Discussion Intel Crescent Island thread

[marees]

Questions
Questions

https://twitter.com/x/status/1978398034112266689

Intel Announces "Crescent Island" Inference-Optimized Xe3P Graphics Card With 160GB vRAM - Phoronix

www.phoronix.com

Intel to Expand AI Accelerator Portfolio with New GPU

newsroom.intel.com

 

[regen1]

It was there on the roadmap. A major objective of their "Inference Workstation" aka Battlematrix(Xeon platform + 4* B60 Dual GPUs) was to advance their software ecosystem before their future stuff releases(Inference GPUs, Jaguar Shores,.....).

 

[regen1]

It has a little(160GB vs 128GB) memory advantage vs Rubin CPX but bandwidth would likely be lower, may be lower TOPs(compute) as well.
This thing would have to undercut Rubin CPX in pricing well enough to compete if it were to release in the same timeline. Again lots of speculation, too far away.

Talk shows Intel still well committed to UALink, UEC contrary to some doubts raised by association with Nvidia deal involving NVLink

An interesting slide, heterogeneous mix of Gaudi 3 and B200

 

[DavidC1]

Crescent Island seems to have 40 blocks on the outside, where the I/O usually is. They are probably memory controllers. I count 40 of them, and that fits 160GB nicely. If each are 64-bits, at 8533MT/s, it would be 2.7TB/s bandwidth, which is quite below competition. The blocks don't seem too small, maybe it can support 128 or 256-bits. That would be 5.4 or 10.8TB/s.

There are 32 compute blocks, with 16 on top and 16 on bottom separated by likely some router. Inside each block, I see 12(or 14) sub-blocks. That could indicate 384 Xe cores. If you scale ~50mm2 for 12 Xe3 in Pantherlake, it would be 1600mm2. Maybe it's doable with 2-Tile Foveros.