Intel "Crescent Island" GPU Die Shows Up: 32 Xe3P Cores and 256 XMX Engines

erek

erek

8-D
2FA
Joined
Dec 19, 2005
Messages
18,872
“Additional details include support for the FP4 data format alongside FP8, as well as large register files of 1 MB per core and an L1 cache of 512 KB per core. In total, there are 256 Vector engines and 256 of the new XMX engines, each divided into groups of eight per Xe3P core. Furthermore, the Xe3P lacks graphics processing elements for outputting imagery, as the silicon is designed purely for AI inference workloads that require long context and agentic tasks. A media engine with four encoders and four decoders is also included, which could assist with multimodal workloads where video generation and understanding are essential.
19Qwqq4xWTEQoTyH_thm.jpgOExCGJ6iQNA8DxW4_thm.jpg

The accelerator is designed as a 350 W, air-cooled PCIe 5.0 x16 card, allowing it to fit into conventional server infrastructure without the need for dedicated liquid cooling. Intel designed this silicon with agents performing thousands of steps, handling a 100K context, maintaining that context, and reasoning, all while serving tokens at high speeds. We have yet to see what real-world performance looks like and what Xe3P graphics IP can do for AI processing.“

Source: https://www.techpowerup.com/354171/...ie-shows-up-32-xe3p-cores-and-256-xmx-engines
 
erek said:
“Additional details include support for the FP4 data format alongside FP8, as well as large register files of 1 MB per core and an L1 cache of 512 KB per core. In total, there are 256 Vector engines and 256 of the new XMX engines, each divided into groups of eight per Xe3P core. Furthermore, the Xe3P lacks graphics processing elements for outputting imagery, as the silicon is designed purely for AI inference workloads that require long context and agentic tasks. A media engine with four encoders and four decoders is also included, which could assist with multimodal workloads where video generation and understanding are essential.
View attachment 833323View attachment 833324

The accelerator is designed as a 350 W, air-cooled PCIe 5.0 x16 card, allowing it to fit into conventional server infrastructure without the need for dedicated liquid cooling. Intel designed this silicon with agents performing thousands of steps, handling a 100K context, maintaining that context, and reasoning, all while serving tokens at high speeds. We have yet to see what real-world performance looks like and what Xe3P graphics IP can do for AI processing.“

Source: https://www.techpowerup.com/354171/...ie-shows-up-32-xe3p-cores-and-256-xmx-engines
Click to expand...

Intel's biggest problem (like with AMD) is the CUDA moat.
It is not just about running code, you also need vendor support for said GPU and the application(s) it runs.

Most vendors I have dealt with have their appcliation tested and verified on CUDA/NVIDIA , going "off the path" is a very risky business move.

You do not want to tell you customer "Yeah, I tried to save costs on the hardware and used notn-certified/non-compliant hardware, so the service you are paying for is not working at the moment and the vendor offers zero support for me. I hope it wasn't business critical for you?"
 
littlegreenmenfromhome said:
Intel's biggest problem (like with AMD) is the CUDA moat.
Click to expand...
Only in a sense for the giant customer.

An openAI and others make their own language-compilers, generate .ptx themselve and let nvidia cuda-stack/compiler optmise from there, for them using intel-amd the second they would be cheaper-better would not be a big deal, there is so much money spent in inference that it would be worth the limited amount of trouble, it is different than for smaller customer.

Intel-AMD biggest problem is that the even if you do not rely on CUDA at all (i.e. made the AMD version well and it work well), it is often cheaper to buy Nvidia hardware and use it than get AMD hardware for free and use it and that despite Nvidia selling it to you with a 80%+ gross margin (which hide their level of giant dominance a lot), if AMD get twice as good and still give it for free, Nvidia can become back a 40% margin company and still win...

Even if you give a proven working solution, the Nvidia-CUDA one is still so much better.

The giant customer like amazon-openai-anthropic-google-meta already have their own chips/full ecosystem or working on them, they are not necessarily affraid-have issue to go without CUDA, they have the money-know how (already did it), that why all of them do use AMD hardware in some capacity, it is just vastly inferior at the moment.

And now, you have the cuda side of things, but also the supply chain-time to market in physical time advantage, Nvidia ability to reserve and use the HBM, copper, fiber and so on over the rest is getting bigger and bigger, as design get more complex-costly having bigger sales mean you are the only one able to do yearly revision and launch (like apple advantage vs AMD-Intel in the cpu space) and has you add dimensions factor, like how important networking-storage system-custom asic accelerator (upcoming sram qroq), there is 7 massive dimensions (each with a custom Nvidia chips) that goes into a full Rubin solution, Nvidia is ahead on all of them and it compound.

