Feb 20, 2024 11:20:00

Jim Keller, the creator of Apple A4 chips and AMD Ryzen, criticizes NVIDIA's CUDA and x86 architecture by calling it a 'swamp'

It was reported that Jim Keller, an architect and engineer who was involved in the development of various chips such as AMD's Athlon , Zen microarchitecture , and Apple A4 , criticized ``NVIDIA's CUDA is a swamp'' on X (formerly Twitter). It is being

Jim Keller criticizes Nvidia's CUDA, x86 — 'Cuda's a swamp, not a moat. x86 was a swamp too' | Tom's Hardware
https://www.tomshardware.com/tech-industry/artificial-intelligence/jim-keller-criticizes-nvidias-cuda-and-x86-cudas-a-swamp-not-a-moat-x86-was-a- swamp-too

You can understand Mr. Keller's career by reading the article below.

An interview with Jim Keller, a genius engineer who worked at Apple, AMD, Tesla, and Intel, is being released, so what exactly was he doing at Intel? -GIGAZINE

In addition, Mr. Keller will launch the semiconductor fab ``Atomic Semi'' in February 2023.

Jim Keller, the genius engineer who created AMD Ryzen and Apple A4, launches a semiconductor manufacturing company 'Atomic Semi' - GIGAZINE

Mr. Keller commented on X, ``CUDA is a swamp, not a moat . The x86 architecture is also a swamp.''

Cuda's a swamp, not a moat. x86 was a swamp too
— Jim Keller (@jimkxa) February 17, 2024

CUDA is a computing integrated development environment using NVIDIA GPUs announced in 2006. CUDA provides a unique programming language, library, and API based on C++, and by programming with CUDA, you can perform calculations on the GPU.

CUDA has long been a competitive moat for NVIDIA, with each software and hardware update adding features while maintaining backward compatibility. However, Keller points out that while maintaining backward compatibility made it possible to run CUDA even in older environments, performance was affected and development became more difficult.

In addition, x86 was originally a 16-bit architecture, but as it progressed to support 32-bit and then 64-bit, various functions such as virtual memory and virtualization were added while retaining old functions to maintain compatibility. They were added. Mr. Keller has also mentioned in past interviews that the x86 architecture and Arm architecture have become more complex over their long history.

Keller posted, ``Basically no one writes CUDA. Even if you write CUDA, it probably won't be fast.''

Basically nobody writes Cuda. If you do write Cuda, it's probably not fast
— Jim Keller (@jimkxa) February 17, 2024

Keller also commented, ``There's a good reason why we have Triton , Tensor RT , Neon , and Mojo .'' For example, NVIDIA's Triton is an open source tool with features such as AI model version control, multi-model provisioning, model concurrency, and optimization of GPU and CPU resource usage, but it does not rely solely on CUDA. , supports frameworks such as TensorFlow, PyTorch, and ONNX. The complexity of CUDA itself is one reason why an environment has been created that allows GPU calculations to be performed without relying on CUDA.