Physix FrontierTesting if non-CUDA stacks can handle medical imaging
A friend recommended a new stack claiming CUDA's moat is gone, so I decided to test how usable it actually is. I've been tinkering with GPU deployments for three weeks and happen to have a small CT image segmentation demo on hand—a similar workflow I kept from my time doing AI imaging at United Imaging. The model isn't big, but the input consists of continuous slice sequences, making it sensitive to VRAM, scheduling, and logs. I figured I'd run it through to see if the product is deployable beyond just loading the model.
First, compare two environments. One is the old familiar CUDA, NVIDIA's GPU computing ecosystem: install drivers, configure environment, confirm GPU availability—basically the same as before. No surprises, but problems are relatively predictable. VRAM usage is visible, error stacks are readable, and crashes usually indicate whether it's memory shortage or kernel failure. The other is the new stack. That recent video discussed how CUDA's moat is being diluted by open backends and compiler ecosystems. I tested on my local workstation; running device queries in the terminal showed the GPU recognized as a compute device. The first surprise was that model import wasn't as troublesome as imagined; exported weights and intermediate formats were recognized. When the "device ready" prompt appeared, I almost thought it was done.
But the real pitfalls emerged during data processing. My tests showed it took about forty minutes from setting up the environment to running the first CT sequence, with most time spent on dependency matching. The first result wasn't bad speed-wise, at least not toy-like. The second also worked. On the third, the process hung directly. Fans spun up, but the terminal showed no errors and no locatable traceback—the kind that tells you which line crashed. After restarting, reproduction attempts with the same data sometimes passed, sometimes failed. This confirms that complaints in tweets aren't exaggerated: drivers are still unstable, lacking debugging entry points during crashes or hangs, and unable to even dump GPU status. For developers, this is a black box.
From a product perspective, running successfully is just the first layer
If looking only at technical demos, the new stack has made progress. It lowers the barrier of requiring specific GPUs. For AI product people, this means procurement and deployment have a second option, avoiding binding all models and inference machines to a single path. Especially for lightweight imaging tools, teaching demos, edge boxes, or scenarios where budgets don't allow high-end cards, it's worth continuing to follow.
But in medical imaging, I immediately ask: Has clinical validation been done? Hospital procurement cares more about actual doctor feedback: Can it connect to PACS? Can reports be written back? Does clicking in the reporting station yield results within tens of seconds? Are there logs if it fails? Can omissions be traced back? Who is responsible when issues arise? The most painful aspect of the new stack currently is its lack of auditability. Crashes have no state dumps, hangs have no reliable reproduction paths. Delivering this to hospitals is like driving a car without a dashboard into the ICU. No matter how accurate the model is, if no one dares to sign off in the workflow, the product cannot be deployed.
So my conclusion is it depends. It's suitable for R&D sandboxes, personal tinkering, budget-sensitive multi-platform verification, and for those wanting to observe if CUDA's moat is truly broken. It's not suitable for hospital delivery, especially when compliance logs, doctor signatures, and remote O&M are required. Pushing doctors' workflows onto it directly is even worse. Action advice is simple: Don't put production models on it first. Start with a desensitized small dataset, run it continuously for a few days, and record startup, crashes, logs, and recovery. Only discuss replacement if it runs stably. I dare not conclude whether CUDA's moat has been dismantled, but in my tests, the door to hospitals hasn't opened yet.
📌 This article is compiled from Hacker News. Original video: https://www.youtube.com/watch?v=TiRxcPQNcBA
Copyright belongs to the original authors. This is a compilation and independent analysis based on public reports.
