Intel SC22 Intel Falcon Shores overview
This is a Friday afternoon release if I’ve ever seen one. In a news brief today, Intel has fundamentally changed its HPC efforts until at least 2026. Or in short, Intel has announced that the current Xeon Max, based on Sapphire Rapids with HBM, and its Ponte Vecchio OAM GPU are practically the last of this old HPC platform architecture “before XPU” (see update).
Intel announces that it is 3 years behind AMD and NVIDIA in XPU HPC
Intel didn’t say the above or include it in a new roadmap, instead analyzing the announcements:
Rialto Bridge being discontinued The next Max series XPU/GPU will be Falcon Shores in 2025
This effectively means that Intel won’t have a new GPU offering between 2023 and 2025.
Update: Intel contacted us to clarify their roadmap. Falcon Shores 2025 will be a GPU architecture, not with CPU cores. Intel told us the vision for Falcon Shores is still XPU, but 2025 Falcon Shores is GPU focused. The “traditional” part is a GPU-only offering, Falcon Shores will offer different IP blocks as part of an overall XPU offering, but that won’t be the 2025 generation.
In 2025 we’re getting Falcon Shores, which should co-package CPU cores, GPU cores, memory and other IP, possibly other accelerators and even Lightbender silicon photonics. Intel has also revealed plans for CPU tiles only with Falcon Shores co-packaged memory.
Intel SC22 Intel Falcon Shores XPU 2
Looking at what this means for the traditional HPC architecture of CPUs sitting in sockets attached to accelerators like GPUs, this model is effectively being phased out, except at Intel.
The AMD Instinct MI300 (2023) packs the CPU, GPU and memory together.
dr Lisa Su AMD ISSCC 2023 AMD Instinct MI300
NVIDIA Grace Hopper (2023) will package CPU, GPU and memory together and have the ability to have CPU only packages.
NVIDIA Computex 2022 HGX Grace HGX Grace Hopper
Intel’s next-gen HPC part Falcon Shores (2025) should package CPU, GPU and memory together.
Intel SC22 Intel Falcon Shores overview
Maybe still, but not at Intel until 2025.
The subtext of today’s announcement is that Intel and its customers see the need for this co-packaging architecture, to the point that Rialto Bridge hasn’t been established as a big winner. Intel is instead focusing on GPU-only Falcon Shores for 2025. This means that by 2026, NVIDIA and AMD will have the lead in this next-gen architecture.
For those wondering about the benefits of co-packaging memory, CPU, and GPU, it’s in many ways the same as what’s happening with the Apple M1/M2.
Apple M2 Pro package and 32GB storage
Moving data from CPU memory to GPU memory and vice versa may require duplicating data, resulting in more wasted memory, but performance is even more important. Last week we featured AMD Talks Stacking Compute and DRAM at ISSCC 2023. The power used to move data between sockets and even between CPU sockets and DIMMs while increasing speed is a big challenge.
dr Lisa Su AMD ISSCC 2023 Energy Efficiency of IO Interconnect
As a result, co-packaging the CPU, GPU, accelerators, and memory using more advanced methods is seen as a way to increase connection speeds at a different performance curve than traditional separate components.
dr Lisa Su AMD ISSCC 2023 bits per joule
In large HPC systems, a huge amount of power is consumed simply to move data within nodes and externally. The reason AMD, NVIDIA, and now Intel are saying next-gen HPC GPU architectures will have co-packaged components is to minimize this “wasted” power consumption. This in turn allows more energy to be used for calculations instead of shifting data.
last words
The HPC market is sizeable, but it is only a part of the overall computer market. We still assume that there will be PCIe and OAM GPUs etc. in the years to come. It just seems like this trend is accelerating for the HPC space to the point where Intel will skip a generation. That saves development resources, but it also means that Intel will use Ponte Vecchio to compete against the NVIDIA H100, NVIDIA Grace-Hopper and AMD Instinct MI300 over the next two years until Falcon Shores is hopefully ready in 2025. This announcement also means it plans its 2025 Falcon Shores XPU to be truly a GPU with the ability to have custom IP blocks in the future.
Intel Aurora Sapphire Rapids and Ponte Vecchio HPE Cray EX Node ISC 2022 2x SPR
Additionally, Intel announced it will discontinue Lancaster Sound, its next-generation Flex-series GPU, and make the next-generation Rekord Melville Sound.
