Software
PTX
In plain English
NVIDIA’s virtual instruction language, which is compiled into architecture-specific GPU machine instructions.
Technical definition
PTX is a virtual instruction-set representation used in NVIDIA’s compilation toolchain, rather than the final machine instruction encoding executed by a GPU.
Engineering details
A compiler lowers PTX to architecture-specific code. Instruction availability, operand constraints, and generated sequences depend on the target architecture and toolchain. Similar-looking PTX operations can therefore produce different machine instructions or scheduling behavior across hardware generations.
Why it matters
Counting PTX instructions does not necessarily count executed hardware instructions. Low-level performance work should inspect the generated machine code and timing rather than infer throughput solely from source-level instruction names.
How to read it in InferenceX
The Blackwell port relates cp.async to LDGSTS and tensor bulk copies to UTMALDG for the measured paths. These mappings belong to that compilation context, not an unconditional rule for all PTX programs.