SM100
A compute capability designation in NVIDIA's Blackwell architecture, corresponding to data-center-class GPUs such as the B100 and B200 series. Compared with Hopper's SM90, it offers marked improvements in low-precision compute throughput, memory bandwidth and NVLink interconnect bandwidth.
Timeline (last 90 days)
SM120 is the compute capability version number of NVIDIA Blackwell architecture consumer graphics cards (such as the RTX 50 series), while data-center professional cards correspond to SM100.
MiniMax officially only supported SM100 professional servers at first; after the uploader added SM120 support for it, speed rose from 45 tokens/s to 51.6 tokens/s.
Currently the hardware supported by DeepSeek-Infer is NVIDIA SM100 GPUs and Huawei Ascend 950.
SM100 corresponds to GPUs with compute capability 10.0 in NVIDIA's Blackwell architecture, such as B200 and GB200; it is an architecture identifier rather than a specific graphics card model.
The FlashAttention 4 integration on SM100 adds FP8 KV cache support and headdim-256 support.
All Facts (5)
SM120 is the compute capability version number of NVIDIA Blackwell architecture consumer graphics cards (such as the RTX 50 series), while data-center professional cards correspond to SM100.
50%UnverifiedMiniMax officially only supported SM100 professional servers at first; after the uploader added SM120 support for it, speed rose from 45 tokens/s to 51.6 tokens/s.
50%UnverifiedCurrently the hardware supported by DeepSeek-Infer is NVIDIA SM100 GPUs and Huawei Ascend 950.
50%UnverifiedSM100 corresponds to GPUs with compute capability 10.0 in NVIDIA's Blackwell architecture, such as B200 and GB200; it is an architecture identifier rather than a specific graphics card model.
50%UnverifiedThe FlashAttention 4 integration on SM100 adds FP8 KV cache support and headdim-256 support.
50%