Nvidia Blackwell Architecture Speculation

  • Thread starter Deleted member 2197
  • Start date
Anyone can make a small cooler. The actual test is how well it cools while remaining silent. After my super quiet 4090, I'm never willingly going back to loud dual slot cooler if I can help it.
 
Anyone can make a small cooler. The actual test is how well it cools while remaining silent. After my super quiet 4090, I'm never willingly going back to loud dual slot cooler if I can help it.
Probably fine in big cases, but SFF cases might be another story. Ironically, it's the only "SFF" FE. I would be more worried about the ribbon cables for display connectors, which might lead to signal issues earlier than proper PCB routing
 
Obviously it remains to be seen how well the FE version dissipates heat, but in theory NVIDIA is not going to release a device rated for 575W but not able to handle the power.
So if the GPU is mainly power limited (instead of temperature limited), the bigger AIB versions are not likely to be overclocked much better, since they are also power limited. Even if they designed a "OC" version allowing more power it's not likely to be much higher (the 12pin connector provides up to 600W, plus the 50W from the PCIe interface, that's at most 650W).
Of course noise is also one potential problem. If the FE version needs to run the fans much faster then it could be much more noisy.
 
I mean it's obviously power limited. Why have a 575W TDP otherwise?
How often you would hit that limit in gaming on your particular PC configuration is a different question.
 
Power limited? At 575W?
Sorry, what I meant was the power slider option for overclocking. Typically FE parts offer either no positive adjustment at all or maybe a 3% uplift option. Aftermarket cards (such as my ASUS Tuf Gaming 4090) offer an 11% power uplift slider, I know others offer even more. That's what I was referring to. I suspect the cooling is entirely sufficient for the rated 575W given the teardown interview I watched from GamersNexus.

My 4070 Super cards, my 3080 Ti, and my 4090 are all overclocked + undervolted at all times, however even with a pretty heavy handed undervolt, my 4090 can still run out of power headroom to sustain >2950MHz clocks (eg not counting the blips where it clocks up and then immediately backs out thanks to hitting the power limit...)
 
So ~20% faster on average in games for 80% more bandwidth and ~30% more flops.

It’ll be interesting to see which games if any push those numbers higher. I expect hardware utilization on the 5090 will be atrocious.
 
  • Like
Reactions: KOF
So ~20% faster on average in games for 80% more bandwidth and ~30% more flops.

It’ll be interesting to see which games if any push those numbers higher. I expect hardware utilization on the 5090 will be atrocious.
My 4090 is GPU limited @4K (99% load) when I am done tweaking settings (turning stuff up to 11 or in "CyberPunk 2077" just max settings), will be fun to see what changes with the 5090 but my bet is that I will find settings to tweak and again be GPU limited as I game @ 4k, not 1080p.
 
My 4090 is GPU limited @4K (99% load) when I am done tweaking settings (turning stuff up to 11 or in "CyberPunk 2077" just max settings), will be fun to see what changes with the 5090 but my bet is that I will find settings to tweak and again be GPU limited as I game @ 4k, not 1080p.

GPU load is a pretty misleading metric though. It just means your card is doing “something” not that it’s fully utilized. I can guarantee your 4090 is not at 99% load.
 
Yah, power draw is probably a better way to estimate real utilization.

Fully loading a GPU will probably raise power draw far beyond TDP limits. E.g. a synthetic async workload that can stress raster/texture/shader/memory/RT all at once. Most games use just a fraction of all those units and rarely at the same time.
 


Q4. With MFG, the core technology of NVIDIA's new DLSS 4, you'll now be able to generate multiple frames instead of just one. This time, you can generate up to 3 frames, but how many frames will you want to be able to generate with AI in the future?

A4. DLSS 4 MFG technology is targeted at 4K 240Hz. It's hard to say how many more frames you can generate, what the technical constraints will be, and what the future will look like. At the end of the day, it's all about balance. We can generate 16 frames at a time, and we'll do so if it ultimately benefits the gaming experience, but for now, we've decided that a maximum of three frames is appropriate.
 


Q4. With MFG, the core technology of NVIDIA's new DLSS 4, you'll now be able to generate multiple frames instead of just one. This time, you can generate up to 3 frames, but how many frames will you want to be able to generate with AI in the future?

A4. DLSS 4 MFG technology is targeted at 4K 240Hz. It's hard to say how many more frames you can generate, what the technical constraints will be, and what the future will look like. At the end of the day, it's all about balance. We can generate 16 frames at a time, and we'll do so if it ultimately benefits the gaming experience, but for now, we've decided that a maximum of three frames is appropriate.
I think someone said it before -- the GPU should be generating as many frames as needed to match the refresh rate of any sample-and-hold monitor (unless the monitor supports BFI, see below). The problem is that sample-and-hold is fundamentally broken. Your sample-and-hold monitor is already generating frames if you're sending a 60 fps signal to a 240Hz display -- it's just repeating the old frame 4 times. It's the most brain-damaged form of frame-generation and you use it all the time.

BFI changes the equation because it kills the hold prematurely, allowing your brain to interpolate between frames (which somewhat emulates CRT behavior). You do lose brightness. I think in practice some mix between GPU-FG and BFI is going to be the ideal recipe.
 
Back
Top