Discussion Intel current and future Lakes & Rapids thread

[dullard]

What exactly was the setup on the original AMD benchmark? Did they release a slide deck with the actual configurations used for the claims? Or could they have been using 'generic software' on Xeon and custom tuned non-public software on Turin to make it look better than it is?

See the bottom bullet point. https://cdn.mos.cms.futurecdn.net/ugUHdfYVaJWDrKmRaToUj7-970-80.png.webp

I don't know if it matters, but different Ubuntu versions and 50% more memory for AMD system. No details on software versions or settings other than using INT4.

Tom's Hardware says it probably comes down to selecting settings that don't use AMX (Advanced Matrix Extensions) math extensions that Xeon has but the EPYC Turin probably does not have.

 

[Khato]

See the bottom bullet point. https://cdn.mos.cms.futurecdn.net/ugUHdfYVaJWDrKmRaToUj7-970-80.png.webp

I don't know if it matters, but different Ubuntu versions and 50% more memory for AMD system. No details on software versions or settings other than using INT4.

Thanks. So slightly different OS versions (probably no effect, but bad/lazy practice) and different memory capacity. Pretty sure the memory capacity isn't having an effect since the Intel results appear to have been with a 512GB system?

But yeah, AMD's notes fail to mention the pertinent aspects of software configuration/optimization as well as the latency. Well, it looks like they intended to mention the latency but someone forgot to put in the number - "All testing with weights quantized to INT4 and with latency under msec."

Really don't understand why AMD feels the need to play benchmarking games. Well, except to convince investors that the grossly inflated stock price is justified.

 

[H433x0n]

See the bottom bullet point. https://cdn.mos.cms.futurecdn.net/ugUHdfYVaJWDrKmRaToUj7-970-80.png.webp

I don't know if it matters, but different Ubuntu versions and 50% more memory for AMD system. No details on software versions or settings other than using INT4.

Tom's Hardware says it probably comes down to selecting settings that don't use AMX (Advanced Matrix Extensions) math extensions that Xeon has but the EPYC Turin probably does not have.

The AI benchmarks AMD showed at their Computex keynote was extremely misleading. They ran their benchmarks with Linux kernel version 5.14. The Linux kernel 5.16 was the first to adopt AMX support so the results against EMR were extremely misleading to say the least.

This is probably why Intel felt compelled to post this since a lot of their sales in data center are due to their AMX extensions and the perception that Turin significantly outperformed it could have hurt their bottom line with those customers.

 

[lightisgood]

See the bottom bullet point. https://cdn.mos.cms.futurecdn.net/ugUHdfYVaJWDrKmRaToUj7-970-80.png.webp

I don't know if it matters, but different Ubuntu versions and 50% more memory for AMD system. No details on software versions or settings other than using INT4.

Tom's Hardware says it probably comes down to selecting settings that don't use AMX (Advanced Matrix Extensions) math extensions that Xeon has but the EPYC Turin probably does not have.

In the first place, AMX provides 8 times the peak perfomance of AVX-512 VNNI.
AMD made up EPYC benchmark score with 2x core count and 1/2 data type (= 2x throughput), of course this is fair.
And then both of them have close benchmark scores through some bottleneck.

P.S.
Anyway, thanks to AMX, EMR is competitive product in AI server market and GNR should beat Turin.

 

[Nothingness]

In the first place, AMX provides 8 times the peak perfomance of AVX-512 VNNI.
AMD made up EPYC benchmark score with 2x core count and 1/2 data type (= 2x throughput), of course this is fair.
And then both of them have close benchmark scores through some bottleneck.

P.S.
Anyway, thanks to AMX, EMR is competitive product in AI server market and GNR should beat Turin.

Sorry if that's already well known. I thought there was one AMX unit per chip, not per CPU core. The picture talks about AMX ops/core. Is that per AMX core or is there really one AMX in each CPU core?

 

[Abwx]

In the first place, AMX provides 8 times the peak perfomance of AVX-512 VNNI.
AMD made up EPYC benchmark score with 2x core count and 1/2 data type (= 2x throughput), of course this is fair.

