Unverified50% confidenceBenchmarkExact time
RTX 4090的理论FP16算力为330 TFLOPS,对于7B FP16模型的Decode实际速度通常在50-60 tokens/s
1
Sources
50%
Confidence
Long-term
Relevance
9/10/2026
First Seen
Sources
Related Entities
Related Claims
Unverified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 TFLOPS79% similarUnverified在显存充足、模型完整驻留GPU时，RTX 4090跑Q4量化的27B模型通常可达15～30 tokens/s；触发CPU卸载后速度可能骤降至1～2 tokens/s甚至更低78% similarUnverified该模型在单张RTX 5090上使用NVFP4方案可达约206 tokens/秒75% similarUnverifiedRTX 4090算力约165 TFLOPS，而单路推理实际只需约54 TFLOPS，算力严重过剩74% 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 TFLOPS73% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/887580API
curl https://kongchang.com/api/v1/knowledge/claims/887580MCP
get_claim(id=887580)