I'm mixed. Definitely a very competent chip, great perf & efficiency gains, a fairly lovely new platform.
$700 for the 16-core 7950X chip is well into HEDT price range, and while this has the CPU oomph everything else is missing. It's more $/core than what EPYC launched at, with a 16 core EPYC 7281 running $650. That had half the MHz & wasn't as good an architecture, but it also 128 lanes of PCIe (3.0 in Naples then 4.0 in Rome) and 8 channels of memory, versus 28 PCIe lanes (5.0!) / 2 channel memory (oof); that's a huge advantage.
Hopefully Intel can start shipping new process nodes again & the competition can come-back.
The Ryzen 9 5950x launched just under two years ago at $800. This is an improvement on those prices.
The improvements in IPC over the original Zen core of the EPYC 7281 are quite substantial, and AMD is in a much better position than it was those five long years ago.
One downside is that the TDP is almost doubled (170/~230 W vs 105/140 W for default/PPT) which may be a dealbreaker - certainly for my SFF system where I'm not sure I can cope with another 100W. These are limited by the socket really (AM4/AM5), but still it's nice to have that many threads in such a low power budget. I'm more curious to see what the 32-thread competition turns out to be in the same TDP bracket.
It's not that made up. 5950X used 120-140 W and I expect the 7950X to use 210-230 W. AMD said that the increased multicore performance comes from a higher power limit.
I’ve no doubt that the newer cpu will use more power, but the numbers are made up. They’re picked by the marketing department using a formula, and calling them “watts” to try to make them look like energy used or heat dissipated doesn’t make them real. They don’t correspond to numbers that you can actually measure in any way.
They measured one “105W” TDP chip maxing out at 86 Watts, and another “105W” TDP chip maxing out at 144 Watts? That’s not “pretty well aligned” with anything.
See the part of the article starting at “We’ll first start with iso (like-for-like) comparisons between Ryzen generations, with iso-power.” That shows “up to +37%” at 105W.
See my original post. I have an SFF case with a water-cooled 5950X+3090. That's already borderline for noise and thermals on a 240mm radiator, so an extra 100 W for the 7950X would probably be untenable. But as people point out - more perf per watt would probably mean cooler performance in a lot of workloads. It'll be interesting to see what the ITX options are for AM5 anyway.
Try to think of TDP as "thermal headroom." It's not efficiency or power consumption. It's room for the CPU to speed up when it's necessary. The boost max going from 4.9Ghz to 5.7Ghz (16%), and at the high end, the power necessary to increase clocks ramps up.
In another slide, they indicated ~62% lower power at the same performance levels of the Ryzen 5000 CPUs. So more efficient, but also able to clock higher and get work done faster.
For a thermally constrained system, you won't achieve the absolute maximum performance if the CPU throttles, but you do very much benefit from a more efficient architecture (+13% IPC) and process node.
imo, it's the opposite. compared to the original thread rippers, this chip has way more pcie (gen 5 vs gen3) and although it has less ram bandwidth, 64mb of l3 cache mean you don't need it as much. this costs as much as the original threadrippers did because they are the replacement for that class of chips.
So for a similar launch price you get >2x the performance, almost identical pcie bandwidth (4x28) and faster memory (less of it though). Seems pretty great to me?
It's good that AMD doubled the lanes on the PCH as 4 was a real limit (some of the higher-end motherboards have 3xM.2s plus Wifi etc hanging off the PCH, so definately a bottleneck).
8x DDR4-2666 -> 2x DDR5-5600 is half the memory throughput too. That was the first thing that made my head scratch; having much bigger cores & still lots of PCIe throughput, but way less memory bandwidth.
That said, 8x DDR4-2666 also scaled up to 64c chips in Naples, this isn't unreasonable.
But you're comparing consumer/desktop to HEDT/server. They prioritize different things - the memory bandwidth is less of a bottleneck on desktop. While your premise is that the pricing is HEDT (based on a first generation part from 2017), the current Threadripper is a $6500 part that does still carry 8 memory channels, up to 3200Mhz.
One would imagine that EPYC based on Zen 4c would have the memory throughput you crave, though I wouldn't hold your breath for $700 pricing. It's still hard to say how committed AMD is to the middle ground of HEDT.
One thing to consider is the growth of cache. If you have better L3 hitrate (which you should given the doubling of L3 cache and 10x growth of L2 cache), you ease up the memory pressure a bit.
Die Area of only 3.84mm2 for Core + L2. This is a similar to Apple A14/A15. On TSMC N5 although they are likely different variant, optimising for clockspeed rather than Low Power.
