News Intel redefines AI strategy — Jaguar Shores to be rack-level design with focus on silicon photonics

Also, our open x86.
Open on what ?
Only AMD and VIA have a license and this date back when IBM forced them to do it in 1982.

Intel can provide products for inference but they are too much behind nVidia and AMD on AI training, they can only be third option if they can supply in quantities. It's unlikely for nVidia and AMD to use Intel Foundry for cutting edge technologies, the market known the story of Apple and Samsung Foundry.
 
Michelle Johnston Holthaus said:
What we are seeing is that customers do love the x86 ecosystem, the software around it. If they can build out an AI infrastructure with x86, they are very interested in doing that.
Uh, that didn't work so well for Xeon Phi. I'd caution Intel against going back to x86, even though they perceive that as a strength.
 
Open on what ?
Only AMD and VIA have a license and this date back when IBM forced them to do it in 1982.
I don't know, but maybe referring to this?

Intel can provide products for inference but they are too much behind nVidia and AMD on AI training,
Gaudi 3 doesn't seem so bad. I wouldn't say their strategy isn't salvageable.

What they definitely need to do is avoid the mistake AMD often makes of trying to beat Nvidia at its own game. Nvidia is currently tied to CUDA, which has been a strength, but also slows down the pace at which they can evolve. AMD foolishly decided to make a CUDA-compatible API called HIP. Intel has been going in a similar direction with oneAPI, and while I like oneAPI, I can see that it's not the path to success in AI that Intel needs.
 
Intel should focus on CPU performance improvements especially for the mobile ones.
Lunar Lake was a good improvement over Meteor Lake, so long as you don't need more than 8 cores. The respin of Meteor Lake they did on Intel 3 also seems like a solid step forward.

I haven't followed iGPU developments as closely, but that also seems to be an area benefiting mobile quite nicely.

Maybe the biggest challenge their mobile chips face is competitive pricing?
 
The SYCL Joint Matrix Extension appears to offer new life for SYCL/OneAPI, and Intel's Battlemage GPUs will be supported.

Info is from the IXPUG presentations, a month ago.

They say the same code will run on CUDA, AMD, Intel GPUs and Intel AMX matrix operations on Xeon.
 
  • Like
Reactions: bit_user