The lock-in software side aspect is more for the small-mid enterprise or at least not really what going on for the hyperscalers-frontier lab with the money and experience to make their affair run on everything, it is running better on Nvidia than forced to run on it, they all run fine on AMD (at least inference wise).
 
LukeTbk said:
Only in a sense for the giant customer.

An openAI and others make their own language-compilers, generate .ptx themselve and let nvidia cuda-stack/compiler optmise from there, for them using intel-amd the second they would be cheaper-better would not be a big deal, there is so much money spent in inference that it would be worth the limited amount of trouble, it is different than for smaller customer.

Intel-AMD biggest problem is that the even if you do not rely on CUDA at all (i.e. made the AMD version well and it work well), it is often cheaper to buy Nvidia hardware and use it than get AMD hardware for free and use it and that despite Nvidia selling it to you with a 80%+ gross margin (which hide their level of giant dominance a lot), if AMD get twice as good and still give it for free, Nvidia can become back a 40% margin company and still win...

Even if you give a proven working solution, the Nvidia-CUDA one is still so much better.

The giant customer like amazon-openai-anthropic-google-meta already have their own chips/full ecosystem or working on them, they are not necessarily affraid-have issue to go without CUDA, they have the money-know how (already did it), that why all of them do use AMD hardware in some capacity, it is just vastly inferior at the moment.

And now, you have the cuda side of things, but also the supply chain-time to market in physical time advantage, Nvidia ability to reserve and use the HBM, copper, fiber and so on over the rest is getting bigger and bigger, as design get more complex-costly having bigger sales mean you are the only one able to do yearly revision and launch (like apple advantage vs AMD-Intel in the cpu space) and has you add dimensions factor, like how important networking-storage system-custom asic accelerator (upcoming sram qroq), there is 7 massive dimensions (each with a custom Nvidia chips) that goes into a full Rubin solution, Nvidia is ahead on all of them and it compound.

The lock-in software side aspect is more for the small-mid enterprise or at least not really what going on for the hyperscalers-frontier lab with the money and experience to make their affair run on everything, it is running better on Nvidia than forced to run on it, they all run fine on AMD (at least inference wise).
Click to expand...
Look at vendor software.
Default is CUDA on NVIDIA hardware in +90% of use cases, especially for interference.
We have zero AMD/Intel GPU's in our datacenters, due to zero vendor support.
 
littlegreenmenfromhome said:
Look at vendor software.
Click to expand...

Yes vendor/traditional enterprise, anything turnkey, anything people can buy, hyperscaler level of internal usage is quite different and usually way more custom than that, Amazon-Google-Meta run their own stuff on their own silicon in a way we would not usually see on a vendor software

littlegreenmenfromhome said:
We have zero AMD/Intel GPU's in our datacenters, due to zero vendor support.
Click to expand...
Pretty much all the big inference players (that make giant amount of inference themselve) run customs inference stack, write their own Triton kernel, they are a bit their own vendor here, they all have some AMD gpu in their datacenter doing inference today, meta, Anthropic, Microsoft, OpenAI, google could be an exception because of their own custom TPU.... Not sure about SpaceX, but the others all made big partnership-acquistion deal with AMD in the past.

Put it this way, OpenAI could become one of the biggest if not biggest owner of AMD in a couple of years if their deals go throught, they will continue to run inference on both using their homemade programming language (Triton) that generate a custom LLVM compiler for each different target both (delivering their own ptx Nvidia-HSACO for amd), like they already do. https://github.com/triton-lang/triton, make dev-install-llvm LLVM_BUILD=amd # AMD LLVM only

The moment AMD become competitive and deliver supply it would not be a big deal for them to just augment the ratio of AMD vs Nvidia gpu in their fleet in the future, they made sure to make all the work and to be ready for this as a way to have some leverage when they negotiate price and condition with nvidia, for the biggest client CUDA is a nice moat but not a lock and it is in the sense that it is faster/cheaper to run than run vs not run on AMD.

When we are talking about company that are all making make their own custom "gpu" to run their own custom inference stack, they are not intimedated at the same level by a lack of external support to help them do anything as the regular buyer and when you run for 20 millions of dollar a day, a giant amount of work to optimise it will be worth it, the second it become possible to run it cheaper than on Nvidia.

littlegreenmenfromhome said:
Default is CUDA on NVIDIA hardware in +90% of use cases, especially for interference.
Click to expand...
More so for training no ?
 
LukeTbk said:
More so for training no ?
Click to expand...
Even more for training, but training and interference are two different types of "beasts".
And ROCm is just sucky (on top of lesser hardware) comapred to CUDA, no way around that..

Do not know about Intel GPU's (oneAPI/UXL) as zero of our vendors offers any of those servers or supported applications.
 
You must log in or register to reply here.
Back
Top