13% increase in IPC, and additional clock speed increase as well from N5. The server variant Zen4C with 128 Core per socket should compete fairly well with Ampere ARM variant.
I just hope AMD is a little more ambitious than they were and book way more 5nm capacity. They have been way too conservatives with their forecast on their CPU side.
Its been a very very long time since I was last excited about anything in the PC /x86 space. For the past 8 to 10 years it has been mostly about Smartphones, ARM and Apple.
Still sad I couldn't hold on to my AMD stock, I bought them at below $3. Sigh.
>I just hope AMD is a little more ambitious than they were and book way more 5nm capacity. They have been way too conservatives with their forecast on their CPU side.
I doubt they've been conservative. AMD has less than $6bn cash on hand. They're prepaying for stuff what they can afford, which they've said in statements totals $300-400MM at any given time. Apple can wire $6bn to TSMC tomorrow if they need to. AMD just can't play in that game.
TSMC demanding tighter payment terms and more up front to reserve capacity is really not good for AMD.
So like 5 years ago when the ryzen came out and made 8 cores accessible to the masses, you didn’t get a little tingle of excitement at the sudden competition?
We'll have to see real world performance and power usage, but their presentation included an estimated ~62% lower power at the same performance as their Zen 3 / 7nm Ryzen 5000 CPUs.
If that's even close to right, that should really start to shine when they have mobile APUs shipping.
This makes these machines a lot more viable in the business and engineering world as well where CPU performance is desired but graphics is not. I've seen in the past AMD Zen machines paired with say a GT710 simply because it was the cheapest GPU option available that would allow for basic apps and drive two screens.
What a coincidence, that was my setup as well but for my next(now current) pc, I changed it to a Radeon R7 250, (nvidia is fine but dealing with the drivers on linux was a pain in the neck)
Yeah, but Intel has the advantage of providing vGPU support for their integrated graphics. I don't think there is anything really stopping AMD but wanting money.
Intel barely supports this anymore to me, and their SRIOV support has not been publicized anywhere from what I can see beyond a commit on GitHub. Furthermore, they are not including this on Arc dGPUs from my understanding, so I imagine Intel is moving this to datacentre GPUs only as well.
Not sure if you were thinking of something else, but both NVIDIA and AMD could really approve in this regard. There is no reason to limit sharing your GPU across VMS except to primarily limit platforms.
I personally suspect that the true reason for that integrated GPU is the USB4 support. Always having that integrated GPU allows for full USB4 directly on the motherboard without needing hacks like looping a DisplayPort cable from the discrete GPU to the motherboard. And I believe it's possible to render in one GPU but present the output through another GPU, so you could in theory even use a discrete GPU for rendering while the output is going through the USB4 socket on the motherboard, without any extra cables between the graphics card and the motherboard.
I hope they also make more advanced models with regular GPUs. I have a 4300GE which runs fine non trivial games from just a few years ago (Far Cry 5, Prey 2017, etc) under Linux+WINE, and its quite low 35W TDP makes it a really nice CPU for home/office use with occasional non high end gaming.
Unfortunately it is quite difficult to build a power-efficient home server as the power consumption of many components is not clear. Take the motherboard for instance, you really have no idea how much is consumed by it. It is reasonable clear for the CPU, but for many other components it is not.
AMD is giving Intel a bit of breathing space at the lower end of the product stack, the 7600X price/perf isn't that great when you factor in the requirement for DDR5 + new AM5 socket.
My 3950 has an economy bios mode, bringing it down to 65 watts (for 16 cores). I expect the 7600X (and the others) will have something similar. If you're running linux, you can also override the cpu scaling governor, and either force the frequency down (or up), or you can set it to conservative mode. Conservative mode will preserve energy when the load is low, at an almost unnoticable cost to responsiveness, but still provide close to full power when the cpu is under real load.
Under load, AMD is still going to be a lot more energy efficient than current gen Intel, even without such a setting. And on idle, the TDP is not really relevant.
Currently have a 105W 5900X. Opening Visual Studio 2022 saw power consumption as high as about 68W for a blink of a second, but compiling a small web project only drove it to 30W. With my off the shelf Corsair AIO water cooling, it's usually around 35° C. For what it's worth!
In powershell, running some python / dbt uses even less power, mostly dipping below 20W. Really takes a lot to make it work hard. I imagine some Cinebench would get it going!
I'm impressesed by the initial performance numbers, the price, PCIe 5.0, Having 8 lanes to the PCH instead of 4, 10gbe on most motherboards. It's good that AMD is continueing the trend where users don't have to upgrade their motherboards every generation.
The main catch is DDR5 being expensive and compatability issues with four DIMMS[0]/ teething issues.
