Intel's Habana Labs Driver Quietly Drops References To The Greco AI Processor

[phoronix]

Phoronix: Intel's Habana Labs Driver Quietly Drops References To The Greco AI Processor

Announced last year at the Intel Vision conference was the Habana Labs Gaudi2 and Greco AI hardware. Since then we've seen a lot of Linux kernel driver work happen for enabling the Gaudi2 second-generation training and inference AI processor while there hasn't been anything real in the way for Greco, which was the successor to the Goya AI processor. Now references to Habana Labs Greco are being removed from the driver...

https://www.phoronix.com/news/Habana-Labs-Driver-No-Greco

[timofonic]

Is Xeon phagocytizing AI stuff? Or will be merged into their dGPUs? Why does Intel buy so much stuff to cancel or abandoni it later?

[brucethemoose]

Is Xeon phagocytizing AI stuff? Or will be merged into their dGPUs? Why does Intel buy so much stuff to cancel or abandoni it later?

Everything, including Gaudi, is supposedly being merged into Falcon Shores in 2025.

I think Intel concluded AI researchers/runners don't want a low power inference architecture and a little training architecture and a big training architecture. They don't want a shared memory server APU or Xeon CPU with an AI ASIC... They want Nvidia GPUs. So Intel canceled pretty much everything and are piling all the IP into a Nvidia GPU-like architecture with Falcon Shores.


AMD seemingly reached a similar conclusion. They took their neat, HPC oriented MI300 APU, stripped the CPU out of it, and replaced it with even more GPU for an AI SKU. Maybe the shared memory MI300 is theoretically a more efficient training platform than the GPU-only MI300 with an Epyc host CPU, but that doesn't matter if no one is going to target it.