Fun seeing performance reviews of new chips coming out. It's a sign to me that the market is recovering after the pandemic. Imagine someone promoting a chip during the pandemic, when you couldn't get them for love or money.
Still, I would have liked to see a graph or other pretty chart showing the results of the benchmark off better so that I could internalize them easier.
I don't know but for what it's worth my main machine is a 6000 series laptop (with Pluton) running Linux and I did not have any compatibility issues. Sure, it sucks to have some Microsoft designed hardware in my CPU but at least it not causing issues (for now).
You have two ring -3 coprocessors with unrestricted DMA, unrestricted disk I/O, and unrestricted access to your network interface. One belongs to the NSA, the other to Microsoft.
Do some traffic analysis on the upstream end of an ethernet cable plugged into that computer while it is hibernating or sleeping some time, you might not like what you what you find.
Are you willing to give up the reason for your secrecy? I'm not sure what you stand to lose if you just say "I saw some traffic to NSA headquarters every 10 minutes that coincide with access entries on my SSD"
Yes. Divulging the destination on a public platform is providing more identifying information than I am comfortable sharing.
It's funny that you'd think real-world espionage by intelligence agencies would be sending that data to headquarters rather than some random commercial VPS set up as collection infrastructure that is deliberately unattributable to the organization behind the espionage.
>Do some traffic analysis on the upstream end of an ethernet cable plugged into that computer while it is hibernating or sleeping some time, you might not like what you what you find.
Is this something you know for a fact, or are you extrapolating from intel ME?
It's behavior I've witnessed on my own system w/ AMD PSP. I can't definitively attribute it to PSP, but I can't attribute it definitively to anything else either.
What method did you use to do the traffic analysis? I mean, under assumption that this kind of traffic can only be sniffed at the wire level, what exactly did you do to accomplish this? I am genuinely interested.
Allowlisting, yes. Keep in mind that even fairly unsophisticated malware has been observed using channels like pastebin and Twitter for exfil/c2.
Blocklisting, that's a cat and mouse game. Go look at how many different URLs and IPs are utilized for commercial telemetry in the likes of Adobe and Microsoft software if you aren't familiar.
Zen 4 (Raphael) desktop CPUs will have it. Right now, Zen 3 (Chagall) and Zen 4 Threadripper (Storm Peak) don't currently have plans for integrating it, but that may be subject to change. I cannot provide proof, but I sure as hell am never buying Zen 4 or later AMD Desktop CPUs ever again for a system that will be connected to the internet.
I would not go that far saying I won't buy AMD ever again, but if Microsoft has a foot in the door inside my systems in form of creepware like Pluton it is there for keeping the door open.
I am also paranoiac about such things, but the part with avoiding the connection to the Internet is easy for a desktop computer, if it also does not run Windows, where you never know what services might be active and with which external servers they might try to communicate.
I have a small computer that is used as the router for the Internet connection and on which I have complete control over the firewall and over the proxies used for connecting to the Internet, so there are no chances for an unwanted connection to the Internet.
For additional protection, one could run the Internet browsers in a Virtual Machine, to be absolutely certain that there is no way for a script run by the browser to access the physical hardware, assuming that you do not trust the sandboxing capabilities of the browser.
VMs are an imperfect solution to preventing programs running in the guest OS from affecting the host OS (imperfect, as VM escapes are a thing).
VMs do not provide protection to the guest OS (or it's processes) from inspection and snooping by the host OS, or from inspection / snooping by the hardware the host OS is running on.
If using a bare metal hypervisor, sed 's/host OS/hypervisor/g'
FWIW it's already present on Zen 3+ (Rembrandt) CPUs. However, at least on lenovo you can disable it (there's an option in the BIOS that allows you to toggle between TPM 2.0 and pluton).
No, Intel's 13th generation (Raptor Lake) chips are slated to have it too.
I suspect that like with Threadrippers, some Xeons will not ship with it, but I am not as confident about this as I am for Threadrippers, for reasons I cannot divulge.
Maybe I'm interpreting this comment incorrectly but I'm begging people to realize if they're a normal person and they have the NSA at the top of their threat model they're bad at internalizing relative risk.
You do appear to be interpreting it incorrectly. This is about Pluton, for which the threat actor is Microsoft. People concerned about the NSA would need to be worried going all the way back to the introduction of AMD PSP back in 2011.