(Personally, I wish AMD released a 5900X/ 5950X with 3D V-Cache. The 5800X3D is nice, but I also need cores.)
To clarify, this is a new platform, but they have committed to maintaining the AM5 socket (and supporting platform) through at least some future advancements, as they did with AM4 from Zen, Zen+, Zen 2 and Zen 3.
It could have gone up (I think the EU/ UK will increase, we'll see). Still haven't heard anything on motherboard pricing (those motherboards with 20 powerstages aren't going to be cheap!).
Given the platform costs, the lowend 7600X doens't make sense. AMD does some weird product placing...
I didn't expect to see them hitting 5.8Ghz [stock], wow!
So the big questions on this and Intel next-gen process releases:
1) so it is just the process that gives the Apple ARM laptops their superior battery life?
2) if so (I doubt it will be), when will there be as good a performing laptop example in x86?
3) if not (as I suspect)... what is the actual response to the M1 gauntlet by Intel/AMD? THeir own ARM chips?
If Wintel tries to go ARM, it is the last opportunity for Linux desktop inroads, or perhaps an opportunity for Windows to go full-core Linux with windows as the UI.
At this point after 25 years of linux desktop being subpar, I welcome a major player, even goddamn Microsoft, and yes I remember every horrible thing Microsoft has done in the last 40 years of my life, to take the bull by the horns.
Microsoft is the lesser of two evils. Apple will dangerously close the freedom of the software on their machines to push everything to app stores. Microsoft, to their credit, I think the first time I have ever praised them, is more open in their platform approach historically.
Microsoft tried everything they could imagine to try and kill off Win32 and force everything and everyone onto Microsoft Store with locked down MetroModernUIImmersiveUWP software.
They all failed because the entire reason for Windows's existence is users can install and use practically anything anywhere anytime.
So no, Microsoft is not open. They do not want Win32 and its open nature. Microsoft is salivating at the kind of walled garden Google and Apple managed to erect, and would try for it again the next chance they get.
Sometimes market is wondering. Imagine anyone said freedom to choose is what Microsoft is. In a sense it was (as it basically copied Apple II open hardware and against macintosh closed system). In a sense it is not (it tried closed on the software part somehow and even get anti-trust on browser). Even then the “embrace, extend and extinguish” is well known. But now. It seems it is that major game. That is crazy. Can linux …
DDR5 JEDEC specification includes essentially a form of ECC so it may not be _as_ big of a deal for many users. It also means that setups on AM5 will have some level of ECC handling functionality on both the chipset and the CPU.
The ECC that is standard with consumer DDR5 is purely on-die, protecting the data at rest but not providing any ECC for the memory bus. It's transparent to the host system. As a result, the CPU and other components do not even have the option of participating in that ECC. It's not a substitute for and does not preclude the use of module-level ECC as in previous generations.
I'd be surprised if it didn't support ECC; every gen since Ryzen 2000 has supported ECC (although it's unadvertised and you have to be very careful with motherboard selection). I currently have proper ECC running on a 3000 series Ryzen with an AsRock motherboard.
I wonder if they ended up adding so many MLCCs to the top of the substrate to try to hit the higher frequency targets (especially with higher multicore power draw), or if they were required due to the different process or uArch changes.
Probably the former, which isn't as interesting, but it is visually striking either way.
I wonder how this is going to look like in real world tests and if the motherboards will come as buggy as for 5XXX series, where it took almost a year to get a stable system.
I am feeling a bit sceptical, as I am not sure if these numbers are that impressing in comparison to previous gen. I am worried that AMD is starting to follow the Intel path. Hyping up the new line up, while selling essentially the same CPU with some cosmetic changes that in real world get you that 1% performance bump on a full moon.
I just got a 5900H laptop. Suspend doesn't work (black screen, no response to any input), nouveau oopses if used, no HDMI out. Until the nouveau driver runs and oopses, it runs hot.
I have 5950X in my desktop and it took about a year before the motherboard got a stable BIOS. Before that bluescreens happened to me on a weekly basis and connecting anything to USB had to be prepended with a deep breath and saving all work in case system crashes.
78 comments
[ 2.4 ms ] story [ 153 ms ] thread$700 for the 16-core 7950X chip is well into HEDT price range, and while this has the CPU oomph everything else is missing. It's more $/core than what EPYC launched at, with a 16 core EPYC 7281 running $650. That had half the MHz & wasn't as good an architecture, but it also 128 lanes of PCIe (3.0 in Naples then 4.0 in Rome) and 8 channels of memory, versus 28 PCIe lanes (5.0!) / 2 channel memory (oof); that's a huge advantage.
Hopefully Intel can start shipping new process nodes again & the competition can come-back.
