implemented the clock frequency control.

Using the 512-bit AVX-512 instructions requires more power than when using the 256-bit AVX-512 instructions, the same as when using e.g. 4 cores instead of 2 cores. In both cases, when doubling the operation width or when doubling the number of active cores, the clock frequency is reduced.

When a program has a large proportion of 512-bit instructions, then the throughput is higher despite the lower clock frequency.

On the other hand, when a program has only a few 512-bit instructions, the execution will be slowed down for almost a second after 512-bit instructions are no longer used, until the CPU decides to power down the upper half of the 512-bit units.

All this problem is caused because the Intel CPU tries to be too smart and decides automatically when to power down the unused units.

In the similar case when using more cores, there is no problem because when the core is no longer used, the program has a halt or a MWAIT instruction which powers down immediately the core, restoring the higher clock frequency.

If Intel had provided an instruction like "end of 512-bit instructions" to power down the upper halves of the execution units immediately, there would have been no problems with the slow down caused by sporadically using a few 512-bit instructions, exactly like there is no problem when launching some extra execution threads, because the clock frequency is restored when the extra threads finish or are suspended.

Because Zen 4 has the same execution units as Zen 3, using AVX-512 on Zen 4 will not cause any kind of slow down that would not have also happened when using AVX on Zen 3.

Thank you Adrian (B) for explaining this thoroughly! I'll use your comment as future reference for me!
You need to improve your micro-architecture to hit those speeds as well. You can't get there without deep pipelines.
The uarch affects reachable clock speeds, a lot.
Base clocks have been basically meaningless for quite a while. My 3700x happily sits at 4GHz on all cores; others see their 5950x at 4.1GHz. It's basically just a number AMD came up with that they want to guarantee under all workloads, not a number that indicates any kind of performance characteristics.
Why is it meaningless? It tells you how long a clock cycle takes, and you can look up instructions per cycle for relevant calls. I admit the picture for what makes a fast CPU is highly application dependent and has a lot of variables to consider, but I don't think it's meaningless.
Meaningless because (adequately cooled) modern desktop CPUs mostly run near their boost clock in single threaded loads and somewhere between base and boost in multithreaded.
As I stated, CPUs don't actually run at the base clock rate. They idle well below and under load they boost above it. How much they boost is of course architecture, load and temperature dependent, but it's generally a fair bit higher than the base clock.
> It tells you how long a clock cycle takes

It does not tell you how long a clock cycle takes. It tells you how long a clock cycle takes when the CPU is massively loaded on all threads, in adverse thermal conditions. If you have better than average cooling, you will never see the base clock under load. If you are running just a single thread, you will always boost to Fmax.

This is not accurate. Per their presentation, the geomean of the measured IPC improvement is 13% at a fixed 4.0Ghz comparing Zen 4 to Zen 3. That combined with increased maximum clockspeeds accounts for the up to 29% single-core performance improvement.

(Base clock speeds are not relevant or used in their performance comparisons.)

Where do you see in the article information about 13% IPC improvement at fixed 4.0Ghz?

(Was the article link changed after posting. Others in this thread are referencing other things I also don't see in the article)

(comment deleted)
IPC means “instructions per cycle”, not “instructions per core”. So the numbers you quote would mean a 0% IPC increase (or slight decrease)
The numbers quoted have an implicit assumption that performance claims are based on "base clocks", which is incorrect.
You compare instructions per clock with the clockspeed.
... you seem pretty confused by words. Micro architecture is where you get IPC improvements, along with some fraction of the speed and energy efficiency improvements (it's not 100% the node, or rather, designing for a node is in feedback with microarch).
(comment deleted)
A practical reason for downvotes: having misinformation being the top comment isn't conducive for anything good.
Base clocks are largely irrelevant. These things will run way above or way above that clock speed when appropriate.

Furthermore, architecture determines what frequency can be achieved. Designers can choose to sacrifice IPC to increase frequency, or vice-versa. We could build a machine with ridiculously high IPC if you don't mind it running at 1 MHz.

Yes. You'll essentially never see a 10% clockspeed boost result in 10% faster completion of a benchmark more sophisticated than dhrystone, because you're still limited by the speed of your memory.
It really depends on the benchmark. Lots of real-world tasks remain highly cache-bound. One important one for me personally is building software. I run big system builds for Zephyr test suites and Chromeos regularly on 64 core devices, and have access to systems with both 4-channel and 8-channel DRAM setups. The doubled memory bandwidth makes for only about 5% higher build throughput.

Honestly I think a lot of modern benchmarkery has ended up too far down the "cover all the use cases" rabbit holes, and there's not enough coverage of the kind of boring scalar workloads that most of these systems are being purchased to run.

A pity the TDP went up a lot, but in order to compete with Intel, they likely had to do that.
I'm confused by this image on this point too: https://images.anandtech.com/doci/17552/Ryzen%207000%20Tech%...

Is the intention that some will cap the voltage at the wall?

They’re showing the performance difference at different points along the power curve, since it’s non linear.
You can limit the maximum boost clocks to reduce power draw and thermal dissipation. Some have reported that they have set this in the motherboard settings.

At any rate, this demonstrates how much more efficient the Zen 4 cores at lower power when compared to Zen 3. If you need high clock speed across cores, you won't gain as much of the efficiency, but the potential here for laptops based on Zen 4 is extremely promising. In those cases, chips are often configured for a maximum TDP between 15W and 65W depending on use case (though some high end workstations and gaming machines are set even higher.)

Not at the wall, but I'm sure some folks will limit the TDP in BIOS/UEFI to lower noise/temperature/power draw.

I also expect some vendors to sell smaller computers that force the CPU to lower TDP due to insufficient cooling.

But, I expect the main reason for that slide is to draw attention away from the huge TDP increase.

AMD allows the user to set the TDP for their CPU. The branding they use for this is "Configurable TDP" or "cTDP". You can set it in your bios, or possibly using some AMD-specific cpu tuning program directly from windows.
They are covering both ends of the power envelope, with greatly increased efficiency (they claim up to 74% improved efficiency at 65W TDP), but also increasing the overall TDP to allow higher multi-core clock speeds and, obviously, higher single-core clock speeds.
Discussion from last night based on Anandtech's coverage:

https://news.ycombinator.com/item?id=32644878 (66 comments)

The Anandtech article[0] includes some more details about efficiency, particular at any given TDP.

> Interestingly, AMD offered performance figures for three different TDPs: 65W, 105W, and 170W. The greatest performance gains are actually at the lowest TDPs, where the 7950X saw a 74% increase in Cinebench R23 MT performance. These advantages actually decreased as TDPs went up, dropping to 37% at 105W, and finally 35% at 170W.

[0] https://www.anandtech.com/show/17552/amd-details-ryzen-7000-...

> AMD Ryzen 9 7950X 16C/32T Up to 5.7 / 4.5 GHZ 80MB Gen 5 170W $699

With ECC, this will be an amazing perf/cost server that can handle substantial load.

Besides Hetzner & OVH (who both offer AMD 59xx series chip), does anyone else offer these variant chips for dedicated hosting needs?

It looks like good competition on the high end, but these numbers would indicate AMD is not much ahead of Intel in performance/price around the 12600k tier. Alder Lake i3/i5 buyers in the last year are probably pretty happy with their purchases.
(comment deleted)
AMD should make a CPU for home-users without the PSP.
TDP starting at 105W, i'll pass
Can I drop it into my ASUS Prime B350 PLUS mainboard from four years ago?
No, these CPUs use the new AM5 socket.
Any confirmation that these include Pluton?