Intel Talks Up Granite Rapids, Falcon Shores & Their Open Software Stack @ SC23

With the SC23 Super Computing conference kicking off today in Denver, Intel has just lifted the embargo on a number of disclosures.

Intel isn't announcing any new products for SC23 but is talking up some of their existing CPU, GPU, and AI products as well as continuing to tease their next-generation wares and highlight their wonderful open-source software work.

During the Intel press briefings, Intel talked briefly on the Aurora supercomputer that is still being brought up... It's still not fully deployed yet. Aurora is still being scaled up for that 2+ ExaFLOPS supercomputer that is still being brought up following a series of delays. The system is fully installed but taking everything online and scaling well is a matter still being carried out. More details when the latest TOP500 list is published.

Intel did continue to highlight the great success of their Data Center GPU Max series hardware.

For scientific workloads the Data Center GPU Max series is working out very well. While I haven't tested any Data Center GPU Max hardware, unfortunately, even when it comes to their consumer Arc Graphics hardware they've been performing very well for GPU compute atop a fully open-source stack.

Intel also is using SC23 to reinforce their great Gaudi 2 AI accelerator hardware as what they claim to be "the only viable alternative to (NVIDIA's) H100." Gaudi 2 is certainly looking good in the performance per dollar department too against NVIDIA's wares.

Gaudi 3 was also teased as their next-generation AI accelerator coming in 2024 and will feature more and faster HBM, greater BF16 capabilities, move to a 5nm process, and sport more integrated networking capabilities.

Intel continues to support the Habana Labs / Gaudi hardware with their upstream Linux kernel driver and open-source SynapseAI components in user-space. It's with the Falcon Shores APU/XPU in 2025 where the Habana Labs and Intel product lines will converge for really interesting possibilities.

On the Xeon Scalable side, 5th Gen Xeon Scalable "Emerald Rapids" processors are launching 14 December as previously disclosed. Emerald Rapids is an incremental improvement over Sapphire Rapids with maintaining platform compatibility but delivering better performance and power efficiency. It will be with Granite Rapids next year where things should get more interesting with more cores, higher memory bandwidth, continued AI improvements, and the first P-core Xeon based on Intel 3.

In the meantime Intel was showcasing their Xeon Max 9480 Sapphire Rapids HBM2e performance. Not too fair, they compared their flagship Sapphire Rapids HBM2e part against the EPYC 9654 for various HPC benchmarks... Much more appropriate would have been the existing AMD EPYC 9684X Genoa-X processor that is similar in its roles to Xeon Max. With Intel's benchmarks the Xeon Max 9480 was just 19% faster overall than the EPYC 9654 Genoa processor. You can see some of my Xeon Max benchmarks on Phoronix for comparisons to Genoa and Genoa-X. There are certainly some workloads where Xeon Max is very capable and competitive while it will be all the more interesting to see second-generation Xeon Max and hopefully allowing a greater HBM2e capacity per core to open up additional possibilities.

For Emerald Rapids, Intel is promoting 1.2~1.4x performance improvements over Sapphire Rapids while at the same power level. It will sure be fun to benchmark Emerald Rapids.

With Granite Rapids is where Intel looks to become ultra competitive with 2~3x performance improvements in different areas. I'm already salivating over the Intel Granite Rapids benchmarking potential.

The presentation concluded with Intel talking up their open-source software stack. Their open-source software stack needs no introductions at Phoronix and continue covering it the most out of any publication on the web... Intel annually makes countless open-source contributions to benefit not only their own hardware but often working equally as well on other platforms. Their oneAPI effort is wonderful (or now the Unified Acceleration Foundation as UXL) and beyond that their incredible investments into the Linux kernel, GCC, LLVM/Clang, and other open-source projects. Intel is unmatched when it comes to their open-source ecosystem contributions and especially with efforts around performance optimizations and early hardware enablement / bringing up new CPU ISA features.