The FASTEST CPU on the Planet... AGAIN

[AdamFromLTT]

Fun Fact: In a 2-socket Epyc 9654 server, you need 4 threads pegged at 100% to get 1% CPU utilization.

AMD's new SP5 socket, and HUGE new Epyc Genoa CPUs are crazy! Using Supermicro's servers, we set a bunch of World Records on a bunch of benchmarks. Intel is gonna be sweating, unless their upcoming Sapphire Rapids platform can magically compete. Intel's Raptor Lake CPUs might win against Ryzen in the mainstream, it's a whole different story in the datacenter.

[Results45]

I wanna see how many FPS 384 threads can crank out in Crysis......please do a follow-up SwiftShader test

[AnonymousGuy]

If you want to know how Intel is going to launch Sapphire, it's going to be a focus on the accelerators built in since the core counts aren't going to match.  Me: "good luck with that".

[nasir91]

Can anyone link me to the rankings website for SuperPi?

[Luscious]

I think a more interesting use case would be a workstation build using a pair of higher clocked 9174F's. That's if Linus still gives a F about showing how these would be deployed in a more user friendly non-server configuration. Windows 11 still has support for dual CPU but you need the "for workstation" version of it - something that also should be made clear for those considering going down this path for a build. It's also quite trivial how many 4U server chassis are just turned sideways to convert into a tower workstation by adding feet and a different cover that can then sit under your desk.

[Xandercusa]

It seems like the y-cruncher submission was removed from HWBot, and I’m surprised the GPUpi submission hasn’t been taken down yet since it doesn’t have CPU-Z or GPU-Z open in the screenshot. 

[AdamFromLTT]

3 minutes ago, Xandercusa said:

It seems like the y-cruncher submission was removed from HWBot, and I’m surprised the GPUpi submission hasn’t been taken down yet since it doesn’t have CPU-Z or GPU-Z open in the screenshot. 

Guess we'll have to beat em again.

[tim0901]

4 hours ago, AnonymousGuy said:

If you want to know how Intel is going to launch Sapphire, it's going to be a focus on the accelerators built in since the core counts aren't going to match.  Me: "good luck with that".

I don't think they even need your good luck. Despite years of AMD topping the performance charts, Intel still absolutely dominates the server space with over 85% marketshare.

 

It wouldn't surprise me if driver quality had something to do with it - I remember LTT encountering some big problems when trying to do virtualisation on AMD server CPUs at one point (probably a couple of years ago now) that they were unable to fix, eventually forcing them over to an Intel-based system, and that's exactly the kind of BS that these companies wouldn't want to deal with either. They'll take the tried-and-trusted Intel any day.

 

I've also heard rumors of supply issues in this space as well from AMD, although I've no idea exactly how true that is. It would make sense though given their lack of supply in other areas as well eg GPUs.

[dogwitch]

6 hours ago, tim0901 said:

I don't think they even need your good luck. Despite years of AMD topping the performance charts, Intel still absolutely dominates the server space with over 85% marketshare.

 

It wouldn't surprise me if driver quality had something to do with it - I remember LTT encountering some big problems when trying to do virtualisation on AMD server CPUs at one point (probably a couple of years ago now) that they were unable to fix, eventually forcing them over to an Intel-based system, and that's exactly the kind of BS that these companies wouldn't want to deal with either. They'll take the tried-and-trusted Intel any day.

 

I've also heard rumors of supply issues in this space as well from AMD, although I've no idea exactly how true that is. It would make sense though given their lack of supply in other areas as well eg GPUs.

2 largest super computers in the usa is all amd this time. from cpu to gpus.

 

Intel’s market share in April 2020 was 78.28%, while AMD’s was 21.71%.

i forget where i read it but its over 25% now.

check out servethehome.

 

[seanyfresh]

On 12/14/2022 at 4:47 PM, tim0901 said:

I don't think they even need your good luck. Despite years of AMD topping the performance charts, Intel still absolutely dominates the server space with over 85% marketshare.

 

It wouldn't surprise me if driver quality had something to do with it - I remember LTT encountering some big problems when trying to do virtualisation on AMD server CPUs at one point (probably a couple of years ago now) that they were unable to fix, eventually forcing them over to an Intel-based system, and that's exactly the kind of BS that these companies wouldn't want to deal with either. They'll take the tried-and-trusted Intel any day.

 

I've also heard rumors of supply issues in this space as well from AMD, although I've no idea exactly how true that is. It would make sense though given their lack of supply in other areas as well eg GPUs.

