A developer has built OpenAssembler, a tool that reverse-engineers NVIDIA's closed-source ptxas compiler by treating it as a black box and decoding the GPU instruction format from its outputs.
Reverse engineering · GPU compilers
Decoding NVIDIA's closed compiler from the outside
OpenAssembler treats NVIDIA's ptxas compiler as a black box. By changing input bits and watching how the machine code shifts, it works out how the GPU instruction format is encoded, with no access to the compiler's internals.
The method
A repeatable black-box loop
Step 1
Alter input bits
Change the PTX input in controlled ways.
→
Step 2
Observe the output
Record how the native machine code changes.
→
Step 3
Infer the encoding
Map which bits control which instruction fields.
↺ Repeat and refine the mapping across more inputs.
What it reveals
An instruction encoding for a production GPU compiler, inferred without documentation
A way for independent developers to study the layer between GPU code and hardware
A systematic application of a standard reverse-engineering technique
What remains unverified
No independent verification, benchmarks, or NVIDIA statement
Unclear how complete the decoded format is
Unclear whether it yields working machine code, not just an описание of encoding
Key questions ahead
Can the findings be reproduced and turned into tools?
Reproducibility
Can others independently confirm the decoded format?
GPU coverage
How well does it span different GPU generations?
Practical output
Will it extend into code generation or analysis tooling?
Scope and licensing
What license terms and limits on analysis apply? Not addressed yet.
NVIDIA has not publicly responded to the work.
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article...