Unverified50% confidenceBenchmarkExact time
The mxfp scheme (stochastic rounding + Hadamard preprocessing) achieves 37.2K tokens/s/GPU, corresponding to 86.3% MFU, with a final training loss 2.11% higher than pure bfloat16.
1
Sources
50%
Confidence
Long-term
Relevance
10/4/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedOn an A100 (80 GB VRAM) GPU, the throughput for training a 7B model is approximately 3,000–5,000 tokens per second.70% similarUnverifiedThe Q8 main model occupies 162 GB, preventing the drafter from fitting into unified memory, rendering DSpark acceleration ineffective; after moving three layers of MoE experts back to the CPU, output only improved from 22.87 to 23.49 tokens/second (2.7% faster), while prefill dropped from 95.38 to 76.65.67% similarUnverifiedSGLang using official NVFP4 weights with D-Spark speculative decoding achieved an average of approximately 173 tokens/second on Blackwell GPUs, approaching 200 for short tasks, while loading a 262k context window.64% similarUnverifiedVM级别隔离（如Kata Containers）会带来约15%-30%的GPU性能损耗64% similarUnverifiedFor the LLaMA-2 70B model, when processing a context of 4096 tokens, the KV Cache requires approximately 20 GB of system memory at FP16 precision.64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/971842API
curl https://kongchang.com/api/v1/knowledge/claims/971842MCP
get_claim(id=971842)