News Intel launches three new Xeon 6 P-Core CPUs, will debut in Nvidia DGX B300 AI systems

[Admin]

Intel has unveiled three new Xeon 6 CPUs, set to debut in Nvidia's DGX B300.

Intel launches three new Xeon 6 P-Core CPUs, will debut in Nvidia DGX B300 AI systems : Read more

 

[Obsidian Blue]

This doesn’t seem particularly impressive, or am I missing something?

The EPYC 9575F boosts many more than 8 cores beyond 5GHz under sustained load - so surely the EPYC CPU’s would still perform better as a GPU server CPU than these new Intel xeons?

Adding v expensive MRDIMM memory may help the Xeon but I doubt it would bridge the performance gap.

 

[thestryker]

This doesn’t seem particularly impressive, or am I missing something?

The EPYC 9575F boosts many more than 8 cores beyond 5GHz under sustained load - so surely the EPYC CPU’s would still perform better as a GPU server CPU than these new Intel xeons?

Adding v expensive MRDIMM memory may help the Xeon but I doubt it would bridge the performance gap.

While it's not in the text of the article the slide says PCT enables customizable turbo frequencies which I'd guess are not limited to the CPU's stock behavior.

 

[JRStern]

Did I miss where it says if these are 18a or made by TSM?

 

[thestryker]

Did I miss where it says if these are 18a or made by TSM?

No change from the rest of the line: Intel 3

 

[TerryLaze]

This doesn’t seem particularly impressive, or am I missing something?

The EPYC 9575F boosts many more than 8 cores beyond 5GHz under sustained load - so surely the EPYC CPU’s would still perform better as a GPU server CPU than these new Intel xeons?

Adding v expensive MRDIMM memory may help the Xeon but I doubt it would bridge the performance gap.

Just as with the x3d CPUs, core counts and clocks aren't everything.
I have no idea what these xeons provide, if anything, just saying that clocks are pretty low on the priority list for servers.
Maybe it's just because of the pci lanes and ram speed...

Intel's Xeon 6 CPUs include up to 128 P-cores per CPU and 20% more PCIe lanes than previous-generation Xeon processors, with up to 192 PCIe lanes per 2S server. Intel also claims Xeon 6 offers 30% faster memory speeds compared to the competition (specifically the latest AMD EPYC processors), thanks to Multiplexed Rank DIMMs (MRDIMMs) and Compute Express Link, and up to 2.3x higher memory bandwidth compared to the previous generation.

 

[Obsidian Blue]

Just as with the x3d CPUs, core counts and clocks aren't everything.
I have no idea what these xeons provide, if anything, just saying that clocks are pretty low on the priority list for servers.
Maybe it's just because of the pci lanes and ram speed...

I'd beg to differ with regards to clock speeds on cores within GPU servers. It's been shown in lab tests that the 9575F can deliver 20% more tokens/sec in a 8xH100 GPU server (running Llama 3.1 - 70B @FP8) compared to an identical server with Intel Xeon 8592+. The 9575F was specifically designed to deliver high boost to a few cores since AI workloads tend not to be sustained and are more 'bursty' in their CPU demands - however the CPU's have to keep feeding the GPU's so when they're called on they need to complete their orchestration tasks as fast as possible.

It will be interesting to see if Phoronix do some independent testing and compare the new Intel CPU's vs EPYC 9575F for GPU workloads

 

[TerryLaze]

I'd beg to differ with regards to clock speeds on cores within GPU servers. It's been shown in lab tests that the 9575F can deliver 20% more tokens/sec in a 8xH100 GPU server (running Llama 3.1 - 70B @FP8)

Is this what every server is running?! Like the one and only thing?!

 

[Obsidian Blue]

Is this what every server is running?! Like the one and only thing?!

Sorry, I'm not sure if I've understood the question. The testing was completed on individual servers and they were just running a Llama 3.1-70B inference benchmark.

There was also a publicly released test that ran a Llama 3.1-8B Training benchmark on 2 8xH100 GPU systems (9575F vs Xeon 8592+) and it shows a +15% training performance improvement with EPYC

 

[TerryLaze]

Sorry, I'm not sure if I've understood the question. The testing was completed on individual servers and they were just running a Llama 3.1-70B inference benchmark.

There was also a publicly released test that ran a Llama 3.1-8B Training benchmark on 2 8xH100 GPU systems (9575F vs Xeon 8592+) and it shows a +15% training performance improvement with EPYC

My question is that I have no idea how representative this is, or if the nvidia servers are made for this one benchmark.
I'm just saying, maybe they bought the intel CPUs because they are running something else.