Unverified50% confidenceEventExact time
The work described in the original article ports and optimizes JFA to the Blackwell platform, with the goal of approaching SOTA-level FA4 performance.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/1/2026
First Seen
Valid until: 12/30/2026
Sources
Related Entities
Related Claims
UnverifiedBlackwell 的第五代 NVLink、HBM3e 显存和专用 FP4 张量核心使其在推理密集型场景下相较 H100 可获得2-4倍吞吐量提升62% similarUnverifiedFlash Attention 4 (FA4) is designed specifically for the NVIDIA Hopper/Blackwell architecture, with core improvements including asynchronous data movement using TMA, pipelined producer-consumer warp separation, and FP8 low-precision path optimization.61% similarUnverifiedOn the H100, FA4's forward pass can achieve an MFU of over 75% of the theoretical peak.59% similarUnverifiedThe main thread of evolution from FA1 to FA4 is the continuous improvement of utilization of hardware asynchronous features (Tensor Core pipelines, asynchronous memory copies), pushing actual throughput closer to the theoretical peak FLOPS.58% similarUnverifiedHBM4E is an enhanced version built on top of HBM4, targeting high-end AI training scenarios that are extremely sensitive to bandwidth.58% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/966302API
curl https://kongchang.com/api/v1/knowledge/claims/966302MCP
get_claim(id=966302)