Now, with regards to the important part of your message, I'm begging people to realize that if they're blindly trusting all code running on all firmware in their personal machine, they are even worse at accurately internalizing risk than those who are proactively paranoid.
I would argue not trusting MS is even sillier. Like it or not, they run most desktops around the world. Even if you only run the purest NetBSD at home, nearly everything someone knows about you will touch Windows at some point. In terms of secureboot et al, MS hasn't done any of the shenanigans people were worried about. You can still turn off SB on every x86 desktop. What is the actual risk here? At the end of the day, MS isn't the government.
Raptor Systems is right there for the truly paranoid.
You're right that MS isn't the government - they have fewer restrictions on behavior than government does, and a financial incentive to snoop on people the government doesn't care about.
If you're aware of the risks, you do you, just don't kid yourself into thinking the third most valuable company on the planet is diligently concerned with protecting your privacy and putting your interests above their own.
based on what I heard from insiders, two batches of reviews for AMD's new chips, those released now are the ones from sites traditionally happy to provide overwhelmingly positive reviews, those going to be released tomorrow are the ones which usually provide more balanced reviews.
good PR tricks - they just sign NDA agreements with different reviewers with slightly different embargo lift dates.
As this specific review is from Phoronix - do you have some proof that their reviews are (heavily) biased? I'm looking at reviews in more details only when I want to buy something, other than that I skip them or just have a quick look. I haven't did any serious comparison either, but I remember Michael Larabel from Phoronix criticising various parties. Many of his reviews are basically measures/benchmarks with some summary on top of that - his benchmarking methodology may be or may be not flawed, I don't remember him shilling for AMD. AMD has probably some goodwill from the Linux community since they open source or provide various specs for their graphics card. At least that used to be the case.
I was just passing on something claimed by the largest computer parts review site in China. They claimed that their site has been required not to release their review until tomorrow, the owner of the site further claimed that there are two batches of reviews and his site is in the second batch for obvious reasons.
FWIW, at the AMD event in Texas last month where the Ryzen 7000 series was announced, everyone was given the same embargo date... Mostly all the US and EU reviewers there.
Maybe the China/Asia review briefings or so are going with a different date based on different retail availability or so? Hadn't heard of some reviewers getting a different embargo date, typically see that usually only for different geo availability or if getting a sample from a motherboard OEM / other partner and not the CPU company review embargo.
Everybody is doing dirty PR tricks these days, cash flow is above morals for a long time. People overall don't mind so don't expect change for the better.
Anyway, enough ranting, nobody sane decides on these 0day benchmarks but waits at least few weeks for overall conclusions, quirks, not so common issues etc. And even better is to wait few months for more stable motherboards and drivers, better availability of compatible RAM etc.
Can you suggest some of these less positive reviewers please? I like to have a broad range of reviews before deciding and all the ones I normally look at came out today (though were not overwhelmingly positive in every area).
chiphell is the largest computer parts review site in China, and that nApoleon guy posting this two batches of review story is the owner of the site.
In a more recent thread, he confirmed again that his site has completed their own review and he had read it, yet they are not allowed to post it until tomorrow.
In a separate thread he posted today, he mentioned that based on their tests, the temp of the zen4 chip is always 95 degrees celsius - no matter whether you use a shitty cheap cooler or a fancy kick ass one.
Would be interesting to know if benchmarks run faster with a better cooler. It could be at the thermal limit in both cases, just at different clock speeds.
Interesting but is delaying negative reviews works? It seems that GN posts review with temperature criticism today (see also thumbnail). https://www.youtube.com/watch?v=nRaJXZMOMPU
Not sure why the 95C thing is bad. If a chip has work to do and can control it's overclocking, why not have it stick to whatever safe temperature it wants?
What are you talking about? E.g Anandtech and GamersNexus (https://www.youtube.com/watch?v=nRaJXZMOMPU) are hardly examples of being "overwhelmingly positive" towards anyone.
Heise Online has released their review (alas, behind a paywall) and it starts very positive. I'm pretty sure they are the most trusted reviewers in the german language.
