Unverified50% confidenceBenchmarkExact time
On the gate/up projection with SwiGLU at 8B shapes, FuTransformer's BF16 path is 1.13x faster than CODA, and its FP8 path is 2.02 to 2.18x faster than the evaluated BF16 CODA path.
1
Sources
50%
Confidence
Short-term (~14 days)
Relevance
10/19/2026
First Seen
Valid until: 11/2/2026
Sources
Related Entities
Related Claims
UnverifiedIn BF16 configuration (two 48GB GPUs), decoding speed ranges from 25.1 tok/s at 1K to 23.9 tok/s at 128K, with almost no degradation.64% similarUnverifiedAt a 32K context length, WavePrune achieves 1.15× prefill speedup and 1.24× decoding speedup compared to FlashAttention-2.61% similarUnverifiedWhen mainstream inference engines run this model on the same hardware, the speed is generally stuck in the 6–8 FPS range.61% similarUnverifiedIn Path C, an RTX 3080 paired with a DGX Spark improves Prefill speed from 1111.3 to 1600 (approximately 44%), but Decode speed drops from 33.36 token/s to 17.62 token/s.61% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1065907API
curl https://kongchang.com/api/v1/knowledge/claims/1065907MCP
get_claim(id=1065907)