Raja Koduri Departing Intel To Start New Software Company

[user1]

Please please please go to nvidia!

Why? Do you want an already pathetic dGPU market to be even more pathetic when it's even more dominated by a greedy monopoly?

[Spacefish]

Raja should be put in charge for the next gen NVIDIA Architecture and Software platform.. please! I have AMD Stocks b.t.w.

[WannaBeOCer]

well, it was always over-promise and under-deliver, regardless whether at AMD or at Intel. I first thought this had to do with the companies, but it seems to somehow be linked with certain personalities, too.

He delivered AMD ROCm, CDNA and RDNA and within his first year at Intel OneAPI. Let’s not forget that he completely flipped AMD’s unstable drives when releasing AMD Crimson when he rejoined AMD.

Hopefully he pushes AI adoption of AI generated physics and locomotion in games.

[kiffmet]

WannaBeOCer he was not responsible for the software, like at all. Arc Graphics is now part of the client computing group under the leadership of Lisa Pearce, who has been the head of the GPU software and driver team before.

Ponte Veccio had so many issues and revisions that its successor Rialto Bridge was outright cancelled…

Subjectively, there have been two major problems with Raja: He cannot execute/deliver projects on time, and he's not very good in communicating with other departments. That was also seen back then at AMD; The driver teams never got the hardware documentation in time; same goes for ES silicon that's also used for driver bringup.

[WannaBeOCer]

WannaBeOCer he was not responsible for the software, like at all. Arc Graphics is now part of the client computing group under the leadership of Lisa Pearce, who has been the head of the GPU software and driver team before.

Ponte Veccio had so many issues and revisions that its successor Rialto Bridge was outright cancelled…

Subjectively, there have been two major problems with Raja: He cannot execute/deliver projects on time, and he's not very good in communicating with other departments. That was also seen back then at AMD; The driver teams never got the hardware documentation in time; same goes for ES silicon that's also used for driver bringup.

When did I mention Arc deivers? He oversaw the development of OneAPI which has been a success. Same with ROCm, CDNA cards are derivatives of Vega. I bet if Raja Koduri was still at AMD at the launch of the RX 5000 series they wouldn’t have had all of those driver issues that lasted months.

He’s been jumping around every 5 years for quite some time. Left AMD, went to Apple to help with their GPU, came back to AMD and left. Similar to many engineers.

[theriddick]

Intel admitting Arc's failure? First AXG is gone, now Raja is gone. Next: Arc will be gone?

Arc is showing signs of success lately (under Windows mostly) thanks to the massive increase in cost for the new AMD/NV cards. A 4060 for $500USD..... come on!

Anyway the Arc hardware isn't the issue; the software and drivers are the core of the problem and are improving significantly overtime.
It's possible that within the next year the Arc GPU's could gain %20-50 more performance thanks to improvements in the drivers.

A lot of people are still looking for 3060-3070 level of performance but at cheaper prices which the Arc GPU's are starting to fill that space.

[brucethemoose]

He delivered AMD ROCm, CDNA and RDNA and within his first year at Intel OneAPI. Let’s not forget that he completely flipped AMD’s unstable drives when releasing AMD Crimson when he rejoined AMD.

Hopefully he pushes AI adoption of AI generated physics and locomotion in games.

IDK what you mean by AI generated physics. Generated mocap?

I'm guessing its mostly gonna be on the developers side of things, as most gamer's PCs arent beefy enough to run generative AI and a game at the same time. Buyers would be pissed if the requirement was like an RTX 3070 or an AMD 7800 and up, and still looked kinda low fi.

[WannaBeOCer]

IDK what you mean by AI generated physics. Generated mocap?

I'm guessing its mostly gonna be on the developers side of things, as most gamer's PCs arent beefy enough to run generative AI and a game at the same time. Buyers would be pissed if the requirement was like an RTX 3070 or an AMD 7800 and up, and still looked kinda low fi.

AI based physics and locomotion barely use any resources. The training does require a lot of resources. Now that the RX 7000 series has dedicated tensor accelerators like the RTX 20 series and newer hopefully we’ll see more games adopt these models. We’ve already seen games with AI generated facial expressions for example in Cyberpunk 2077. Physics has already been too hardware intensive, AI makes it a much easier workload to run.

https://youtu.be/atcKO15YVD8

https://youtu.be/8oIQy6fxfCA

Edit: Found a new video by Unity which is quite interesting.

https://youtu.be/Fa49SQtLOsA

[Quackdoc]

He delivered AMD ROCm, CDNA and RDNA and within his first year at Intel OneAPI. Let’s not forget that he completely flipped AMD’s unstable drives when releasing AMD Crimson when he rejoined AMD.

Hopefully he pushes AI adoption of AI generated physics and locomotion in games.

RocM is a success story!?!? more like a joke. it's preformance is mediocre, and unstable. I had a fresh fedora crash when using the rocm opencl runtime. ROCM is laughably bad

[Paradigm Shifter]

AMD seem to be doing something right at least partially, because they are making waves in the supercomputer world both with CPUs and GPUs, but their support for more consumer accessible hardware in their compute stack isn't just a bad joke, it's continuing to beat the glue that was the remains of the horse a decade ago. Their continued false starts in the compute space allowed nVidia to open up a lead (and then achieve total dominance) which is going to be almost impossible to break. OpenCL, heterogeneous computing, Vulkan, Mantle, ROCm...

ROCm is the latest, of course, and right now the closest AMD has come to actually following through on their promises, but it's still a nightmare to get working reliably, depending far, far too much on BIOS support, specific hardware (most of which are old and now next to impossible to actually source!), and what OS and version of said OS you are running. Getting ROCm to work successfully feels like having the stars align after performing some arcane and esoteric ritual. For consumers interested in compute, it really does feel like over promising and under delivering.

On the Raja front, I never really fully understood why he seems to have such near-mythical stature among some. But the last decade seems to show that either he was extremely lucky earlier in his career, was extremely unlucky for the last decade, or, possibly, just worked in a team with dozens or hundreds of other people and is actually nowhere near as mythic a figure as cast.

I wish him and his future endeavours well, but given that we appear to be staring a financial crisis which will absolutely dwarf the clusterfarce that 2008 was right in the face... well, either something will happen and the new company will be positioned to take advantage of a competitors weakness, or it will get ground into dust inside five years.