If no Windows support is planned for 5090 as it wasn't for 4090, pass along to corporate to mention this in advertisement

[NeedsMoar]

I didn't buy my 4090 for AI, it was more of a distraction, but I strongly recall TransformerEngine hardware from hopper being mentioned in at least a couple places, probably the arch whitepaper and I think somewhere on the product page.

I run Windows on this machine, and have no desire to install WSL2 and chew up the space, dual boot linux, or even enable the hypervisor due to the performance hit of having the OS running inside one which is small but existent.

If continuing lack of support for Windows (as per the closed bugs on the matter) is the policy for TransformerEngine which is the only easy way to make use of that portion of the hardware, it should probably be mentioned as an asterisk somewhere that only Linux is officially supported; without test framework in place there's not really any validation that custom user-written code for this feature works properly under the windows toolchain / drivers, and the build certainly didn't work for this project the last time I tried. It might work fine but given that fp8 kinda needs it to be as fast as it should be it's an important caveat for people buying cards for ML. I can only assume the same situation will occur with fp4 and blackwell when those cards are released shortly and it'd be nice of NVidia to inform people of it clearly. I doubt it'll drive anybody away from a purchase but it might save people time and an OS install if they're planning on doing a bunch of ML stuff and get the impression from everything else that Windows is fully functional with what might also be getting justified as a gaming card.

Personally I don't have much stake, I got bored with stable diffusion months ago and bought the 4090 to drive Karma and Renderman in Houdini and play the rare game which it's working great at; I'll probably get a 5090 too if they're not artificially price inflated by the time I check availability, but I remembered this from a while ago and thought I'd bring it up here. It might not be the place but I don't have the time to dig around looking for a corporate email on the NVidia site and I'd assume this team has some kind of contact with somebody who can mention an asterisk to marketing / docs. :-) Cheers.

[johnnynunez]

@NeedsMoar it's easy to have transformer engine for windows with wsl2 or natively.

I have running transformer engine for my 5090
https://hackmd.io/U6MULBYNRGixSu2nmehhZw

[johnnynunez]

Also Nvidia is working to support blackwell and fp4. Be patient
https://github.com/NVIDIA/TransformerEngine/tree/release_v2.0

[ksivaman]

@johnnynunez You're correct that this branch supports blackwell, but it supports mxfp8 and not fp4

[johnnynunez]

@johnnynunez You're correct that this branch supports blackwell, but it supports mxfp8 and not fp4

Yeah! But I mean.. in the future for sure :)
There must be an adaptation process and hard work in the libraries. Everything is just coming out—CUDA 12.8, CUTLASS, cuDNN 9.7, etc.