Unverified50% confidenceBenchmarkExact time
DeepSeek published a sparse prefill figure of 1350 TFLOPS (absolute value) for the NVIDIA B200, which uses a different metric from the 95% peak utilization figure cited for Ascend.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/1/2026
First Seen
Valid until: 12/30/2026
Sources
Related Entities
Related Claims
UnverifiedThe DeepSeek distilled versions range in parameter count from 1.5B to 70B.72% similarUnverifiedFlashMLA's prefill throughput for the sparse attention operator on the Ascend 950 reaches 410 TFLOPS (95% of the hardware's theoretical peak), and decode reaches 360 TFLOPS (83%), both figures as published by DeepSeek itself with no third-party measurements.71% similarUnverifiedA 428B dense model requires approximately 856 TFLOPs per token for forward propagation, while M3 requires only approximately 46 TFLOPs, reducing the amount of computation by approximately 18 times.69% similarUnverifiedAt FP16 half-precision computation, the RTX 4090 has a theoretical performance of approximately 330 TFLOPS, while the M2 Ultra's GPU delivers approximately 27.2 TFLOPS.69% similarUnverifiedRTX 4090算力约165 TFLOPS，而单路推理实际只需约54 TFLOPS，算力严重过剩68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/963641API
curl https://kongchang.com/api/v1/knowledge/claims/963641MCP
get_claim(id=963641)