The improvements in IPC over the original Zen core of the EPYC 7281 are quite substantial, and AMD is in a much better position than it was those five long years ago.
In another slide, they indicated ~62% lower power at the same performance levels of the Ryzen 5000 CPUs. So more efficient, but also able to clock higher and get work done faster.
For a thermally constrained system, you won't achieve the absolute maximum performance if the CPU throttles, but you do very much benefit from a more efficient architecture (+13% IPC) and process node.
That said, 8x DDR4-2666 also scaled up to 64c chips in Naples, this isn't unreasonable.
One would imagine that EPYC based on Zen 4c would have the memory throughput you crave, though I wouldn't hold your breath for $700 pricing. It's still hard to say how committed AMD is to the middle ground of HEDT.
https://www.amd.com/en/products/cpu/amd-ryzen-threadripper-p...
https://www.servethehome.com/amd-ryzen-threadripper-pro-5995...
13% increase in IPC, and additional clock speed increase as well from N5. The server variant Zen4C with 128 Core per socket should compete fairly well with Ampere ARM variant.
I just hope AMD is a little more ambitious than they were and book way more 5nm capacity. They have been way too conservatives with their forecast on their CPU side.
Its been a very very long time since I was last excited about anything in the PC /x86 space. For the past 8 to 10 years it has been mostly about Smartphones, ARM and Apple.
Still sad I couldn't hold on to my AMD stock, I bought them at below $3. Sigh.
I doubt they've been conservative. AMD has less than $6bn cash on hand. They're prepaying for stuff what they can afford, which they've said in statements totals $300-400MM at any given time. Apple can wire $6bn to TSMC tomorrow if they need to. AMD just can't play in that game.
TSMC demanding tighter payment terms and more up front to reserve capacity is really not good for AMD.
If that's even close to right, that should really start to shine when they have mobile APUs shipping.
https://wiki.archlinux.org/title/Intel_GVT-g
Not sure if you were thinking of something else, but both NVIDIA and AMD could really approve in this regard. There is no reason to limit sharing your GPU across VMS except to primarily limit platforms.
Under load, AMD is still going to be a lot more energy efficient than current gen Intel, even without such a setting. And on idle, the TDP is not really relevant.
In powershell, running some python / dbt uses even less power, mostly dipping below 20W. Really takes a lot to make it work hard. I imagine some Cinebench would get it going!
The main catch is DDR5 being expensive and compatability issues with four DIMMS[0]/ teething issues.
(Personally, I wish AMD released a 5900X/ 5950X with 3D V-Cache. The 5800X3D is nice, but I also need cores.)
[0] https://www.youtube.com/watch?v=uu9U7TVNImI
Given the platform costs, the lowend 7600X doens't make sense. AMD does some weird product placing...
I didn't expect to see them hitting 5.8Ghz [stock], wow!
1) so it is just the process that gives the Apple ARM laptops their superior battery life?
2) if so (I doubt it will be), when will there be as good a performing laptop example in x86?
3) if not (as I suspect)... what is the actual response to the M1 gauntlet by Intel/AMD? THeir own ARM chips?
If Wintel tries to go ARM, it is the last opportunity for Linux desktop inroads, or perhaps an opportunity for Windows to go full-core Linux with windows as the UI.
At this point after 25 years of linux desktop being subpar, I welcome a major player, even goddamn Microsoft, and yes I remember every horrible thing Microsoft has done in the last 40 years of my life, to take the bull by the horns.
Microsoft is the lesser of two evils. Apple will dangerously close the freedom of the software on their machines to push everything to app stores. Microsoft, to their credit, I think the first time I have ever praised them, is more open in their platform approach historically.
But... nah, it won't happen.
They all failed because the entire reason for Windows's existence is users can install and use practically anything anywhere anytime.
So no, Microsoft is not open. They do not want Win32 and its open nature. Microsoft is salivating at the kind of walled garden Google and Apple managed to erect, and would try for it again the next chance they get.
(And I don’t just mean it can use the RAM, I mean for it to gain the actual benefits of ECC)
On-die ECC also isn't new with DDR5; it's been used for years on at least some LPDDR4 and LPDDR4x memories (eg. https://www.micron.com/about/blog/2017/february/the-advantag...) and even longer for special-purpose DRAM parts.
Probably the former, which isn't as interesting, but it is visually striking either way.
I am feeling a bit sceptical, as I am not sure if these numbers are that impressing in comparison to previous gen. I am worried that AMD is starting to follow the Intel path. Hyping up the new line up, while selling essentially the same CPU with some cosmetic changes that in real world get you that 1% performance bump on a full moon.
Hoping things will get better.