Unverified50% confidenceFactExact time
The A100 wastes more than 97% of HBM bandwidth during 4K-token prefill; HBM accounts for 18% of the A100's manufacturing cost, and as much as 45% on the B200.
1
Sources
50%
Confidence
Long-term
Relevance
10/19/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedWSE-3的片上SRAM容量达到44GB，SRAM访问延迟约1ns，比HBM约100ns低10-50倍69% similarUnverifiedA100的L2缓存为40MB，H100的L2缓存为50MB67% similarUnverifiedFor 1 million tokens, the KV Cache alone can consume hundreds of GB of VRAM at FP16 precision, far exceeding the 80 GB capacity limit of a single A100/H100.66% similarUnverifiedThe latest generation of HBM3E can include 8 to 12 DRAM die per stack, providing over 1 TB/s of bandwidth, which is more than 20 times the bandwidth of conventional DDR5 memory (50–60 GB/s).64% similarUnverifiedUnder the above assumptions, an 8192-token prompt's transfer time is about 30% of prefill time on a 100GbE link, about 15% at 200Gb, and about 7% at 400Gb.64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1070131API
curl https://kongchang.com/api/v1/knowledge/claims/1070131MCP
get_claim(id=1070131)