Efficiency Index
GPU cost and speed on replayed agent workloads. Measured and simulated points are marked.
LLM serving benchmarks.
Throughput, latency and cost per token by GPU, on chat and agent workloads.
Models
Workloads
GPU prices
Rental price, $ per GPU hour. Used by the Rent basis.
Token prices
$ per million tokens. Sets revenue.
Config i
Frontier: no other point is better on both axes.
Compare two GPUs
vs
◂Drag▸
Efficiency score
Solved tasks per dollar of GPU time, for each model on each GPU setup.
How the costs are built
Kernel profiling
Time per decode step by operation, one row per profiled run.