The new NVIDIA RTX50 series gaming cards have been announced.
I know the dev team recommends the use of pro cards so in theory none of these will be any good for Flame. Still, the RTX5090 has some impressive specs for a gaming card.
Did you watch the keynote and the live render demo at the beginning?
By all appearance this card is not just a generational upgrade, but sizable step function. In particular the integrated use of ray tracing and inference. Out of the live render demo, ray tracing only rendered every 4th frame, the other three were inferred by the GPU with the use of a purpose trained model.
For some use cases that can be a significant change. Unclear what the use cases outside of gaming might be?
One scenario that could get around bandwidth limitations of streaming services could be a ‘lossy’ high compression codec. Train a model when releasing some content. Then when a consumer starts streaming, it downloads the model, and every 4th frame from the server. Your TV will infer the other 3 frames locally, leading to higher image quality at a fraction of the bandwidth.
Of course it would absolutely kill the control freaks that obsess over every pixel. You couldn’t actually QC this content, since every stream will every so slightly variable.
Interesting times.
1 Like
I watched some highlights and saw that. Really got the impression that that was more for gaming than streaming but so so interesting.
The price for the RTX5090 is insane when comparing it to a RTX5000 ada. 1/3rd the price but multiple times faster. Why is it that Flame needs to utilise the pro cards again?!!
It is recommended but not mandatory.
I’ve been using high end gaming card on flame for a while and haven’t faced any drawback, it is quite stable and reliable (In my case), in fact, I’ve never had a OoM gpu crash like I’ve heard from Alan with quadro card. And you may ask why i’m using a gaming card instead of a quadro? simple, it is faster. Also because I like to use some AI tools which almost all of them rely on the gpu speed because of the tflops at certain precision requierements (fp16, fp32), sometimes more than the amount of vram your gpu has.
RTX 4090 inference is faster than RTX6000Ada even though rtx6000 has twice amount ram. (They both are same latest generation) You can even have 4x tesla M40 24GB in SLI (now NVLINK) running a LLM and they run slower than a single RTX 4090 24GB.
It is well known that “pro” quadro and tesla cards are slower than xx80/xx90 gaming cards. The “pro” cards benefits are nothing but the amount of (ECC) ram and that they have way less power consumption than high end gaming cards…so you got to choose between performance and single ram.
You can also combine/stack multiple gpus and get a wild performance, like many Resolve and Nuke users do…why? because time is money.
I’m planning to buy one those new 5090 and see the beast performance they have. Really sure it will make me scream.
The era of QUADRO ONLY has gone, like Lustre many years ago…It is a bit pitty that this still happens but it is what it is!
1 Like
I assume it’s less the hardware and more the driver that’s the difference. But some of this may be hangover from times past.
Back in the days of the 2090 and Quadro there were actual hardware differences that mattered. The Quadro had two copy engines, the GeForce only one. For gaming where the data flow primarily one way that’s no issue, but if you load textures and them bring them back to copy them to DeckLink the 2nd copy engine enabled different dataflows.
No longer the case as of the A5000.
I mean, my TV already does that and it takes me ten minutes of menu diving to shut it off.
As a medium-skill gamer, I don’t know that frame interpolation is ideal. “AI interpolation” feels like marketing hogwash to me. I mean, I trust it does what it says, but is it good for games? It seems like it would be a real problem if it’s being rendered at 15fps but temporally upscaled to 60.