Jim Keller has shared a string of posts over at X where he has called out NVIDIA’s CUDA for being a swamp and also said the same was true for x86.
NVIDIA’s CUDA Isn’t a Moat, Rather A Whole Swamp, Says Jim Keller, x86 Was One Too But Has Seen Improvements
In the current day and age, NVIDIA is dominating the AI segment not because it has high computing power on its back but because its CUDA platform has been optimized in a way to leverage the performance of the company’s AI accelerators. CUDA is seen as NVIDIA’s primary catalyst for the firm’s massive success in the data center segment. However, industry personalities such as Intel’s CEO Pat Gelsinger and Tenstorrent’s CEO Jim Keller have criticized Team Green’s approach with CUDA, claiming it’s a trap for the firm’s sustainability in the AI segment.
Cuda’s a swamp, not a moat. x86 was a swamp too
— Jim Keller (@jimkxa) February 17, 2024
Jim Keller reached out on the social media platform X to express his views on the ongoing “CUDA dominance,” interestingly, he categorized the resource as a swamp and related it to how the x86 architecture was progressing back in the day. Now, how a swamp works is that it restricts an individual in itself when he enters in, and now replicating this situation with CUDA, Jim Keller meant that its limitations could potentially lead to CUDA’s downfall. While he didn’t specify the reason behind this categorization, we have some “educated” guesses.
If we look at how x86 progressed with time, instead of receiving fundamental changes, the architecture was developed in light of prioritizing backward compatibility. While this might sound beneficial, it led to the creation of layered, intricate architectures in the long run, which made future development quite hectic. The case with NVIDIA CUDA might seem similar as well, and Jim’s analysis probably makes sense here since building upon existing foundations sometimes means carrying inefficiencies.
Meanwhile, Jim Keller’s Tenstorrent has already pledged to compete against NVIDIA’s AI GPUs using its very own RISC-V-based AI chips which are backed by Buda and BudaM software suites. More on those here.
Apart from this, NVIDIA’s reliance on CUDA isn’t as high as what we would expect since the firm does utilize other resources, such as Triton and TensorRT, for specific tasks, suggesting limitations in CUDA’s general-purpose performance. Since CUDA isn’t the primary weapon of NVIDIA’s AI software arsenal, this could lead to potential performance degradation, as introducing overhead layers isn’t sustainable moving into the future.
The statements on CUDA by industry experts highlight the technology’s flaws. Still, we all know how NVIDIA “reveals its cards” on the deck, which means that moving ahead, we could see a dynamically modified CUDA, but of course, this is for the future to decide. The company has its upcoming GTC event planned for next month where it will be talking a lot about AI and software developments so expect to hear more from Team Green soon!