x.com/teortaxesTex/status/2080147241956770231/photo/1
1 correction found
when V3 trained, it still used NVIDIA GPUs, but no longer used NVIDIA's ecosystem...
DeepSeek-V3 did use NVIDIA’s broader ecosystem during training. DeepSeek’s own V3-related training stack relied on NVIDIA-specific technologies including CUDA, NVLink, NVSHMEM, and Hopper/PTX support.
Full reasoning
DeepSeek’s public materials do not support the claim that V3 training had already left NVIDIA’s ecosystem.
- DeepSeek’s official V3 repository says the model required 2.788M H800 GPU hours for full training and describes an FP8 mixed-precision training framework built around that hardware.
- DeepSeek’s own DeepEP documentation says it follows the DeepSeek-V3/R1 pretraining setting on H800 with NVLink and RDMA, and lists Ampere/Hopper GPUs, CUDA, NVLink, and NVSHMEM as requirements.
- DeepSeek’s DeepGEMM library is explicitly described as a “CUDA codebase” for NVIDIA GPU kernel optimization techniques, and it requires NVIDIA SM90/SM100 GPUs plus the CUDA Toolkit.
Those are all parts of NVIDIA’s ecosystem, not just NVIDIA chips. So while V3 was indeed trained on NVIDIA GPUs, the public evidence shows its training stack still depended on NVIDIA-specific software and interconnect technologies rather than having already moved beyond NVIDIA’s ecosystem.
3 sources
- GitHub - deepseek-ai/DeepSeek-V3
Despite its excellent performance, DeepSeek-V3 requires only 2.788M H800 GPU hours for its full training. ... We design an FP8 mixed precision training framework...
- DeepEP/docs/legacy.md at main · deepseek-ai/DeepEP
We test normal kernels on H800 ... and we follow the DeepSeek-V3/R1 pretraining setting ... Requirements: Ampere (SM80), Hopper (SM90) GPUs ... CUDA ... NVLink ... DeepEP V1 depends on NVSHMEM.
- GitHub - deepseek-ai/DeepGEMM
DeepGEMM is ... a single, cohesive CUDA codebase ... a clean and accessible resource for learning NVIDIA GPU kernel optimization techniques. Requirements: NVIDIA SM90 or SM100 architecture GPU ... CUDA Toolkit.