A Chinese language GPU startup has acquired funding from the nation to develop chips that function compatibility with NVIDIA’s CUDA & OpenCL programming fashions.
China Startup On Its Approach To Develop NVIDIA Rivaling & CUDA-Suitable GPUs After Securing Funding From China Web Funding Fund
The Shanghai-based startup, Denglin, was based in 2017 and has now acquired funding from the CIIF (China Web Funding Fund) which was additionally initiated collectively by the State Our on-line world Administration of China and the Ministry of Finance to develop chips for the home market which might be each CUDA and OpenCL appropriate. The chips will primarily be used for GPGPU utilization however will be additional utilized in extra premium markets akin to HPC and AI after their growth succeeds.
Denglin will probably be creating their chips on the GPU+ know-how which is alleged to supply a software-defined on-chip heterogeneous computing structure and can assist each CUDA and OpenCL, stories JonPeddie.
As for the product lineup, the China-based GPU maker initially unveiled 4 merchandise which ranged from gaming to AI coaching. The highest resolution is the Goldwasser chip which is designed for AI-based compute acceleration and will probably be getting edge and cloud computing platforms. The UL32 and UL64 are 10-15W options, providing as much as 64 TOPs INT8 and 16 TFLOPs FP16 compute with assist of as much as 16 GB reminiscence and 64-channels video encoding at 1080P@30FPS.
The corporate states that its 2nd technology GPUs have been taped out in 2022 and mass manufacturing is predicted this 12 months. The brand new GPUs will ship a 3-5 time efficiency acquire in transformer-based fashions and tremendously cut back the {hardware} value in workloads akin to ChatGPT and Generative AI.
Specification | Goldwasser UL32 | Goldwasser UL64 |
Energy consumption | 10W | 15W |
Interface | MXM3.1 | MXM3.1 |
INT8 | 32 TOPS | 64 TOPS |
FP16 | 8 TFLOPS | 16 TFLOPS |
Video decoding (H.264/H.265) | 32 channels, 1080p@30 FPS | 64 channels, 1080p@30 FPS |
Reminiscence (configurable) | 4G/8G/16G | 4G/8G/16G |
A chipmaker that gives CUDA programming capabilities and high-end GPUs for AI processing could be a massive deal for China’s home market. Whereas NVIDIA and Intel are each pushing to provide extra chips inside the nation, the accelerators themselves are cut-down variants of what’s out there in different areas resulting from US-based restrictions. Moreover, it’s changing into tougher to accumulate these chips resulting from growing demand and better prices with an H800 GPU priced above 250,000 RMB or over $35,000 US.
Presently, there are 13 GPU builders in China and we must wait and see whether or not Denglin can really get these chips out out there or if they’d find yourself with the identical destiny as Biren Tech who have been poised to grow to be a serious GPU producer and compete towards NVIDIA inside the HPC section however in the end noticed no progress resulting from US rules.
Originally posted 2023-07-15 15:15:06.