Unverified50% confidenceFactExact time
Hopper supports FP8, but its FP8 scaling is at tensor granularity, requiring a global scaling factor to be pre-computed before each matrix multiplication.
1
Sources
50%
Confidence
Long-term
Relevance
9/16/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe Krea team simultaneously released specially optimized FP8 weights at the time of the model launch, rather than simply performing post-hoc quantization of the BF16 weights62% similarUnverifiedNVFP4 introduces a per-group scaling factor mechanism, typically sharing one FP8 scaling factor per group of 128 weights61% similarUnverified模型量化是将模型权重从FP16压缩至INT4/INT8以节省显存59% similarUnverifiedTechnical approaches to mitigating the inference wall include model quantization (reducing FP16 to INT8 or INT4), speculative decoding (using a small model to predict large model outputs to accelerate generation), and KV Cache optimization.57% similarUnverifiedTensor contraction is a general operation in linear algebra that sums multidimensional tensors along specified dimensions; matrix multiplication is essentially a form of tensor contraction56% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/927332API
curl https://kongchang.com/api/v1/knowledge/claims/927332MCP
get_claim(id=927332)