2016+ CPU/GPU News

[revelator]

oh boy it had to happen https://www.techspot.com/review/2652-dlss-vs-fsr-performance/

mighty big can o worms opened in that review hehe.

[revelator]

so turns out FSR is just as fast as DLSS in a few cases a little faster DLSS looks better though.

FSR does have the advantage that it can be used by pretty much any gfx card though a gtx 10 series is probably the oldest card where it makes sense. great news for 1080 ti users as it actually makes raytracing viable on these cards.

 

[jaxa]

I think more people care about the quality than the performance, and one of the assertions is that DLSS is better than FSR when working with a low input resolution. If you're doing something like turning 1440p into 4K (~66.7% render scale), any upscaling technique will work well. Then frame generation is a whole 'nother can of worms.

[revelator]

i agree was a bit fun to see the DLSS entusiasts going ape over comparing FSR performance on a RTX card though, omg sacriledge . on the older gtx models it also hurts some that these cards dont have any RT cores for raytracing but it does make some games who uses light raytracing playable at 4K.

not really sure what to make of frame generation, guess we will have to see how much of an impact it makes and how visible it will be.

[revelator]

hmm my guess is after reading some about it would be that the tensor cores would use data from the previous frame and the current frame and then try to guess pixel positions using AI to fill in to the generated frame ? i wonder if it would be all that visible but some minute corruption might be possible to notice if the AI guesses wrong (scenes with high movement).

kinda reminds me of how we converted ntsc to pal here in europe though a whole lot more complex.

ntsc to pal conversion was originally done by playing the film fast at 25 fps the two sequential fields were created from the same frame and displayed 1/50 of a second apart, the conversion is then done by repeating every 5th field with a process called 3:2 pulldown.

in essence the frames are roughly doubled and most of them contain the same data as the previous frame, this is in most cases inpossible for the human eye to notice

[revelator]

ah allmost right https://nvidianews.nvidia.com/news/nvidia-introduces-dlss-3-with-breakthrough-ai-powered-frame-generation-for-up-to-4x-performance seems it goes a bit farther

[revelator]

hmm according to reviews it seems to only have a positive affect if the card can allready do high framerates > 100 fps

on entry level cards such as the 4060 and 4060 ti it actually causes quite a load of visual artifacting also affects the midrange 4070 and ti models though not as pronounced even the 4080 is not completely free from it only the 4090 seems to have the nessesary power to let it really shine. newer drivers might real in some of the artifacting though i suspect anything less than a 4080 might not cut it if you want flawless visuals.

[revelator]

doh just found out my board does not support rebar by default, i had to use a hack with the rebar code injected into the bios and a small tool to enable it in windows it does help quite a bit though.

[nbohr1more]

Yep, it's a race:

Who will win:

Silicon graphene hybrid ( already demoed high volume low defect production, not sure about remaining obstacles )

Boron Arsenide ( still has high defect rates )

High-NA EUV ( uses silicon, expected to have medium to high defects initially, will probably be adapted to use alternate substrates like those above but will take more time )

[jaxa]

High-NA EUV is obv independent of the substrate. It's going to get used regardless. The scenario where it doesn't get used is if an older process node could be used, which was one of the ideas behind nanotube-based 3DSoC.

So, the brand new Ryzen 8000G desktop APUs are out. The consensus is they don't make sense compared to CPU + dGPU:

https://www.tomshardware.com/pc-components/cpus/amd-ryzen-7-8700g-cpu-review
https://www.techspot.com/review/2796-amd-ryzen-8700g/
https://www.anandtech.com/show/21242/amd-ryzen-7-8700g-and-ryzen-5-8600g-review
https://www.phoronix.com/review/amd-ryzen7-8700g-linux
https://arstechnica.com/gadgets/2024/01/ryzen-8000g-review-an-integrated-gpu-that-can-beat-a-graphics-card-for-a-price/
Gamers Nexus video review

Neat though.

Edited January 30 by jaxa

[revelator]

Yeah seems the concensus has been that these new techniques for getting gfx performance are being used as a bit of a crutch instead of working toward better hardware capabilities, i kinda agree.

the new hybrid materials could be very interresting to that effect down the road though.

[jaxa]

1 hour ago, revelator said:

Yeah seems the concensus has been that these new techniques for getting gfx performance are being used as a bit of a crutch instead of working toward better hardware capabilities, i kinda agree.

