News Legendary GPU architect Raja Koduri's new startup leverages RISC-V and targets CUDA workloads — Oxmiq Labs supports running Python-based CUDA appli...

[Admin]

Raja Koduri's Oxmiq Labs develops RISC-V–based GPU IP and a hardware-agnostic software stack.

Legendary GPU architect Raja Koduri's new startup leverages RISC-V and targets CUDA workloads — Oxmiq Labs supports running Python-based CUDA appli... : Read more

 

[bit_user]

We may be the first new GPU startup in Silicon Valley in 25+ years,
GPUs are not easy.

Vivante was at least headquartered there. Not sure if they did any development there or not.

The oddball Bolt Graphics also lists Sunnyvale as its mailing address. I'd say they're probably more of a GPU than what Koduri is doing.

BTW, Nvidia's NVDLA uses RISC-V with (probably nonstandard) vector extensions for AI. So, he's already got some stiff competition. Think-Silicon and Imagination also have RISC-V -based GPUs.

 

[ejolson]

The ability to add custom GPU-style vector instructions as needed to RISC-V is exactly the idea behind an open ISA. Yeah!

It's more difficult to get excited about another cross-platform heterogeneous compute architecture. Everyone except Nvidia has tried to create cross-platform tools: OneAPI, OpenCL and HIP are examples.

On the other hand DeepSeek training was made possible by using PTX assembler tuned specifically to the hardware.

 

[Giroro]

Since when are we calling Koduri a "legendary architect".
He's the legendarily incompetent C-Suite bean counter who killed ATI, later returned to put a dead-stop on AMD's ability to innovate their GPUs (which they haven't completely recovered from) before being forced to move to Intel, where in under 2 years he nearly killed their dGPUs before they ever got off the ground (a critical loss which could still kill off Intel entirely).

He's a poison pill. The only company that's ever benefitted from this guy's career is Nvidia, because he's never worked for them.

 

[thisisaname]

Since when are we calling Koduri a "legendary architect".
He's the legendarily incompetent C-Suite bean counter who killed ATI, later returned to put a dead-stop on AMD's ability to innovate their GPUs (which they haven't completely recovered from) before being forced to move to Intel, where in under 2 years he nearly killed their dGPUs before they ever got off the ground (a critical loss which could still kill off Intel entirely).

He's a poison pill. The only company that's ever benefitted from this guy's career is Nvidia, because he's never worked for them.

He has managed to get funded 🤯

 

[bit_user]

Since when are we calling Koduri a "legendary architect".
He's the legendarily incompetent C-Suite bean counter who killed ATI, later returned to put a dead-stop on AMD's ability to innovate their GPUs (which they haven't completely recovered from) before being forced to move to Intel, where in under 2 years he nearly killed their dGPUs before they ever got off the ground (a critical loss which could still kill off Intel entirely).

I think you're too harsh, but I do suspect that both AMD and Intel kicked him to the curb. In fairness, work on RDNA should've already started while he was still at AMD, so maybe he deserves partial credit for it?

I think he was probably a decent engineer, early in his career, which is how he rose through the ranks. What we might be seeing is the Peter Principle at work. In short, it observes that success is rewarded by promotion. Absent other considerations, one would therefore expect an employee to continue getting promoted until they reach a position for which they're no longer competent.

Of course, even if that happens once, you don't tend to get another job at that level, unless you're also a shameless self-promoter. So, I'm not trying to say Raja is merely a victim of his early successes.

He's a poison pill. The only company that's ever benefitted from this guy's career is Nvidia, because he's never worked for them.

LOL, you might be on to something!
: D

 

[bit_user]

He has managed to get funded 🤯

Not with very much. As the article noted, $20M is toy money in the hardware game.

Not sure how much of it is his own or Jim Keller's, either. Maybe the Tenstorrent deal also had a lot to do with securing that seed funding.

I don't really know what Jim sees in Raja, but I would observe that the last time they actually worked together was at Apple, in the late 2000's.

 

[SkyBill40]

He has managed to get funded 🤯

A fool and his money are soon parted. 🤷‍♂️

 

[umeng2002_2]

Legendary? Have you seen AMD GPUs under him?

 

[Alvar "Miles" Udell]

Raja sure is legendary. He messed up Vega so badly it almost bankrupted AMD, then he messed up Intel's GPUs so badly it almost bankrupted the company...

 

[wwenze1]

Another startup developing GPUs that are not meant for graphics

AMD GPU during the later GCN era: People buy them to mine crypto

Intel ARC: People buy them to... er... something

Also, Graphics Processing Units that are not meant for graphics lol

 

[Pierce2623]

Considering the US government gave Nvidia a ruling to prevent a CUDA translation layer from going commercial, I doubt they’ll be able to sell this in the US, if Nvidia decides to prevent it.

 

[bit_user]

Considering the US government gave Nvidia a ruling to prevent a CUDA translation layer from going commercial,

Courts issue rulings. Congress passes laws. Which was it? Got a link?

When people talk about "the government doing something", they usually mean Congress or the Executive branch. Courts only hear cases that are brought before them, so the question of initiative is reversed.

I doubt they’ll be able to sell this in the US, if Nvidia decides to prevent it.

Seems to me that this SCOTUS ruling paves the pathway for work-alike API implementations:

Google LLC v. Oracle America, Inc. - Wikipedia

en.wikipedia.org