I work in infrastructure sales engineering in the datacenter. The most expensive thing in an on-premise datacenter these days is software licensing. This is overwhelmingly metered by counting physical cores. This why you will almost never see systems with 192 physical cores in a virtualization host. Resiliency is king and redundancy for that would require another host with another 192 cores and your software license costs just doubled. Why two hosts at 192 cores each (384 total) instead of 7 hosts with 32 cores each (224 total)? Another issue is actually thermals in 1U or 2U chassis. CPU's get more powerful but cooling, at least in the mainstream DC's, has not come close to keeping up.

The overwhelming majority of my sales in the last two years have come from platforms running the Intel 6326, 6346, and 6254. Speed and core count is great and fits for the vast majority of virtualized workloads, including MSSQL.

Long story short: fewer, faster cores is the way to cost optimize an on-premise datacenter, because while hardware is expensive, software is expensive-r.

But I can also say without any question the Sapphire Rapids strategy of "lets make the CPU speeds slower than the previous generation and market accelerators" is going to supremely backfire. AMD 6124 is going to be my default CPU for the foreseeable future once it becomes adopted and supported by the hypervisors.

[tim0901]

11 hours ago, seanyfresh said:

But I can also say without any question the Sapphire Rapids strategy of "lets make the CPU speeds slower than the previous generation and market accelerators" is going to supremely backfire. AMD 6124 is going to be my default CPU for the foreseeable future once it becomes adopted and supported by the hypervisors.

What makes you say that raw CPU speeds will be slower this gen? I can't see how having a greater number of more modern cores at similar clocks is going to result in worse performance than current parts. That just sounds absolutely ludicrous. 

 

And even if you don't want the high core counts, the switch to the same golden cove cores found in Alder Lake should still provide a substantial uplift over Ice Lake.

 

Sure the accelerators are definitely the big thing this gen, but I can't see for a minute how the raw CPU be power will be getting downgraded here.

 

And why knock the accelerators before you try them?

[danwat1234]

I have a question. Can EPYC CPUs be stably undervolted for more efficiency? Can The super duper Xeons whatever they r called be undervolted as well? Or r these 'locked' and protected against PlunderVolt yuck?

[Lorenzo Neri]

Hello to everybody,

I'm a researcher; I purchased one Supermicro Hyper AS-2025HS-TNR with two AMD 9654 and 1.5TB of RAM. It cost me 30k€.

It was purchased for research purposes. I'm developing a parallel code that computes plasma behaviour.

 

It is obvious that Supermicro and AMD spend money on YouTubers who tell us they are at the edge of technology.

However, I trust Linus's videos because he never hides the difficulties of configuring such many cores.

 

Since the beginning of my machine configuration, I discovered that Supermicro is not fully compliant with AMD EPYC 9004.

I tried to follow the HPC optimization guide for Epyc 9004 processors, here attached, and I discovered that many important sets are missing in the Supermicro Bios. The truth is that it is difficult to configure so many cores and impossible to configure them to their maximum performance.

In detail, from pages 17 and 18 of the guide:

4-link xGMI max speed: 32Gbps  => Supermicro Bios enable only up to 22Gbps => -31% of performance.

BoostFmaxEn: Manual => Supermicro Bios don't have this parameter

BoostFmax: 3700 => Supermicro Bios don't have this parameter

HSMP Support: Enable => Supermicro Bios don't have this parameter

Workload Profile: HPC Optimized => Supermicro Bios don't have this parameter

CPU Speculative Store Modes: More Speculative => Supermicro Bios don't have this parameter

DRAM Survive Warm Reset: Enable => Supermicro Bios don't have this parameter

 

I opened a ticket with Supermicro, and they confirmed that these settings are not in the BIOS and can be asked as a "CUSTOM Bios configuration" through my Supplier representative. I have asked, and they replied that Supermicro is not ready to provide a BIOS with this setting.

 

Without the Boost functionality but setting the governor mode to "performance", the CPU frequency seems to be kept as high as possible, but I'm not sure if I have obtained the maximum CPU frequency. The limited xGMI speed alone reduces the performance by 31%. Furthermore, I'm not able to quantify the effect of other settings.

 

I'm writing here because I have followed Linus's YouTube channel since the beginning and trust in his professionalism.

I have four questions for Linus and Linus staff, in order of priority:

1) Can you help me to obtain this machine's maximum performance? Science will tank you.

2) Can you ask Supermicro to satisfy their declared full compatibility?

3) Can you inform AMD that the seller of the platform they chose to launch, Epyc 9004, is not fully compatible with their cpus?

4) Would you like to make a video with the real-life difficulties generated by this rush in launching the new versions? There is only the challenge to promising Eden. The role of the Free Reviewer is to tell the whole truth. 

 

If you need, I can provide more details about my attempts.

 

Thank you in advance for considering my message.

With Best Regards,

Lorenzo

 

 

58002-epyc-9004-tg-hpc.pdf