evernessinceI would absolutely avoid integrating CUDA libraries in anything as critical as Cyber-security after Nvidia pulled the rug on all 32-bit CUDA applications, libraries, plugins, etc. on 5000 series and later without providing a fallback. Absolutely destroys trust for companies developing long term solutions.

igormp32-bit CUDA support has been announced as deprecated since around cuda 9, over 7 years ago.

igormp5000 series are pretty irrelevant in this regard, and Ada-based products and older should still have a long support window.

Real-time inference: AI models running on GPUs can analyze network traffic in real time to identify zero-day vulnerabilities

evernessinceThe 5000 series are the first Nvidia cards unable to run 32-bit CUDA code since 32-bit support was added. The fact of the matter is they just recently removed a feature from the GPU without providing any fallback. Even Microsoft knows better than this, hence why you can't still run 32-bit windows apps despite the OS not natively supporting them since windows 10.

evernessinceGiven that the 5000 series is the first generation unable to run 32-bit CUDA code, I'd say it's 100% relevant. You seem to be under the impression that the last change to 32-bit CUDA status was years ago but as pointed out, that's not the case.

evernessinceAside from gaming, certain industries such as medical imaging, finance, and manufacturing use specialized 32-bit CUDA software that has been stable for years.

evernessince(as limiting to older hardware only means they can only acquire from a constantly shrinking pool) or cost them a lot of capital in replacing the system unnecessarily.

unwind-protectHow do you train ML models on exploits for holes you don't know about yet?

unwind-protectHow do you train ML models on exploits for holes you don't know about yet?

igormpThat "feature" has been deprecated for quite some time, as I said already. Within CUDA 9 it was already not recommended to build 32-bit stuff.

igormpIf for some reason you have stuff that relies on CUDA 9 and 32-bit, the fact that 5000 series do not support 32-bit is actually irrelevant, because the newest GPU you can get CUDA9 stuff running is Volta. It simply won't work with Turing and newer.

igormpAnd by the point your newer GPU with the newer current compute capability is not supported by your previous CUDA version, you'll likely need to rebuild it against a newer CUDA version, which also makes it a non-issue to jump into 64-bit.

evernessinceYou don't seem to be in tune with how software development works

evernessincewe've already seen examples of modern software still using 32-bit CUDA code.

evernessinceCase in point, it's the reason the 4090 is faster than the 5090 in PassMark Direct Compute:

evernessinceNo, you can in fact run 32-bit code on up to the 4000 series. Have you not been following this issue at all? There are dozen of examples of 32-bit PhsyX running on the 4000 series.

evernessinceOk so you clearly have not developed software at a professional capacity if you think jumping from 32-bit to 64-bit is easy. Only in rare cases would it be trivial to recompile your code with minimal work for 64-bit CUDA. First off, you are making the assumption that every bit of code, every library, and every plugin that utilizes 32-bit CUDA; the devs have the source code for. That is exceedingly rare. You are also assuming that the toolchains being used also support 64 bit CUDA. Their build pipelines, testing frameworks, and deployment strategies need to accommodate 64-bit binaries.

evernessinceLast but not least, the code needs to be combed through and tested for the following issues (and this is not an all-inclusive list):

Pointer size differences, Data Structure Alignment, Casting Issues, Kernel optimization changes, Register pressure (64 bit operations tend to consume more registers), Mixed precision issues (if you mix different data precisions, behavior may change in a 64-bit environment), etc.

All the fun little bugs that cause big headaches, particularly if we are talking about well aged software. Software of which may very well have been written by now retired wizards (aged developers) using a variety of programming languages or approaches whose knowledge has been lost to the ages, thus further complicating anything short of starting from scratch.

There's really no telling how many bespoke applications are out there that rely one way or another on code that has now been deprecated. It's beyond naive and harmful that you think it's a simple matter to update the code in these instances, especially considering we are talking about the medical, science and engineering fields that use CUDA.

evernessinceHence why fallbacks are typically provided for deprecated APIs, middleware, etc. That's just standard practice. Imagine if Microsoft had just told people to take a hike or to reprogram every 32-bit app when it officially dropped 32-bit support instead of what they did in including an emulation layer. They would have hemorrhaged businesses and customers.

evernessinceInference is running an already trained AI model so in this case they are not talking about training.

Things covered by this article would be running the AI to detect known threats / exploits in addition to analyzing traffic for IDS and IPS. It's pretty much replacing existing algorithms that do that with AI.

unwind-protectI *am* talking about training. I am asking how NVidia trains for as-of-now unknown exploits.