> For the complete documentation index, see [llms.txt](https://brindha.gitbook.io/mylearning/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://brindha.gitbook.io/mylearning/7-days-genai-learning-challenge/is-just-true.md).

# Is Just True

\##1\
This is the moment I decided to go against the grain.
\
Everyone is abstracting GPU programming away. More Python. More layers. More distance from the metal.

I'm picking the other direction with my SASS King project.
\
I don't know if it is a wise choice but future will tell.

SASS is NVIDIA's native GPU machine code.
\
What actually executes on the hardware.
\
And it's a black box.

CUDA C++ has documentation.
\
PTX has documentation.
\
For SASS, the last systematic public work is Jia et al. (Citadel) on Volta and Turing, 2018-2019.
\
Nothing for Ampere. Nothing for Hopper. Nothing for Blackwell.
\
The goal is to close that gap.
\
Five architectures. Three major ruptures to analyze and document.

Three tracks of methodology:
\
Top-down: pattern recognition on real kernels, SASS dumps side by side.
\
Bottom-up: microbenchmarks per instruction, latency, throughput, dual-issue rules.
\
Compiler analysis: how ptxas translates source into actual SASS.

Claude accelerates the iteration loop on kernel variants.
\
I've already analyzed over 20 kernels.
\
Target: a thousand.

My objective: understand SASS entirely and document every compiler behavior.
\
In a world where cost and sustainability are the a primitive concerns, I want to contribute at the lowest level possible.\
\
\--

The core claims in your post are well-grounded. Here's a factual breakdown:

**What's accurate:**

* SASS is the low-level, semi-documented machine code generated when compiling CUDA source code with nvcc or when translating PTX with ptxas. "Semi-documented" is the key word — it's real and the gap is real.
* PTX is NVIDIA's virtual ISA compiled by ptxas into native SASS. PTX has forward compatibility — a CUDA program compiled to PTX years ago can still run on modern GPUs like Blackwell — but it may not exploit the latest hardware features efficiently. This is exactly why SASS matters: PTX abstracts away the performance-critical details.
* The architecture ruptures you describe are real. Ampere added 8 instructions and removed 2. Hopper added 36 and removed 1. Blackwell added 48 and removed 8. Those are large, substantive changes across generations.
* The documentation gap is real. While NVIDIA does publish instruction tables in its binary utilities docs, the coverage is limited to listing valid instructions per architecture — no latency, throughput, dual-issue rules, scheduling constraints, or encoding details.
* Some recent work is closing parts of the gap. A 2025 microbenchmarking study on the Blackwell architecture studied latency, throughput, cache behavior, and scheduling details, including new SASS-level instructions like OMMA and QMMA for FP4/FP6 datatypes. And SemiAnalysis spent months benchmarking Blackwell PTX and SASS instructions, noting that no detailed public datacenter Blackwell microbenchmarking study existed until their work.

**Where to be slightly precise:**

The Jia et al. (Citadel) framing as "the last systematic public work" is largely fair, though there have been partial follow-ups. A paper on Ampere did measure PTX-to-SASS mappings and clock cycle latencies, explicitly noting that no such studies had been done on Ampere at the time. So the field moves in patches, not systematically — which actually reinforces your point rather than undermining it.

**Bottom line:** Your description of SASS as a black box with a meaningful documentation gap across Ampere, Hopper, and Blackwell is accurate. The recent Blackwell microbenchmarking work shows the field is starting to move, but it's fragmented and focused on ML workloads — nobody has done what you're describing (systematic, architecture-spanning documentation of compiler behavior and ISA semantics). The gap is real. The project is well-motivated.