Next Gen GPU's will be even more expensive

[las]

This isn't a measure of power efficiency. What is is you now have a 200 W GPU achieving roughly the same a 320 W of previous generation did.

Not really, 4070 performs like 3080 overall in raster but 4070 wins in most new games and have 2GB more VRAM and options for Frame Gen on top + DLSS 3 support, which will then massively beat 3080/6800XT from last gen, at just 200 watts, meaning cool and quiet operation. Down from 350-400 watts. Radeon 6800/6900 series also had massive power spikes which was fixed with 7000 series. AMD still struggles with multi monitor and video playback power usage. Using like 2-3 times as much as Nvidia.

Also, 4070 SUPER exist, at 599 MSRP, where 3080 was 699 MSRP but closer to 1000+ in reality because of GPU mining craze.

The best way to compare GPUs are, and will always be, on performance per watt. Ada have like way way better performance per watt than Ampere and Radeon 6000 series as well. Beats Radeon 7000 too. With full support for DLSS 3 and FG.

AMD probably won't be able to deliver 4090 performance till 2026+ with RDNA5 thats how far behind AMD is. Had my 4090 since Sep 2022... Bought for 1500 dollars, sold my 3090 for 1000 dollars and used a temp 6800XT for a few months, which is why I know exactly how far behind AMD is on drivers/features. Wonky experience with lack of drivers for new games on release is the norm for AMD.

AMD is cheaper for a reason, and still don't sell.

When you factor in the much lower resell value and higher power usage (idle, video, multi monitor + gaming), it is simply not worth it for 90% of people to even consider an AMD GPU and AMD is at like 10% dGPU marketshare now. Lets see if RDNA4 will make them regrab some. They will need very aggressive pricing to do that and FSR needs to be improved.

They will even go back to monolithic with RDNA4, showing MCM failed with 7000 series.

 

[igormp]

And this is also why 5090 is rumoured to get like up to 25.000 cores which is 50% more than 4090

5090 is probably going to be a beast, maybe MCM, maybe not, however watt usage will be high and price will be too.

5090 with 25k cores? You wish. That's the high-end GB202, which won't appear in the consumer market, similar to how the top AD102 never did.

More likely scenario is something around 20k cores, which is like a 25% jump from the 4090, coupled with a probably bit larger memory bus, with way faster GDDR7.

I also doubt it'll be a MCM design.

Nvidia changed almost nothing in the architecture with Ada, and if this continues with Blackwell, then all advancements (or the lack of them) will likely go hand in hand with node shrinks.

They did change some pretty nice stuff... for machine learning stuff.
Same goes for blackwell. Nvidia's improvments have mostly been focused on where the money is at.

Nvidias focus is on AI and Enterprise and they saw with 3000 series that they could easily beat AMD in the gaming market using a cheap process

That's a great example, since GA100 was on TSMC 7nm, while the rest of the chips were on Samsung's crappy node.

it is simply not worth it for 90% of people to even consider an AMD GPU and AMD is at like 10% dGPU marketshare now. Lets see if RDNA4 will make them regrab some.

AMD is not even trying that hard, as others already said. Their main focus is on CPUs, and the remaining GPU work is meant for either mobile (like laptops) or their enterprise stuff.

 

[Zach_01]

AMD is not even trying that hard, as others already said. Their main focus is on CPUs, and the remaining GPU work is meant for either mobile (like laptops) or their enterprise stuff.

Has anyone heard rumors of AMD switching (yet again) to UDNA (RNDA+CDNA) after a couple of series? (Probably after RDNA5)
If its true... jeeez make up your minds!!

 

[oxrufiioxo]

Has anyone heard rumors of AMD switching (yet again) to UDNA (RNDA+CDNA) after a couple of series? (Probably after RDNA5)
If its true... jeeez make up your minds!!

I think people are tired of them flip flopping every couple generations unlike the Ryzen/Epyc side their gpu divisions doesn't seem to know what they're doing hopefully this unification helps but it didn't seem to lead to better products during the GCN/Vega generations.

I'm honestly taking it as we don't care about the Desktop gpu market and rather just focus on the AI market and just castrate whatever doesn't sell for the Desktop market.

Hopefully I'm wrong but I'm not holding my breath they'll get their shit together.

 

[wheresmycar]

Has anyone heard rumors of AMD switching (yet again) to UDNA (RNDA+CDNA) after a couple of series? (Probably after RDNA5)
If its true... jeeez make up your minds!!

If Nvidia is successfully managing a single architecture which serves both gaming and DC/AI, and does it by reaping rewards with cost efficiency, streamlined software, less development headaches, etc.... it does seem to be the more approachable long term strategy. Dropping specialized architectures will have trade-offs but perhaps AMD needs an ALL-IN-ONE solution for a more efficient ecosystem to more capably compete with Nvidia at the highest level. As the saying goes: "if you can't beat em, join em".

 

[AvrageGamr]

Placing lighting, RT or otherwise isn't as simple as just placing the light. You still have decide the light type (spot, directional, point, ect), light color, and the many other light properties. Certain lights will disperse photons in a specific direction with a very sharp cutoff while others will disperse them around the lighting fixture. More realistic light propagation simulation ala RT doesn't change these requirements. You are still going to want to adjust these parameters regardless of whether you are using RT or rasterized lighting. In addition, some devs may want to bake lighting when fully dynamic lighting isn't required and other times they will need to use different types of reflections (aside from SSR, UE5 supports cheaper reflection captures) because having a ton of RT reflections isn't remotely feasible. Devs are going to use a gambit of reflection types according to their requirements including RT, SSR, and reflection captures (which includes planar, box, ect).

