CASS: Nvidia to AMD Transpilation with Data, Models, and Benchmark
The paper introduces CASS, the first dataset and model suite for cross-architecture GPU code transpilation (CUDA to HIP and SASS to RDNA3), which achieves state-of-the-art translation accuracy and performance preservation while outperforming commercial baselines like GPT-5.1 and Hipify.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine you have a massive library of instructions written in Nvidia's secret language (CUDA). These instructions tell powerful computer chips (GPUs) how to do complex math for things like AI, video games, and scientific simulations.
The problem? Nvidia's chips only understand their own language. If you try to run these instructions on an AMD chip (a different brand of computer brain), it's like trying to play a PlayStation game on an Xbox. The hardware is different, the "dialect" is different, and the game just crashes.
For years, companies have been stuck with Nvidia because rewriting all their code for AMD is a nightmare. Existing tools are like clumsy translators that often get the meaning wrong or miss crucial details.
Enter CASS (CUDA-AMD Assembly and Source Mapping). Think of CASS as a super-smart, bilingual architect that doesn't just translate words, but understands the blueprints of the building itself.
Here is the breakdown of what they did, using some everyday analogies:
1. The Problem: The "Tower of Babel" of Chips
Nvidia and AMD speak different languages at the lowest level.
- Nvidia (CUDA): Uses a proprietary language called SASS.
- AMD (HIP/RDNA3): Uses a different language called RDNA3.
- The Gap: Existing translators (like a tool called Hipify) are like a dictionary that only translates the surface words. They miss the deep grammar rules. If you try to translate a complex sentence, they might get the meaning right but break the sentence structure, causing the code to fail.
2. The Solution: Building a Massive "Translation School"
To teach a computer how to translate perfectly, you need a huge textbook of examples. But no such textbook existed for GPU code.
- The CASS Dataset: The researchers built a massive library containing 60,000 pairs of instructions.
- One side is the original Nvidia code.
- The other side is the perfectly translated AMD code.
- How they got it: They didn't just copy-paste. They built an automated robot factory that:
- Scoured the internet for existing code.
- Used AI to invent new types of code (like making up new recipes for a chef) to ensure they covered every possible scenario.
- Translated them, compiled them, and tested them to make sure they actually worked on both types of chips.
- Analogy: Imagine a chef who writes 60,000 recipes in French, then hires a team to translate them into Italian, cook them, and taste-test them to ensure the Italian version tastes exactly like the French one.
3. The Teacher: The CASS Model
Once they had this "textbook," they trained a specialized AI model (a Large Language Model) on it.
- The Result: This model learned the deep grammar and "muscle memory" of both chip languages.
- The Score:
- Source Code Translation (High-level): It got 88% accuracy. (Existing tools and even super-smart AI like GPT-5 only got around 66-85%).
- Assembly Translation (Low-level): This is the hard part—translating the raw machine code. The model got 69% accuracy.
- The Competition: The best commercial AI models (like GPT-5.1 or Claude) only scored around 22% on the low-level translation. They were basically guessing. CASS was like a master mechanic compared to a guesser.
4. The Proof: Does it Actually Work?
Translating code is useless if the result is slow or breaks.
- Performance: The researchers tested the translated code. In 95% of cases, the AMD chip ran the translated code just as fast and used the same amount of memory as the original Nvidia code.
- The "CASS-Bench": They created a final exam (a benchmark) with 369 different tasks to prove their model works. It's like a driving test where the car has to navigate 18 different types of terrain (math, graphics, AI, etc.).
Why Does This Matter?
- Freedom from "Vendor Lock-in": Right now, if you build a supercomputer with Nvidia chips, you are stuck with Nvidia forever. If you want to switch to AMD (which might be cheaper or more efficient), you have to rewrite everything. CASS makes that switch possible without the headache.
- Democratizing Hardware: It allows researchers and companies to shop around for the best hardware without being trapped by one company's language.
- The Future: This isn't just about Nvidia and AMD today. It proves that AI can learn the "deep physics" of computer hardware, paving the way for future tools that can translate code between any type of computer chip.
In a Nutshell
The paper introduces CASS, a new system that acts as a universal translator for computer chips. By training a specialized AI on a massive, verified library of code, they taught it how to speak both "Nvidia" and "AMD" fluently. The result is a tool that can take code written for one brand of graphics card and instantly make it run perfectly on another, saving companies millions of dollars and unlocking new hardware choices for everyone.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.