Unverified50% confidenceBenchmarkExact time
HMA-Serve uses Tenstorrent Blackhole p150b cards for prefill and NVIDIA HBM GPUs for decode, achieving 3.2x goodput and 4.8x goodput per dollar on Qwen3 4B to 32B with no quality loss.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/19/2026
First Seen
Valid until: 1/17/2027
Sources
Related Entities
Related Claims
VerifiedThe HBM memory bandwidth of the NVIDIA H100 is approximately 3.35 TB/s, and its utilization rate in autoregressive generation tasks is often below 10%.69% similarUnverifiedThe decode unit of DeepSeek V3 uses 144-card EP, amortizing the weight load overhead across large batches and freeing up GPU memory to increase batch size.63% similarUnverifiedThe most advanced HBM3e memory bandwidth is approximately 4.8 TB/s, yet GPU compute utilization during inference often reaches only 30%–60%.62% similarUnverifiedThe HBM3 memory bandwidth of the NVIDIA H100 is as high as 3.35TB/s, while the Ryzen AI Halo, constrained by LPDDR5x bandwidth, typically operates in the range of 100–300GB/s.62% similarUnverifiedGPU HBM bandwidth is typically on the order of 2–3 TB/s (3.35 TB/s for H100), while the effective bandwidth of Host DRAM connected via PCIe 5.0 is only about 64 GB/s—a difference of nearly 50x.61% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1070130API
curl https://kongchang.com/api/v1/knowledge/claims/1070130MCP
get_claim(id=1070130)