I read 128C vs 2 x 64C, nowhere i see 2 x the core count, they would state 256C for Turin if that was the case.

Reaktion auf AMD-Vorstellung: Intel korrigiert AMDs Turin-vs.-Emerald-Rapids-Benchmarks

Eine Benchmarkkorrektur wenige Tage nach Veröffentlichung? Das gab es erst bei AMD vs. Nvidia, nun folgt Teil 2: AMD vs. Intel.

www.computerbase.de

 

[lightisgood]

Is that per AMX core or is there really one AMX in each CPU core?

Each Xeon P-core has AMX accelerator as TMUL (tile matrix multiply unit).

Please see also this video.

Intel® Xeon® Scalable processor Max Series

Technical overview of the Intel® Xeon® Scalable processor Max Series.

www.intel.com

We know that Intel can easily enhance Xeon's AI performance because AMX was carefully designed to secure performance headroom.
It is simply achieved by additional TMUL.
I'm looking forward to see 2nd gen AMX.

 

[dullard]

I read 128C vs 2 x 64C, nowhere i see 2 x the core count, they would state 256C for Turin if that was the case.

I read on that graph

• 2P AMD EPYC "Turin" 128C
• 2P Intel Xeon 8592+ 64C

On their detail slide, I read:

• 2P 5th Gen EPYC 128-C pre-production silicon (128C/256T)
• 2P Xeon Platinum 8952+ (64C/128T) production system

https://cdn.mos.cms.futurecdn.net/ugUHdfYVaJWDrKmRaToUj7-970-80.png.webp

Is there terminology where 2P stands for one processor for AMD but two processors for Intel? In other words, why would they double 128C to 256C for Turin, but not double 64C to 128C for Xeon?

 

[Markfw]

I read on that graph

• 2P AMD EPYC "Turin" 128C
• 2P Intel Xeon 8592+ 64C

On their detail slide, I read:

• 2P 5th Gen EPYC 128-C pre-production silicon (128C/256T)
• 2P Xeon Platinum 8952+ (64C/128T) production system

https://cdn.mos.cms.futurecdn.net/ugUHdfYVaJWDrKmRaToUj7-970-80.png.webp

Is there terminology where 2P stands for one processor for AMD but two processors for Intel? In other words, why would they double 128C to 256C for Turin, but not double 64C to 128C for Xeon?

I will wait for 3rd party benchmarks before I comment further, as each one of these two setups appear to be optimized for that companies hardware.

 

[Abwx]

I read on that graph

• 2P AMD EPYC "Turin" 128C
• 2P Intel Xeon 8592+ 64C

On their detail slide, I read:

• 2P 5th Gen EPYC 128-C pre-production silicon (128C/256T)
• 2P Xeon Platinum 8952+ (64C/128T) production system

https://cdn.mos.cms.futurecdn.net/ugUHdfYVaJWDrKmRaToUj7-970-80.png.webp

Is there terminology where 2P stands for one processor for AMD but two processors for Intel? In other words, why would they double 128C to 256C for Turin, but not double 64C to 128C for Xeon?

Well, if that s 2 x 128 vs 2x 64 the comparison doesnt make sense even in a marketing POV.

They would have been better to compare 2 x 64C systems given that a 64C Turin retain 63% of the 128C perf at same TDPs, the picture would had been more honnest and even more advertisable marketing wise since 2.4x/3.7x/1.57x at same core count is more impressive than 3.9x/5.9x/2.5x at 2 x the core count.

 

[H433x0n]

as each one of these two setups appear to be optimized for that companies hardware.

That’s categorically untrue. This is the equivalent of Intel touting RPL getting big wins over Zen 4 in floating point benchmarks but they hid the fact that they had AVX512 disabled in the endnotes.

Really don't understand why AMD feels the need to play benchmarking games.

Perhaps the loss of Robert Hallock is having a real effect. This latest keynote had 2 really egregious examples. They also had the Zen 3 XT gaming benchmarks where they used an RX 6600 GPU to test against a 13700K equipped with DDR4 to make it appear that the 5900 XT is on par with RPL in gaming performance.