I don't think they aren't delivering better hardware. 8700G is obviously much better than 5700G, Nvidia seems to be pushing top-end by at least 60% per gen. AMD might stagnate with RDNA4 but deliver 4080-ish performance at much lower $400-600 prices.

Most of the reviewers are testing without upscaling or are being transparent about where they are using it. It's a must to turn-on only with raytracing it seems.

AMD's APUs are clearly kneecapped by the memory bandwidth, and from the leaks we've seen there is no solution in sight for the mainstream APUs (next one is Strix Point). Strix Halo on the other hand will have a doubled memory bus width, but I wouldn't count on a mini PC with it to beat CPU+dGPU combos in price/performance unless we see some really generous pricing.

They could work magic with it if they wanted to, particularly a cheaper 8-core model. Strix Halo is presumably using up to 2x 8-core Zen 5 chiplets, so 6/8/12/16 is planned according to leaks.

https://videocardz.com/newz/amd-ryzen-8000-zen5-apus-leak-strix-halo-with-rdna3-5-igpu-to-compete-with-up-to-rtx-4070-max-q

That leak is almost a year old, but from what I can gather we'll probably see this:

16-core Zen 5, 40 CUs RDNA3.5, 256-bit memory (270 GB/s)
12-core Zen 5, 32 CUs RDNA3.5, 256-bit memory (270 GB/s)
8-core Zen 5, 24/32(?) CUs RDNA3.5, 256-bit memory (270 GB/s)
6-core Zen 5, 20 CUs RDNA3.5, 128-bit memory (135 GB/s)

All coming with a 45-50 TOPS XDNA2 AI engine. Sometime in 2025.

Those memory bandwidth figures may be assuming around DDR5-8200 to DDR5-8400, not sure.

A 7600 XT has 288 GB/s and 6700 XT has 384 GB/s, so there's still starvation but doubling that memory bus should greatly alleviate the problem.

[revelator]

sorry did not mean on the cpu side where things have indeed escalated towards better performance, i ment in regards to things such as frame generation for gfx cards which as it turns out does not really help all that much unless you can allready get high framerates, upscaling also though to a lesser degree but still it is kinda trying to get more umph without actually well getting it. sure AI can restore much but it isnt a miracle toy to be used for letting underperforming hardware achive things it normally cant, to many downsides... i would not be surprised if people went for it though they might come to regret it

[jaxa]

I think they have extremely different attitudes towards upscaling and frame generation, particularly Nvidia. Clearly their marketing department has gone wild for it, often with claims of "3X... NO, 4X THE PERFORMANCE!".

They want upscaling and frame generation to be an automatically applied solution that gamers use without thinking much about it, making unplayable scenarios work, particularly heavy raytracing. It will get used often for 4K, and 8K if that takes off, since there's already a lot of detail to work with at the higher resolutions. Upscaling 540p to 1080p could be a disaster, upscaling 1440p to 4K or 4K to 8K probably isn't. Frame generation can be used to fill out 1080p/480Hz or 4K/240Hz without feeling too laggy. Nvidia will not stand still on either approach, there will be a "DLSS 4", "DLSS5"... and even a "DLSS10"?

At the lower end, some people are very satisfied that DLSS2 can help low-end cards like a 3050 Ti (mobile) hit higher frame rates, or hold at a 60 FPS cap with reduced heat and power consumption. Quality loss not apparent, except for maybe Ultra Performance mode. Your pick anyway.

EDIT:

NVIDIA Intros The GeForce RTX 3050 6 GB GPU At $169 – Designed For 1080p Budget Gaming, No Power Plug Required

Quote

The RTX 3050 6GB is a great upgrade for GeForce GTX 1650 and GTX 1050 users wanting to experience RTX. The RTX 3050 6GB is up to 4x faster than its predecessors at the same power and it also supports DLSS Super Resolution for superb performance and image quality.

There's that magic 4X again!

Edited February 2 by jaxa

[revelator]

unfortunatly frame generation does feel laggy unless you have the nessesary horsepower according to reviews so im not to sold on that .  the idea is fine but it does come with some downsides but perhaps when drivers get more mature it will even out some, only time will tell.

[jaxa]

@revelator

 

[revelator]

interresting got some of those games and i pretty much agree with his assesment.

the talos principle ??? wow thats an old title to get the DLSS treatment.

now for alan wake 2... the can it really run crysis crysis game. mesh shaders sound interresting but i think they kinda shot themself in the shoes making a game that litterally requires a 4080 just to play on high settings and if you still got a 1080 ti like mine runs at 15 fps at the lowest setting hehe.

my 2080 ti can atleast play it at medium with around 35 fps but DLSS is a must even at 1080p.

