News Intel releases new software for GPU-powered Project Battlematrix workstations — Arc Pro B-series GPUs get LLM Scaler 1.0 software to optimize perfo...

[Admin]

Intel announces its first container for its upcoming software that will power Project Battlematrix workstations. Intel has released its first version of its LLM Scaler on GitHub, which features early enablement support and optimizations for several AI model types for Arc Pro B-series GPUs.

Intel releases new software for GPU-powered Project Battlematrix workstations — Arc Pro B-series GPUs get LLM Scaler 1.0 software to optimize perfo... : Read more

 

[cyrusfox]

will be interested to see the benchmarks that come out of here in terms of tokens per second and whether the full memory bandwidth can be utilized to load accurate models. Winning on pricing for the memory, depends on performance metric and how quick the software can come to market.

 

[gg83]

Im noticing in medicine and tech that there's really only specialists without anyone person or company that can do it all. Intel used to dominate fab and design, doctors used to diagnose and treat patients. Now the fabless are beating Intel. When hospitals loose a specialist, neuro doctor for instance, the hospital can no longer offer those types of services. Philips used to do everything and same with IBM, now they are very specialized. I don't even know what Philips actually still owns, maybe the medical division.
Anyway, I really want Intel to get back to top tier design and manufacture.

 

[abufrejoval]

From what little I know about making GPU scale-out work generically for LLMs, this project was a fool's errand from the very start.

But of course, it's one of those projects, where I'd love to be proven wrong, because if the approach were to actually work, you could generalize that with more popular gamer dGPUs: no Intel hardware required.

But now those Intel software engineers will never get a chance to even prove the point, since they were likely dismissed already...

Intel's dGPU stories are hard to read, because my head just automatically starts swiveling in a "no, no, no!" motion as soon as I see "Intel" and "GPU" together...

And this "product" builds on a base that is already the worst in the market: just how did they come to believe that stacking them would improve the result? Perhaps, whoever blocked the project back then, got fired first?

They could have created a bit of a stir, if they had brought out big (& cheap) RAM dGPUs at Battlemage launch a year ago. But as a last hooray? (head starts to swivel uncontrollably again...)