 

[igor_kavinski]

Well, if that s 2 x 128 vs 2x 64 the comparison doesnt make sense even in a marketing POV.

I suspect that AMD wanted to show that the Turin 256 core setup can be had for the same or lower price as the Xeon 128C server but somehow failed to do that. The Turin server may also be consuming power equal to or less than the Xeon server but that's also not mentioned. Have to call it here. Pretty dumb comparison. Never a good idea to go against your competitor without doing your homework first.

 

[igor_kavinski]

Pity AMD didn't read this review: https://www.tomshardware.com/pc-com...8592-review-64-cores-320mb-of-l3-and-350w-tdp

Plenty of benchmarks there where the Xeon gets spanked by Bergamo and Genoa but AMD just HAD to choose the one benchmark where the Xeon actually performs well. AMDumb indeed.

 

[DrMrLordX]

Pity AMD didn't read this review: https://www.tomshardware.com/pc-com...8592-review-64-cores-320mb-of-l3-and-350w-tdp

Plenty of benchmarks there where the Xeon gets spanked by Bergamo and Genoa but AMD just HAD to choose the one benchmark where the Xeon actually performs well. AMDumb indeed.

Because AI.

 

[coercitiv]

Perhaps the loss of Robert Hallock is having a real effect. This latest keynote had 2 really egregious examples. They also had the Zen 3 XT gaming benchmarks where they used an RX 6600 GPU to test against a 13700K equipped with DDR4 to make it appear that the 5900 XT is on par with RPL in gaming performance.

Recent AMDs marketing slides are in shambles, they contain intentional and unintentional mistakes that make them close to worthless for any serious analysis.

They spent years to build trust, especially in their CPU marketing efforts, and now they caught the AI fever and will waste all that trust capital on worthless wins. I used to trust their CPU slides because they were consistently proven correct (as much as marketing slides can be ofc), now I can't even be sure about their end notes.

On the subject of 256 cores vs. 128 cores, I think it would be fair game to highlight such an advantage if the platform allows this config for one vendor and not the other, but it would also need further details about cost, power, etc. That being said, once you start highlighting the differences between the platforms, then the lack of AMX extensions becomes an even bigger question mark.

On a lighter note, as @Nothingness already noted, Intel complaining about the competition not being fair with software optimizations is a sight to behold.

 

[yuri69]

AMD is an entity whose aim is to generate profit. That means its marketing dept. is meant to do anything - in legal boundaries - to generate profit. So its marketing slides are aiming to maximize the generated profit.

Odd comparisons, mixed up footnotes, omissions, etc. - things like that matter for tech geeks not the target mass market which generates the profit. In other words, spending $$$ on accurate marketing is might not be the goal.

 

[igor_kavinski]

In other words, spending $$$ on accurate marketing is might not be the goal.

They need to fire EVERYONE in marketing and hire competent people with technical knowledge. Their incompetency is ruining the hard work of everyone in their engineering departments. At the very least, get someone competent to review the marketing materials before disseminating them to the press and the public.

 

[dullard]

Plenty of benchmarks there where the Xeon gets spanked by Bergamo and Genoa but AMD just HAD to choose the one benchmark where the Xeon actually performs well.

That is the thing, AMD server chips are great performers in general terms. Intel knew that they had nothing to compete with AMD directly due to a much larger node size. So, Intel added a ton of accelerators for various niche uses. In those niche uses, Xeons either win or are at least quite competitive. It all boils down to the specific use case to know which chip should be chosen. This AI use case is one of those niches where the Xeons have the edge.

 

[Hans Gruber]

That is the thing, AMD server chips are great performers in general terms. Intel knew that they had nothing to compete with AMD directly due to a much larger node size. So, Intel added a ton of accelerators for various niche uses. In those niche uses, Xeons either win or are at least quite competitive. It all boils down to the specific use case to know which chip should be chosen. This AI use case is one of those niches where the Xeons have the edge.

The node advantage AMD has had for years will be gone by the end of the year. All the hopes and dreams for a brighter tomorrow may be gone. Intel said they were going to put AMD in their rear view mirror. Will we see AMD fire sale prices again? One can only hope that is the case.