I recall that back in the 90s, besides higher clock speeds, one of the greatest reasons for excitement about a new CPU was instruction set extensions (MMX, 3DNow, etc), which could give outsized performance gains when software was updated to work with them. This latest release from AMD has similar cause for excitement: AVX-512 is more than just larger vectors, it also doubles the number of registers, adds per-lane mask registers, and other enhancements; unlike with Intel processors, there should be no worry of slowing down the clock for the whole processor when using these instructions with larger vectors, and unlike with Intel, we can reasonably expect that every AMD processor from now on will always have these instructions enabled.
I'm personally curious about AVX512 support in pyTorch and numpy. Numpy depends on BLAS libraries for performance, and I don't know how much dev resources are behind openBLAS (i.e. how soon I can expect performance increases), especially given that AMD is a much smaller company and doesn't spend as much as Intel on software.
Interestingly even software (oneDNN) that is using Intel's brand new OneAPI is running excellent one Zen 4 CPUs, often faster than on Intel's own CPUs:
There are those rumors. In fact, there's no announced 7800X yet. The speculation is they launch a 7800X3D a few weeks to a few months from now with 3D cache from the beginning in that 7800 name space. To me that seems likely. There's a rather more optimistic rumor that all the 7000 series will eventually have 3D cache versions, which I find maybe too optimistic.
No need for rumours, AMD already said they will be bringing out 7000 series chips with V-cache. This was during a investor presentation, the specific products haven't been announced yet.
> unlike with Intel processors, there should be no worry of slowing down the clock for the whole processor when using these instructions with larger vectors
because they don't implement "true" 512bit registers for AVX-512. This is basically to support instruction set rather than SIMD performance boost.
> On average for the tested AVX-512 workloads, making use of the AVX-512 instructions led to around 59% higher performance compared to when artificially limiting the Ryzen 9 7950X to AVX2 / no-AVX512.
It appears that Zen 4 implements around 192 registers of 512 bits. However, the latest Intel model (Sapphire Rapids) implements more 512-bit registers, around 332 such registers.
For most operations, Zen 4 has the same execution units as Zen 3, it has not expanded them to 512 bits, but the execution units of Zen 3 are enough to match the 512-bit throughput of the Intel CPUs. The only exception is that Zen 4 can do one 512-bit FMA and one 512-bit FADD per cycle, while the most expensive of the server and workstation Intel CPUs have a second 512-bit FMA unit, so they can do 1 FMA + 1 FMA, instead of 1 FMA + 1 FADD. The cheaper Intel CPUs with AVX-512 are worse than Zen 4, because they can do only 1 FMA, not also an FADD, like Zen 4.
On the other hand Zen 4 implements a true 512-bit shuffle unit, while the current Intel CPUs have only 256-bit shuffle units, slowing down the shuffles that must cross the register halves.
As tested at that link, while there are a few cases when Zen 4 has slower implementations for some 512-bit instructions, in many cases the AVX-512 implementation in Zen 4 is better than in the current Intel CPUs.
What would be the purpose? If you want to purposefully limit power you're forcing the chip to underperform. You can just get the lower cost chip that does run at that power. For example the 5600x being 76W versus the higher chips.
> What would be the purpose? If you want to purposefully limit power you're forcing the chip to underperform
The rationale is that each architecture has a sweet spot consumption/performance, and increasing performance over a certain threshold has increasing costs (AFAIK, one significant factor is the increase of voltage in order to push frequencies higher).
I don't know the numbers, but AMD has been offering the 65W threshold for a while, so it can be reasonably assumed that 65W is the sweet spot performance/consumption.
The reason for not buying directly a cheaper model is the number of cores - one may choose, say, a 65W-throttled CPU with 16 cores over an unthrottled 8-cores (in some scenarios, this can make sense).
In addition to this, you can choose to run your top-of-the-line multi-core processor in eco mode most of the time, and then consciously turn it back to stock settings when you know you'll need the extra speed/throughput, only bearing the extra heat and cost of that performance when you really need it.
As shown in the anandtech review above, the 7950x limited to 65W (ECO Mode) handily beats the 7600x at (uncapped) 105W, and anecdotally everything that Intel has to offer as of today...
Both are very significant limiting factors for non-gaming applications.
Top gaming rigs are the dragsters of computing: no power consumption is too high, no hardware contraption is too large. But most people, and most businesses, don't drive dragsters.
