Hacker News Reader

Comment by DannyBee | Hacker News Reader

I'm not sure who this is really for. I mean that honestly.

At least on the workstation side - i do a lot of solidworks-based 3d modeling. The A6000 can easily, for example, raytrace in real time anything i can even find to throw at it. Like models that Solidworks still has trouble opening and rendering normally can raytrace instantly.

What part of workstation software is still GPU bound at this point on high end GPU's?

I guess if CAM was taking more advantage of GPU's, i could see it useful there, but on the modeling side, i honestly don't get it.

View on HN

Color correction uses as much GPU memory as it can get. I used to maintain DaVinci Resolve systems with expansion PCIe chassis to handle all of the exapnsion cards. Each year, we'd upgrade the GPUs with more cores and memory. Each upgrade allowed the system to be much faster. However, this used to be on Macs, so an Nvidia card is no longer useful on that platform. Might still be a cool thing to play with under WinX/Linux

As someone who is totally ignorant about film color grading, why does it require so much VRAM?

I’ve done color correction in games and it’s computationally very cheap to throw a LUT in there, so clearly something else is going on?

First off, throwing a LUT in there is not color grading!!!!!!!! People making this claim are appreciated as it lets me know immediately that they are not familiar with the subject at all. It's like a shibboleth. Edit: I mean this in the nicest way. It's just a pet peeve of mine. You already stipulated you didn't fully grok the topic.

In a full color grading session, you can have power windows. You can have those power windows tracking. You can have noise reduction. You can have other filters. You can do so so much. All of that takes up memory. You can be dealing with 8k footage. You can have many many many layers of that footage. You can be grading in >16bit color space.

You can do all of that with very little VRAM at a snails pace, or you can do it in real-time with all of that VRAM and keep the client in a supervised session happy. With all of that VRAM, you can export at faster than real-time for delivering the content to the client at the end of the session rather than the next day after an overnight render.

Access to more memory opens up higher resolutions in ML video and image workflows. I am pushing close to these numbers for GPU memory on the 64G M1 Max.

What are you using to build your models out of interest? I've no idea what is optimised for machine learning on M1 chips yet?

All of the responses are about game development, and that sort of thing which I totally understand. but all of their quotes are from like cad developers. so I definitely feel like they are mistargeting at this point

The A6000s have become very popular for ML engineering, especially in the NLP space since those models are enormous. 48GB VRAM and 2x faster training time is tempting for those applications!

With 16GB I can easily run out of VRAM and encounter serious slow downs when working with multiple 4K texture sets in Substance Painter.

That's working on Hero assets for current-tech games. UE5 Nanite assets and film assets are much heavier.

I totally buy the game developer need here, as well as some of the live video processing folks. but that convinces me even more that these are totally mistargeted given their sales points

Photorealistic renders for catalogs and architectural visualization can still take many hours on the best GPUs available (if the scene even fits into VRAM). Even a slight performance improvement can be worth a lot of money for companies doing this kind of work.

Although an small audience, this is exactly the kind of card that live video processing (IMAG) needs. Software such as Hippotizer, Pandora's Box, Resolume, and vMix rely heavily on GPU.

That's not really ray-tracing. That's just regular GPU rendering with physically-based materials and some ray-tracing effects for extra chrome.

software grows with the hardware, in a few years your A6000 will be barely good enough to play Battlestation 2096

Reply on news.ycombinator.com