Article Intel cans Omnipath

[NTMBK]

Intel Starts to Close Omni-Path: OPA1 Xeon CPUs on EOL, OPA2 Axed

www.anandtech.com

Phi is dead, Omnipath is dead. Intel's HPC visions have changed drastically in the last few years!

 

[Yotsugi]

Intel's HPC visions have changed drastically in the last few years!

It's not like either of those were any good.
(Being fair, Slingshot is by far the best HPC interconnect there currently is).

 

[DrMrLordX]

Phi is dead, Omnipath is dead. Intel's HPC visions have changed drastically in the last few years!

They must be banking heavily on Xe by this point.

 

[NTMBK]

They must be banking heavily on Xe by this point.

Yeaaaah... Good luck with that. I can't see a new GPU competitor unseating CUDA at this point. AMD have had to basically give up and make a CUDA compatibility layer.

 

[DrMrLordX]

AMD have had to basically give up and make a CUDA compatibility layer.

I would expect Intel to do the same. Either that or they can fork OpenCL and update it (or do both after a fashion).

 

[Yotsugi]

I would expect Intel to do the same

Nah, they're leaning heavily on OCL and SYCL.

 

[NTMBK]

Nah, they're leaning heavily on OCL and SYCL.

I repeat... good luck with that one.

 

[Yotsugi]

I repeat... good luck with that one.

Intel has enough manpower to wrestle absolutely anything or anyone in sw.

 

[jpiniero]

Intel has enough manpower to wrestle absolutely anything or anyone in sw.

For that to work though, Xe would have to be pretty competitive hardware wise.

 

[Yotsugi]

For that to work though, Xe would have to be pretty competitive hardware wise.

True that, but the point still stands.
Intel *can* make good sw.

 

[moinmoin]

Intel *can* make good sw.

That's not the issue. Getting customers to then use (and ideally actively ask for) it is the problem.

 

[ThatBuzzkiller]

Intel has enough manpower to wrestle absolutely anything or anyone in sw.

Even if they do how exactly is it even a good idea to lean on OCL/SYCL for a compute stack ?

Why would they subject themselves to the torture of less powerful higher level APIs like OCL/SYCL in comparison to higher performing lower level APIs such as CUDA or HIP ? Intel is already going to be losing some performance on the table in the HPC space because they now have to deal with overhead due to the mismatch between SPIR-V kernels and their native hardware ...

Also by introducing an intermediate representations like SPIR-V, Intel risks making things far harder with potentially dealing SPIR-V compiler bugs ...

If this is their 'vision' for GPU compute then they'll soon figure out just how hard it truly is for industry to adopt standards like AMD has in the past with trying to push OpenCL and that ended up in failure ...

 

[Yotsugi]

Even if they do how exactly is it even a good idea to lean on OCL/SYCL for a compute stack ?

Why would they subject themselves to the torture of less powerful higher level APIs like OCL/SYCL in comparison to higher performing lower level APIs such as CUDA or HIP ? Intel is already going to be losing some performance on the table in the HPC space because they now have to deal with overhead due to the mismatch between SPIR-V kernels and their native hardware ...

Also by introducing an intermediate representations like SPIR-V, Intel risks making things far harder with potentially dealing SPIR-V compiler bugs ...

If this is their 'vision' for GPU compute then they'll soon figure out just how hard it truly is for industry to adopt standards like AMD has in the past with trying to push OpenCL and that ended up in failure ...

Portability is king now.
End of the story.

 

[ThatBuzzkiller]

Portability is king now.
End of the story.

If only this were true ...

AMD isn't coming to save them and Nvidia especially coming to save Intel ...