Unverified50% confidenceFactExact time
Tokens per watt throughput is the core metric for measuring the energy efficiency of AI inference chips, referring to the number of tokens a chip can output per second under unit power consumption, usually measured in tokens/watt.
1
Sources
50%
Confidence
Long-term
Relevance
10/12/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedMFU等于实际每秒浮点运算次数除以硬件峰值FLOPS69% similarUnverifiedThe formula for local AI inference speed is: Token/s ≈ memory bandwidth / amount of data read per token69% similarUnverifiedEngram dedicates one quarter of the entire checkpoint's capacity to n-gram memory, reaching a scale of 197B parameters.66% similarUnverifiedFor models with more than 10B parameters, 4-bit quantization can typically reduce memory usage by approximately 50% while keeping capability loss within 5%.66% similarUnverifiedTaking the LLaMA-3 70B model as an example, with 80 layers, head_dim=128, and BF16 precision, each token occupies approximately 80KB of KV cache.65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1006384API
curl https://kongchang.com/api/v1/knowledge/claims/1006384MCP
get_claim(id=1006384)