[jaxa]

Looking at the performance charts, I feel the move is to turn it on, cap framerate at 60. Most of the minimums were around 60 with DLSS Performance on, so it would be flawless.

But here I am at glorious 720p quality. Also my main system is out of commission for some reason, I'm going to try updating the BIOS.

[revelator]

hmm with FSR in performance mode  i could boost it to 25/30 fps at the lowest setting on the 1080 ti so from slideshow to barely playable at 1080p . strangely FSR does not seem to get this old card a whole lot of extras in this game, compared to the callisto protocol where i can actually keep it above 60 fps with raytracing on at 1080p, it is a whole lot more detailed compared to the latter though so that might explain it to some degree. also the 1080 ti does not fully support dx 12 ultimate so the mesh shading might not work all to well with it.

[revelator]

does not work at all it turns out the game even warns about it just did not notice that part when i first fired it up.

lowering the resolution does not get you much either i went as low as 640x480 and got 1 fps whaaaat!.

there is a vulkan mod that alledgedly nets you a whopping 2 fps improvement at 1080p but it does not work for all...

well atleast i can play it but its quite laggy on the 1080 ti, the 2080 ti runs it quite ok at medium but any higher and things go very very badly.

[revelator]

hmm well turns out there is an FSR 3 mod out now that actually helps quite a bit with the 1080 ti, still not getting stellar performance but i can play it at medium settings now with around 45 fps on the old card.

the 2080 ti can manage it at medium with raytracing at low and still get 60 fps or high with raytracing off to keep it at around 60 fps, drops to around 45 fps with raytracing at low but still playabe due to it being very stable performance wise. fps only fluctuates by 2 to 3 frames.

have to say the game looks really good the story is a bit weird though. there are actually elements from control in it you even get to work with the FBC.

[revelator]

well the fsr3 mod from nukem gave the old 2080 ti a massive boost. i can play it at high with medium raytracing now at 117 fps . there is some minimal ghosting in some scenes with framegen on but overall its very good looking. at 4k i suspect things would look very different fps wise.

[jaxa]

I think the writing is on the wall. Advanced upscaling will be adopted as widely as possible as the free performance band-aid for the gaming industry. The majority of players will probably run it automatically without even noticing.

Recently we've seen rumors of Microsoft working on a Windows upscaler (which may be similar to AMD's RSR in that the game developers don't need to touch it) and Sony may include an NPU in a PlayStation 5 Pro for their own bespoke console-level upscaling solution (not an FSR 3/4, although those can be supported).

The irony would be if Nvidia ended up killing the demand for gaming GPUs faster by marketing DLSS so hard, that there's less "need" for new and top-end GPUs. But they won't care because they prefer to chase more lucrative markets like AI, datacenter, automotive.

I say "faster" because there is some point in the future when additional hardware can't push the boundaries of graphics, or faster hardware can't be created. We'll see an evolution of Unreal Engine 5's photorealism approach, adoption of 8K resolution, possibly 16K for VR, and a push to the 240-1000 FPS range. Generated frames could be used for a free doubling if not quadrupling of FPS to hit those high numbers, and upscaling tends to work better when your input/target resolution are already very high. For VR specifically, foveated rendering can slash hardware requirements, possibly by 80% or more if the implementation is good enough.

On the hardware side, there's still free lunch to be had with a few additional node shrinks. Stacked L2/L3 cache could be extremely beneficial, think the 3D V-Cache version of Infinity Cache (Nvidia has gone with big L2 with Lovelace). We don't see adoption of High Bandwidth Memory in consumer GPUs because it is in such high demand for AI/enterprise products, but there's no technical reason it can't be used. We will see the blossoming of mega APUs this decade.

[revelator]

I agree, it had to happen with the move towards ultra reality .

I suspect the comming years will be quite interresting

[revelator]

damn i just broke the mainboard on my gaming pc switching PSU i did not notice that the cable for the 8 pin cpu power was the wrong type for that PSU and now it wont boot at all sigh...

strangely it does light up the leds but the fans and qcode display stay of no matter what, there are 2 leds at the buttom that light up but according to the manual they just show the board has power so lol.

luckily my 2080 ti still works and the ram also seems to have survived, havent tested the cpu or disks yet.