But a 7950x is just a 7900x (They have identical silicon) that is just "lucky" enough to survive more power draw. If you make the TDP the same you just end up with the same CPU
Anyway fewer turned on? They need something to sell when one or more of the cores don't work. Would be nice to let you have 15 cores when that many of them work.
This is my request for CPU review. It's good way to know core efficiency. TDP or turbo watts is now artificially set very high by manufacturer but not efficient. Maybe reviewers are too busy to take many benchmarks before embargo.
No, I mean that it is physically possible to perform an unlimited amount of computation on a fixed amount of energy, if you are willing to wait long enough. Nobody has built hardware that has this property.
That is a useful measurement, but it's not quite what I'd want.
In that test, the CPUs are running on an open-air test bench with a 360mm AIO liquid cooler with the fans running at 100%. I don't think any of us are going to be running the CPUs like that, and that type of cooling encourages the CPUs to draw way more power - because they can draw much more power without overheating.
What I'd want in a test is to take some reasonable cooling solution (reasonable varies from person to person, of course) like a high-quality 140mm tower cooler or a 240mm AIO, in a computer case with the fans running at a silent or slightly above silent level. Then see what type of power and performance you get out of it.
I just want a CPU freq to power chart, so I know how Intel/AMD are pushing TDP and at what freq things start getting exponential. It would be interesting to see what is considered "normal".
There was a test done on 12th gen Intel (12900 variant IIRC, but my memory is hazy). They limited TDP incrementally down to single-digit numbers, the efficiency kept increasing the farther down you go. My takeaway was that there isn't a "normal" or "optimum".
Presumably that's when you hit the minimum voltage the CPU is allowed to run at. There's still a few other gotchas though.
As CPU frequency goes down, it waits less cycles for RAM. Practically it's a bit of a copout since keeping your RAM at high clocks/voltages doesn't effect more CPU TDP but does use more power. It might explain some of the irregularities in testing at least.
You can run it in 65W or 105W Eco mode. It will get the power consumption down a lot. Single thread performance will be the same and in multi thread it will still beat 12900K even in 65W Eco mode.
I was impressed the 7950 did so well at 65 watts, 81% of the perf at 38% of the power. I'd expect the lower end chips with less cores to have even less of a compromise to get to 65 watts.
I'm a life long Linux user, it's not even a thing for me it's just what I've used since my teens. But phoronix saying something has good benchmarks in Linux means absolutely squat to me. That site _always_ has good performance metrics on Linux.
I do some gaming and I can tell you their posts never reflect my reality.
Not Phoronix, all my experience with them is just extremely good performance on Linux and always worse performance on Windows. The whole site seems to be written and edited by one dude. That's why I prefaced by saying I not just love Linux, it's all I've known for over 25 years. So I have every reason to be biased, but even I think there's something shady with phoronix.
According to other reviews, the 7000 series destroys the competition (both Intel and 5000 series) at 65W.
Also, 7000 series is designed for 95C. It'll run forever at 95C, and will always run at 95C if it has the workload and power budget. IOW, giving it more cooling will cause it to run faster rather than making it cooler.
On other chips 95C might indicate it's on the edge of failure, but AMD asserts that 95C won't shorten the 7000's life.
Ryzen scales up exponentially in power draw as you try to hit peak frequency numbers.
You can run it at 95% speed and usually something like 70% voltage.
My 3900x can do around 4.5ghz at 1.4v or 4.3ghz at 1.05v all core. I imagine Zen 4 will be exactly the same. Don't have the power draw off the top of my head, but that is essentially 125w for 4.5ghz and 90w for 4.3ghz.
Does anyone know how to apply to be a 'reviewer' for pre-released CPUs from AMD or Intel? I am building a new kind of data management platform that is highly threaded so it can do DB queries or other data manipulation functions in parallel. I will benchmark it against other common systems like Postgres (https://www.youtube.com/watch?v=OVICKCkWMZE) but I only do that when I upgrade my own personal desktop or laptop. I would love to get my hands on loaner hardware to run benchmarks and publish them.
That would only really work if I already had the hardware for a meaningful benchmark. Right now my desktop has a 5950x and my laptop has a 6700hq. Showing a benchmark between those two systems would not be that interesting. I would like to show how my system performs on a 7950x vs my 5950x but I don't want to go out and buy another $2000+ computer just to get those numbers.
