1. News
  2. SCALE: CUDA on AMD GPUs

Spectral Compute’s SCALE: NVIDIA’s CUDA Now Functions on AMD GPUs | Parts of Computer and Their Functions | Gpu vs cpu for Gaming | Computer Hardware | Turtles AI

Spectral Compute’s SCALE: NVIDIA’s CUDA Now Functions on AMD GPUs
Spectral Compute introduces a GPGPU toolchain breaking CUDA’s exclusivity

Highlights:

  • SCALE allows NVIDIA’s CUDA to function on AMD GPUs without code porting.
  • Developed by Spectral Compute, SCALE does not use NVIDIA code but is compatible with CUDA.
  • Tested on various applications, SCALE eliminates the need for code translation.
  • NVIDIA has shown resistance to tools like SCALE, but Spectral Compute continues to develop and test the tool.

 

British startup Spectral Compute has unveiled SCALE, a GPGPU toolchain that allows NVIDIA’s CUDA to function seamlessly on AMD’s GPUs without the need for code porting.

 

Spectral Compute has introduced "SCALE," an advanced GPGPU toolchain that allows NVIDIA’s CUDA technology to be used seamlessly on AMD GPUs. This tool promises to break NVIDIA’s historical exclusivity, opening new cross-platform development prospects. This innovation follows previous projects like ZLUDA, which began the process of porting CUDA libraries to AMD. However, SCALE aims for a deeper and more versatile level of integration.

 

Spectral Compute’s CEO, Michael Sondergaard, stated that their goal is to create an open-source environment for GPUs, similar to CPUs. SCALE acts as a bridge to close the compatibility gap between CUDA and other hardware manufacturers, eliminating the need to rewrite code.

 

Main Features of SCALE

Feature                   Description

Compatibility          Allows the use of CUDA on non-NVIDIA GPUs, such as those from AMD.

Development          In development for seven years, it does not use NVIDIA code but builds a CUDA-compatible toolchain.

License                      Accessible through a free software license, not open-source.

Tested                        ApplicationsBlender, Llama-cpp, XGboost, FAISS, GOMC, STDGPU, Hashcat, NVIDIA Thrust with AMD’s RDNA 3 and RDNA 2 architectures.

 

According to Sondergaard, SCALE represents a GPGPU toolkit analogous to CUDA, capable of using binaries for non-NVIDIA GPUs when compiling CUDA code. This eliminates the need for translation layers and allows developers to maintain a single version of their code. SCALE is designed to be highly adaptable, avoiding code porting and enabling developers to work with a single codebase compatible with CUDA.

 

NVIDIA has shown resistance to resources like SCALE that allow CUDA to function on external components, issuing warnings in their EULA against such platforms. CUDA has been a crucial element in NVIDIA’s dominance in the AI market, and the company does not want to lose this exclusive status easily.

 

Despite these challenges, Spectral Compute has continued to develop and test SCALE across a variety of applications, demonstrating its versatility and potential. If implemented on a large scale, SCALE could significantly transform the GPGPU development landscape, promoting greater interoperability and flexibility for developers.

 

Final thoughts: The integration of CUDA on non-NVIDIA platforms through SCALE could open new opportunities for AI software development, enhancing collaboration between different GPU technologies. However, it remains to be seen how NVIDIA will respond to this new challenge to its supremacy.