Unverified50% confidenceBenchmarkExact time
在4K×4K×16K矩阵乘法规模下，M5 Ultra达到91 TFLOPS，M5 Max和M3 Ultra分别为34和22 TFLOPS
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
9/23/2026
First Seen
Valid until: 12/22/2026
Sources
Related Entities
Related Claims
UnverifiedM5 Ultra的UltraFusion晶片间带宽从M3 Ultra的约2.5TB/s提升至4.4TB/s，连接密度提升6倍75% similarUnverified以FP16半精度计算，RTX 4090理论算力约为330 TFLOPS，M2 Ultra的GPU算力约为27.2 TFLOPS69% similarPartially VerifiedGPU重度负载下M3 Ultra功耗约100W，M5 Ultra飙到400W以上68% similarUnverified428B 稠密模型每 token 前向传播约需 856 TFLOPs，而 M3 仅需约 46 TFLOPs，计算量降低约 18 倍66% similarUnverifiedTPU v4单芯片峰值算力约275 TFLOPS（BF16精度），TPU v4-32 pod由32个芯片通过ICI高速互联组成，总算力约8800 TFLOPS66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/945111API
curl https://kongchang.com/api/v1/knowledge/claims/945111MCP
get_claim(id=945111)