Hacker News Reader

Comment by PeterisP | Hacker News Reader

IMHO the requirements for that to happen would be a major slowdown in the "GPU race" where consumers consider that they have "enough" power and features, companies are not able to convince them otherwise and get pushed to commoditize and standartize the products as every product "ticks all the checkboxes", and one of those checkboxes happens to be good standards support. Alternatively, extreme fragmentation might force that, it happened decades ago when there very many video card manufacturers, but it's impossible now given the huge barriers of entry for competitive GPUs.

But while companies are able to differentiate their products feature wise (or even if they can't, but strongly believe they can), the incentives are set against it, so it's not likely to happen, the GPU vendors will push for vendor-specific aspects.

View on HN

I'd say raytracing was NVIDIA trying to distract everyone from the fact that their GPUs had in fact been 'enough' for a while. "RTX ON" was widely meme'd when the first RTX cards came out (edit: partly because at the time game publishers weren't really making good RTX games. Raytracing is mostly pointless if the game publisher doesn't bother to put a lot of effort into scene design, materials, and lighting. Control is a great example of a game that is gorgeous even without raytracing, and reportedly looks even better with raytracing turned on.)

I don't think the current state could be described as a "race".

NVIDIA makes good stuff and great drivers but charges a fortune (+$1500 for a high end GPU?!) AMD has made space heaters that happen to have video-out ports and their cards are useless for compute work because so many projects support CUDA and nothing else. Their current lineup cheats benchmarks by overclocking the cards and thermally throttling shortly after the time it takes to run a typical benchmark.

Edit: since I'm now being 'corrected': look at the TDP for same-generation, same-product-tier NVIDIA and AMD cards. AMD's cards have always used more power for similar performance.

People think certain GPUs "run hot" because

- they buy the cheapest GPU they can which means the heatsink and fans are terrible

- they don't adjust the stock fan curve (most GPUs these days don't even turn on the fans until they hit 60c and only max out the fans when the GPU die temperature is very hot because people care mostly about noise

- they don't have a well-ventilated case (note I said well-ventilated, not "stuffed with fans)

Buy a properly ventilated case, a mid-or-high-tier card from a reputable manufacturer, and adjust the fan curve. Find the temperature the card sits at during light desktop use and set the fans to kick in slightly above that. Then force the fans to run at full speed and benchmark, seeing what the die temperature maxxes out at. Set the fan curve to 100% at slightly below that temp.

NVIDIA GPUs also run extremely hot, and also overclock themselves when under any substantial load.

They even have a name for it: "GPU Boost"

I wouldn't say they run "extremely hot", my 1050 Ti powers both my 1440p144hz display and my 1080p75hz one, and it idles just over 32c for the most part. Under load though, like all GPUs, they will push their cooling to the limit to get as close to their junction temp as possible without causing lasting damage.

I was speaking to TDP. The actual temperature the card runs at is highly dependent upon room temperature, how good the cooling solution is on the card, how well-ventilated the case is, and of course the workload. Just because windows says the GPU is "100"% doesn't mean it's using maximum power; I've had games that make the fans go crazy at "100%" and games that at "100%" the fans never leave their default.

Every vendor makes multiple tiers of a particular GPU and you generally get what you pay for, though with diminishing returns for someone who isn't looking to overclock to the edge and get a few extra FPS. Good vendors starting a tier or two above their budget tier tend to produce cards with thermal solutions that are fairly quiet at stock speeds/voltages when gaming in a properly ventilated case in a comfortably cool room.

On 1xxx series cards for example, MSI's Gaming X line come with an absolutely massive heatsink. I had an MSI 1060 Gaming X that was overclocked and fully loaded the fans weren't noticeable, much less loud, and I had a terribly ventilated case.

My current card is a 1070 ti, from a reputable manufacturer and it's a mid-tier card. I have a pretty aggressive fan curve on it because I want longevity (capacitor lifetime drops appreciably with temperature.) Even then, I almost never notice the fans, but I have upgraded to one of the most well-ventilated $100-ish cases.

Stay well clear of single fan GPUs if you intend to do any gaming or compute work beyond "very occasionally." The fan noise will drive you nuts and the card will quickly thermally throttle.

If you care about your GPU lasting, lower the fan shutoff point to be more like 35-40c instead of the usual 60c, or just set them to run at the minimum speed all the time. Fans are easy to replace; caps aren't...and a quality card, the fans should be nearly silent at minimum speed.

> Just because windows says the GPU is "100"% doesn't mean it's using maximum power; I've had games that make the fans go crazy at "100%" and games that at "100%" the fans never leave their default.

There are even some games where the "GPU load" is 30 % but the card is running at 150 % power and destroying its VRMs :)

> Edit: since I'm now being 'corrected': look at the TDP for same-generation, same-product-tier NVIDIA and AMD cards. AMD's cards have always used more power for similar performance

Benchmarks disagree with that assertion:

https://www.techpowerup.com/review/nvidia-geforce-rtx-3080-t...

RTX 3000 series has consistently worse perf/watt than their AMDs RX 6000 series peers

Yeah, isn't AMD on TSMC 7nm vs. Nvidia with Samsung 8nm? Samsung has much higher power usage than TSMC from what I've heard.

Comparing the nm between different fab factories is not very useful and can be completely non-descriptive. (Not only mean different companies different things when they say Xnm, the things they mean are also often not at all describing the performance/characteristics of it well.)

> Samsung has much higher power usage than TSMC from what I've heard.

Is possible, but weather or not Nvideas bad power efficiency comes from that or architectural decisions in their design can't really be said.

> I'd say raytracing was NVIDIA trying to distract everyone from the fact that their GPUs had in fact been 'enough' for a while

"Enough" maybe for 1080p gaming, but 4k remains still heavy for the top tier GPUs even without raytracing; there are plenty of titles where 3080ti will not hit steady 60 fps on 4k, nevermind about 120 or 144 fps.

The same discussion will happen once 8k becomes mainstream, and then for 16k. In which case performance will never be "enough"

>I'd say raytracing was NVIDIA trying to distract everyone

That remains to be seen. The idea is that RT once fully adopted should be a lot easier for Graphics Designer while giving higher image quality. At the expense of GPU Compute.

Considering AAA Games are getting more expensive every year and Graphics Asset are by far the largest cost centre. I say on paper this seems like a very good trade off.

I dont know if 3D Models from Camera would shift the cost down. But I would really like to see some innovation in Graphics Modelling to bring the cost down by a few order of magnitude.

Reply on news.ycombinator.com