It’s annoying when hardware sites compare new CPUs to a big list of older or less capable ones, but precisely all of them are basically so new that if you have one of them you probably aren’t upgrading now. If you have Ryzen 2/3/4 or 9/10th gen Intel you have to guess what kind of performance leap you’d get with this.
118 comments
[ 4.1 ms ] story [ 182 ms ] threadStill, I would have liked to see a graph or other pretty chart showing the results of the benchmark off better so that I could internalize them easier.
https://www.phoronix.com/review/ryzen-5900x-5950x
Do some traffic analysis on the upstream end of an ethernet cable plugged into that computer while it is hibernating or sleeping some time, you might not like what you what you find.
It's funny that you'd think real-world espionage by intelligence agencies would be sending that data to headquarters rather than some random commercial VPS set up as collection infrastructure that is deliberately unattributable to the organization behind the espionage.
Is this something you know for a fact, or are you extrapolating from intel ME?
Blocklisting, that's a cat and mouse game. Go look at how many different URLs and IPs are utilized for commercial telemetry in the likes of Adobe and Microsoft software if you aren't familiar.
I have a small computer that is used as the router for the Internet connection and on which I have complete control over the firewall and over the proxies used for connecting to the Internet, so there are no chances for an unwanted connection to the Internet.
For additional protection, one could run the Internet browsers in a Virtual Machine, to be absolutely certain that there is no way for a script run by the browser to access the physical hardware, assuming that you do not trust the sandboxing capabilities of the browser.
VMs do not provide protection to the guest OS (or it's processes) from inspection and snooping by the host OS, or from inspection / snooping by the hardware the host OS is running on.
If using a bare metal hypervisor, sed 's/host OS/hypervisor/g'
(I know about Talos.)
I suspect that like with Threadrippers, some Xeons will not ship with it, but I am not as confident about this as I am for Threadrippers, for reasons I cannot divulge.
Now, with regards to the important part of your message, I'm begging people to realize that if they're blindly trusting all code running on all firmware in their personal machine, they are even worse at accurately internalizing risk than those who are proactively paranoid.
Raptor Systems is right there for the truly paranoid.
If you're aware of the risks, you do you, just don't kid yourself into thinking the third most valuable company on the planet is diligently concerned with protecting your privacy and putting your interests above their own.
good PR tricks - they just sign NDA agreements with different reviewers with slightly different embargo lift dates.
sadly, the same can't be said for hundreds/thousands other much smaller reviewers.
See the url in my replies.
Maybe the China/Asia review briefings or so are going with a different date based on different retail availability or so? Hadn't heard of some reviewers getting a different embargo date, typically see that usually only for different geo availability or if getting a sample from a motherboard OEM / other partner and not the CPU company review embargo.
Anyway, enough ranting, nobody sane decides on these 0day benchmarks but waits at least few weeks for overall conclusions, quirks, not so common issues etc. And even better is to wait few months for more stable motherboards and drivers, better availability of compatible RAM etc.
The source of the claim is from here (in Chinese) -
https://www.chiphell.com/thread-2445527-1-4.html
chiphell is the largest computer parts review site in China, and that nApoleon guy posting this two batches of review story is the owner of the site.
In a more recent thread, he confirmed again that his site has completed their own review and he had read it, yet they are not allowed to post it until tomorrow.
https://www.chiphell.com/thread-2446331-1-1.html
In a separate thread he posted today, he mentioned that based on their tests, the temp of the zen4 chip is always 95 degrees celsius - no matter whether you use a shitty cheap cooler or a fancy kick ass one.
https://www.chiphell.com/thread-2446302-1-1.html
https://www.phoronix.com/review/amd-ryzen-7900x-7950x-linux/...
I'm surprised Intel has basically abandoned the idea, and AMD just markets it for "gaming".
Checking out the numbers for these non-3d models, the large cache SKUs are going to be pretty amazing. Likely expensive too
Most speculation/leaks point to CES (early January) would be the announcement so not this year.
because they don't implement "true" 512bit registers for AVX-512. This is basically to support instruction set rather than SIMD performance boost.
> On average for the tested AVX-512 workloads, making use of the AVX-512 instructions led to around 59% higher performance compared to when artificially limiting the Ryzen 9 7950X to AVX2 / no-AVX512.
