Will there be 300W Discrete GPUs in 5 years? 10?
- Thread starter Jawed
- Start date
Many years ago, I had a 68000, a Blitter, a Copper and some other co-processors too. Today I have a i7, a PCIe extender and many GPUs.
The history (and my personal crystal balls) tells me that, in 10 years, I will still have co-processors around my Central Processing Unit (may be not 300W discrete GPUs but some other kind of accelerator/co-processor).
The history (and my personal crystal balls) tells me that, in 10 years, I will still have co-processors around my Central Processing Unit (may be not 300W discrete GPUs but some other kind of accelerator/co-processor).
A1xLLcqAgt0qc2RyMz0y
Veteran
This has been answered many times already.
The GPU on the SOC will not be 300 watts because the CPU cores and the North/South bridge and other devices on the SOC will take many watts.
Thus a 300 watt SOC will always under perform a 300 watt dGPU.
Not really, the cpu and gpu will need to communicate a lot once programming models adapt to take advantage of the silicon. Eventually, there won't be a reason to have a discrete gpu due to latency between the 2 processors. No matter how powerful your 300 watt gpu is, it won't keep up with the interconnect of share silicon.
Maybe if quantum coprocessors become a thing but everything will become integrated eventually.
I think VR will drive the limits of GPU technology for a long time yet to come. Certainly more than 5 more years and almost certainly more than 10. A minimal VR experience of 1280 x 1024 at 90 fps (and only 90 deg fov) (minimum - since vsync must be on and frame drops are not acceptable - the framerate must remain silky smooth to maintain presence) per eye is just barely possible right now with very carefully built games using the highest end descrete GPUs (and ideally with 2 of those GPUs, one per eye) and a lot of short cuts. To reach retinal quality in VR you need to achieve a resolution of about 8k x 8k per eye and ideally at a framerate closer to 120 fps (minimum again with vsync enabled) with the fov reaching 180 to 270 deg. That's about 2 orders of magnitude more GPU power just for the extra resolution. Then there is the enormous amount of extra processing need to make "higher quality" pixels, not just more of them, such as full real-time global illumination, cinematic quality rendering and realistic special effects, full 3d geometry (normal maps, and most other graphical shortcuts fail to work well with VR, you need actual 3d geometry everywhere), etc. On top of all the graphical improvements, there is the necessary improvements in physical and character simulation which will take everything you can throw at it for a long time to come.
As a comparison between a real-time game and a typical CGI film "Cars 2" required 12,500 concurrent CPU cores running to render the film at 11.5 hours per frame. We have a long, long way to go to get to photo-real CGI quality at 8k x 8k at 120 fps per eye.
As a comparison between a real-time game and a typical CGI film "Cars 2" required 12,500 concurrent CPU cores running to render the film at 11.5 hours per frame. We have a long, long way to go to get to photo-real CGI quality at 8k x 8k at 120 fps per eye.
Back in 2013-2014 the whole cpu gpu conectivity related computing for gaming was examined quite in depth for months and there are a number of articles, exploration of what it could mean for the nextgen consoles. Overall these is some benefit but not a whole lot in terms of raw performance. One of the big drawbacks is reduced memory bandwidth not just because of splitting of mem bandwidth between cpu and gpu but the additionally memory contention issue that plagues unified memory system.
The only suitable memory for a gpu would be gddr solutions or stacked memory such as hybrid memory cube or high bandwidth memory. Gddr5 would have to be soldered to the motherboard, not sure about stacked memory.
Also there is the problem of the corporate landscape. I don't see Nvidia ever eliminating its discrete lineup of gpus because they formed some partnership to put their designs and technology on intel chips. And I don't see intel creating gpus more powerful than Nvidia & AMDs discreet offerings for in a single intel cpu/gpu chip solution, their larabee and knights-falling aren't exactly designed to be competitors against nvidia and amds best in the gaming scene.
When PS5 & Xbox6 (Redmonds brilliant naming strategy to surpass 5 from ps5 on their 4th Xbox console) launch in 2019-2021 they will continue to push the graphics limits on those systems, meaning there will still be games that run at 30fps to achieve maximum fidleity, in a $400-$500 pricepoint console and (and like this generation only $100-130 of which is on the cpu/gpu silicon), meaning there is alot of room for better performance from more expensive silicon, for those willing to spend $250+ on discreet gpus.
As the ps5 & Xbox6 age by 2022-2023 with decreasing costs to manufacture at smaller nodes at the various fabs, we will see discreet gpus continue to gain performance at all price points, while console performance remains static.
Last edited:
So make the SoC (300+however_many_watts_you_need_for_the_CPU_plus_I/O) watts.
The SOC would have to be physically quite large to effectively dissipate 400+W. Doesn't seem viable for the vast majority of SOCs which if anything will get smaller with time.
I don't think anyone expects 300+W products to be the vast majority of anything.
A1xLLcqAgt0qc2RyMz0y
Veteran
No doubt, I was simply responding to the post "So make the SoC (300+however_many_watts_you_need_for_the_CPU_plus_I/O) watts."
And a dGPU with the same power envelope will again outperform it.
Last edited:
That's what I'm saying. dGPUs will be with us for many years for the same reasons we have them today.
But then the platform would use more power. At the same platform power, there's no reason why a dGPU should be faster.
A1xLLcqAgt0qc2RyMz0y
Veteran
Why are you moving the goal posts?
No mention was ever made about platform power only that a 300 watt SOC would make a 300 watt dGPU obsolete which is false because a discrete CPU and a discrete GPU would always outperform it.
I'm not moving anything: platform power is what matters. I don't know about you, but my electricity bill doesn't have a separate column for "dGPU energy consumed".
You're right of course. The 12500 CPU cores was the size of the render farm. They obviously computed more than one frame at a time on the render farm. A better characterization of the amount of computation it takes to make a CGI film is to compare the total render time (in render-machine hours) for the entire film vs the running time for the film and the difference in frame rates between film and VR.
Here is a link to the article: http://www.datacenterknowledge.com/archives/2011/07/11/turbo-charging-digital-render-farms/
The article mentions Kung Fu Panda 2 requiring 55 million render hours for a 91 minute film (obviously per rendering machine). That means that it would take a single one of their rendering machines 36 million times longer to render the movie than to watch it (at 24 frames per second) or 180 million times longer to render it than to watch it at 120 frames per second, and thats assuming typical film resolutions (~2k x 1k in this case), though it was rendered in stereo 3D. As I mentioned above, GPUs have got a very long way to go to reach VR retinal quality at 8k x 8k x 120 fps per eye for even Kung Fu Panda 2 CGI quality, let alone complete interactive photo realism.
