Benchmark metrics
Microbenchmark
In plain English
A focused test that isolates an instruction, resource, or small operation rather than measuring a complete application.
Technical definition
A microbenchmark deliberately restricts the workload to expose a particular latency, throughput limit, dependency, or resource interaction.
Engineering details
The harness chooses instruction shapes, concurrency, repetitions, and cache conditions. These choices can suppress costs that dominate production, such as launch overhead, short dependency chains, or communication. Timing also needs a clearly stated scope: a single instruction, one SM, a device, or a complete invocation.
Why it matters
Microbenchmarks are useful for diagnosing bottlenecks and testing architectural hypotheses. They do not by themselves predict requests per second or user-visible latency because application scheduling combines many resources and dependencies.
How to read it in InferenceX
Both accelerator ports use isolated tests alongside broader discussion. Blackwell’s deep in-flight MMA queues and Huawei’s warm-cache operator tests need their original configuration boundaries attached to any comparison.
See the concept in real benchmarks
All articlesThe Huawei-DeepSeek Symbiosis
Ascend 950PR architecture and microbenchmarks, DeepSeek’s Ascend software stack, and the DeepSeek-V4.1-Flash execution profile.
Dissecting 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.