See the other thread on HN, which points to https://www.mersenneforum.org/showthread.php?p=614191
It appears that Zen 4 implements around 192 registers of 512 bits. However, the latest Intel model (Sapphire Rapids) implements more 512-bit registers, around 332 such registers.
For most operations, Zen 4 has the same execution units as Zen 3, it has not expanded them to 512 bits, but the execution units of Zen 3 are enough to match the 512-bit throughput of the Intel CPUs. The only exception is that Zen 4 can do one 512-bit FMA and one 512-bit FADD per cycle, while the most expensive of the server and workstation Intel CPUs have a second 512-bit FMA unit, so they can do 1 FMA + 1 FMA, instead of 1 FMA + 1 FADD. The cheaper Intel CPUs with AVX-512 are worse than Zen 4, because they can do only 1 FMA, not also an FADD, like Zen 4.
On the other hand Zen 4 implements a true 512-bit shuffle unit, while the current Intel CPUs have only 256-bit shuffle units, slowing down the shuffles that must cross the register halves.
As tested at that link, while there are a few cases when Zen 4 has slower implementations for some 512-bit instructions, in many cases the AVX-512 implementation in Zen 4 is better than in the current Intel CPUs.
Also, there is no idle power comparison, something that may also differ significantly and affect Linux.
The rationale is that each architecture has a sweet spot consumption/performance, and increasing performance over a certain threshold has increasing costs (AFAIK, one significant factor is the increase of voltage in order to push frequencies higher).
I don't know the numbers, but AMD has been offering the 65W threshold for a while, so it can be reasonably assumed that 65W is the sweet spot performance/consumption.
The reason for not buying directly a cheaper model is the number of cores - one may choose, say, a 65W-throttled CPU with 16 cores over an unthrottled 8-cores (in some scenarios, this can make sense).
https://www.anandtech.com/show/17585/amd-zen-4-ryzen-9-7950x...
(2) Cooling budget on servers.
Both are very significant limiting factors for non-gaming applications.
Top gaming rigs are the dragsters of computing: no power consumption is too high, no hardware contraption is too large. But most people, and most businesses, don't drive dragsters.
Basically how many watts required to do some fix unit of work (to render 1 frame using Blender CPU rendering)
You can do an unlimited amount of computation on one joule. One watt-second, if you like.
Not an infinite amount of computation, but any finite amount.
In that test, the CPUs are running on an open-air test bench with a 360mm AIO liquid cooler with the fans running at 100%. I don't think any of us are going to be running the CPUs like that, and that type of cooling encourages the CPUs to draw way more power - because they can draw much more power without overheating.
What I'd want in a test is to take some reasonable cooling solution (reasonable varies from person to person, of course) like a high-quality 140mm tower cooler or a 240mm AIO, in a computer case with the fans running at a silent or slightly above silent level. Then see what type of power and performance you get out of it.
As CPU frequency goes down, it waits less cycles for RAM. Practically it's a bit of a copout since keeping your RAM at high clocks/voltages doesn't effect more CPU TDP but does use more power. It might explain some of the irregularities in testing at least.
https://youtu.be/uks4qQ2MXrM?t=1087
https://www.anandtech.com/show/17585/amd-zen-4-ryzen-9-7950x...
I do some gaming and I can tell you their posts never reflect my reality.
Also, 7000 series is designed for 95C. It'll run forever at 95C, and will always run at 95C if it has the workload and power budget. IOW, giving it more cooling will cause it to run faster rather than making it cooler.
On other chips 95C might indicate it's on the edge of failure, but AMD asserts that 95C won't shorten the 7000's life.
You can run it at 95% speed and usually something like 70% voltage.
My 3900x can do around 4.5ghz at 1.4v or 4.3ghz at 1.05v all core. I imagine Zen 4 will be exactly the same. Don't have the power draw off the top of my head, but that is essentially 125w for 4.5ghz and 90w for 4.3ghz.
And then there is a greater chance that phoronix will include it in its tests; or others will run it and publish the results.
so everybody can verify and compare the results.
If you create a benchmark; everybody can run .. and you can ask your networks for run this public test and upload the results.
On the other hand you can wait for ZEN4 Epyc ... and rent for 1-2 hour in a cloud.