Benchmark metricsSoL
Speed-of-Light throughput
In plain English
A stated performance ceiling used to express how close a measured operation gets to its relevant hardware limit.
Technical definition
Speed-of-Light throughput is a normalization against a specified theoretical or attainable performance ceiling for the operation and resource being measured.
Engineering details
The denominator must match the data format, instruction shape, clock basis, resource count, and sparse or dense arithmetic convention. A percentage without that denominator can conceal incompatible comparisons. Compute and memory ceilings also describe different limits and should not be substituted for each other.
Why it matters
High SoL in an isolated instruction test does not establish high application utilization. A model may spend much of its time in other operations, dependencies, communication, or idle gaps.
How to read it in InferenceX
The Blackwell port reports different fractions of peak at deep and shallow MMA queue depths. Use the queue depth resembling the real kernel before interpreting a saturation result as practically attainable.
See the concept in real benchmarks
All articlesDissecting Nvidia Blackwell - Tensor Cores, PTX Instructions, SASS, Floorsweep, Yield
Blackwell SM100 microbenchmarks of PTX and SASS instructions, tensor-core shapes, memory copies, cluster scheduling, and floorsweeping.
The Huawei-DeepSeek Symbiosis
Ascend 950PR architecture and microbenchmarks, DeepSeek’s Ascend software stack, and the DeepSeek-V4.1-Flash execution profile.