 

[igor_kavinski]

Will we see AMD fire sale prices again? One can only hope that is the case.

Maybe but I hope it will make AMD compete harder with better products. They are a leaner organization than Intel. While Intel can afford to throw money, resources and people at the problem, AMD chooses the shortest path to success.

 

[Khato]

That is the thing, AMD server chips are great performers in general terms. Intel knew that they had nothing to compete with AMD directly due to a much larger node size. So, Intel added a ton of accelerators for various niche uses. In those niche uses, Xeons either win or are at least quite competitive. It all boils down to the specific use case to know which chip should be chosen. This AI use case is one of those niches where the Xeons have the edge.

Yup. One major reason why Intel's server sales haven't dropped as precipitously as many forum dwellers predicted based on general performance/power benchmarks are those accelerators. While the accelerators are 'niche' to general computing, some are quite common use cases for servers. And unlike AMD, Intel has the software presence to get those accelerators supported.

Add the competent performance/power of SRF with those accelerators and the tables begin to turn. With the performance improvements Intel is stating for Skymont it's easy to see CWF offering 1.3-1.4x the performance/power of SRF. Though we'll of course have to see where Turin ends up in actual benchmarks once it shows up to market - the lack of any Genoa/Bergamo to Turin performance comparisons in their presentation stands out.

 

[Hitman928]

Yup. One major reason why Intel's server sales haven't dropped as precipitously as many forum dwellers predicted based on general performance/power benchmarks are those accelerators. While the accelerators are 'niche' to general computing, some are quite common use cases for servers. And unlike AMD, Intel has the software presence to get those accelerators supported.

Add the competent performance/power of SRF with those accelerators and the tables begin to turn. With the performance improvements Intel is stating for Skymont it's easy to see CWF offering 1.3-1.4x the performance/power of SRF. Though we'll of course have to see where Turin ends up in actual benchmarks once it shows up to market - the lack of any Genoa/Bergamo to Turin performance comparisons in their presentation stands out.

Intel servers has been hanging on traditional enterprise sales where they are undercutting on price to the point where they don’t make any money. AMD is over 50% in cloud and enterprise is finally starting to open up too. The next releases from Intel will certainly help them but I’m not sure it will be enough to stop AMD from taking market share.

 

[jpiniero]

Intel servers has been hanging on traditional enterprise sales where they are undercutting on price to the point where they don’t make any money.

Depends on how you look at it

But they obviously cannot do that kind of price undercutting on any node past 10 nm.

There is also of course the apathy angle too or won't buy AMD for whatever reason.

 

[Tup3x]

Perhaps the loss of Robert Hallock is having a real effect. This latest keynote had 2 really egregious examples. They also had the Zen 3 XT gaming benchmarks where they used an RX 6600 GPU to test against a 13700K equipped with DDR4 to make it appear that the 5900 XT is on par with RPL in gaming performance.

To be honest, that was so misleading that it should be illegal to mislead people like that.

 

[Khato]

Intel servers has been hanging on traditional enterprise sales where they are undercutting on price to the point where they don’t make any money. AMD is over 50% in cloud and enterprise is finally starting to open up too. The next releases from Intel will certainly help them but I’m not sure it will be enough to stop AMD from taking market share.

No disagreement that Intel has been selling pretty much at cost in order to keep the fab utilization up and minimize market share loss. But keep in mind that such is in large part a result of their products up until SRF still being on the Intel 7 node. Even if Intel sells a 64 core Emerald Rapids at the same price as a 64 core Genoa that's 2x750mm^2 dies versus 8x72mm^2 compute + 400mm^2 IO. The massive die size on Intel's side is financially viable if it could be sold around the $10k suggested retail price, but that's where the 128C Bergamo and 96C Genoa reside. So instead Intel's flagship has to sell closer to the next tier down around the $5k mark. (Note that these are by no means accurate absolute numbers, but should be reasonably close in relation to one another.)

One question though is what's the source on AMD being over 50% in cloud?