You may have misunderstood what they said, implementing RT lighting over rasterized lighting might be easy because you've already done most of the work setting the correct lighting properties but implementing RT lighting does not absolove you of doing that work in the first place. In addition you may still have to go through and adjust lighting parameters for ideal visibility under RT. I've noticed in many RT enabled games the devs don't do this and often text or objects that are supposed to be visible are harder to see. If the dev you talked to really did think that you can simply place a generic light source throughout their entire game without editing it's properties then I very much doubt they were much of a dev to begin with.



Star Wars outlaws having RT on by default doesn't mean the game doesn't use rasterization (it does, hence why you get ok performance on cards like the 1080 Ti). It also uses hybrid RT, so a portion of the work is done on tradition shader cores.

I'm sure we'll move to RT only at some point but people have been saying "once mid tier hardware is capable of it" since the 2000 series (heck there were even some bold individuals claiming mid tier 2000 series cards were it already) and here we are, mid-tier GPUs now costing $700 USD and requiring upscaling just to run modern games with RT enabled at a decent frame-rate. The progress has been less than impressive.

It's not even the first time I've heard someone say 'nice mid-tier hardware is capable of RT well get RT only games'. That was 6 years ago now and we are still waiting. I'd MUCH rather have ray-traced sound, much less resource intensive and extremely beneficial. Even better, dynamically generate sounds based on how sound propogation work. Every pot should make a different sound based on the weapon hitting it, where it hits it, ect. Literally anything other than graphics for once. Decent AI for example sure would be nice. Modern AAA games are as wide as an ocean and as shallow as a puddle.

Outlaws uses software rt at all times. There is no crappy ambient occlusion in this game.

 

[igormp]

I'm honestly taking it as we don't care about the Desktop gpu market and rather just focus on the AI market and just castrate whatever doesn't sell for the Desktop market.

The current ISA split is awful, while the actual feature difference is minimal.
Focusing on a single ISA is a way better approach, just see Nvidia.

That doesn't mean that the halo enterprise products need to have the exact same hardware as the other chips, just see how the x100 chips are widly different from the rest of the stack, but the software stack is still a single piece. So the CUDA code that I write for my 3090 will work just as fine on a A100/H100, something that's not possible with AMD's current offerings.
Heck, apart from their MI200/250/300 chips, the rest of their stack has awful ROCm support.

But yeah, this move is totally meant to make enterprise stuff easier. Their CDNA products also are quite few, so it's an entire software stack just for 1 or 2 SKUs, while all the other RDNA SKUs have shit support and people who don't have a MI300 can't properly validate their code.

If Nvidia is successfully managing a single architecture which serves both gaming and DC/AI, and does it by reaping rewards with cost efficiency, streamlined software, less development headaches, etc.... it does seem to be the more approachable long term strategy. Dropping specialized architectures will have trade-offs but perhaps AMD needs an ALL-IN-ONE solution for a more efficient ecosystem to more capably compete with Nvidia at the highest level. As the saying goes: "if you can't beat em, join em".

Adding to this, and as I said above, while Nvidia has a single high-level arch the actual chips can be wildly different, but the software stack remains the same, something that is really lacking on AMD's side.

 

[Zach_01]

If Nvidia is successfully managing a single architecture which serves both gaming and DC/AI, and does it by reaping rewards with cost efficiency, streamlined software, less development headaches, etc.... it does seem to be the more approachable long term strategy. Dropping specialized architectures will have trade-offs but perhaps AMD needs an ALL-IN-ONE solution for a more efficient ecosystem to more capably compete with Nvidia at the highest level. As the saying goes: "if you can't beat em, join em".

Thing is that AMD had a unified arch and split it in 2 after Vega (and RadeonVII?)

I guess it was a bet that didn’t pay off. Also guessing that nVidia forsaw the future better like 5+years ago.

Let’s hope they do it right this time

 

[igormp]

Thing is that AMD had a unified arch and split it in 2 after Vega (and RadeonVII?)

Yes, CDNA is pretty similar to the GCN found in Vega. But each CDNA gen (3 gens so far) only had 1~4 SKUs each.

I guess it was a bet that didn’t pay off. Also guessing that nVidia forsaw the future better like 5+years ago.

Nvidia has been doing this for over 15 years now, and going a different high-end chip under the same software stack since Pascal (P100 in 2016).
They really played the long game to be where they're at now.

 

[wheresmycar]

Adding to this, and as I said above, while Nvidia has a single high-level arch the actual chips can be wildly different, but the software stack remains the same, something that is really lacking on AMD's side.

Won't be an easy transition and as always the biggest fear is at the software/code sync level. AMD will have to ensure the existing software/dev tools remain compatible whilst furthering the cause for the the new arch's optimizations/toolchain updates/APIs/application support, etc. Can AMD pull it off.. why not! Will it be a well-polished early release, i wouldn't hold my breath. I hope AMD gets it right the first time and early performance issues, driv/soft incompatibilities, bugs/glitches, power management issues, etc are quickly resolved in a respectable time frame. If not they will have shot themselves in the foot with no big or fast enough bandaid to mask/shelter the madness.

Thing is that AMD had a unified arch and split it in 2 after Vega (and RadeonVII?)

I guess it was a bet that didn’t pay off. Also guessing that nVidia forsaw the future better like 5+years ago.

Let’s hope they do it right this time

Yep a punt which defied the obvious.... a unified arch was always going to help efficiently solve problems caused by having different/incompatible hardware designs. With ML/DL+AI driving big changes, and multi-die GPUs potentially seeing broader mainstream adoption, you've got a whole host of additional challenges to contend with. So its a good time to switch up and simplify tackling everything from a single front (UDNA). NVIDIAs been at it for some time, hence lots of ground to cover for AMD. We can only hope they quickly adapt without the long pitstops or support being MIA.

 

[igormp]

Won't be an easy transition and as always the biggest fear is at the software/code sync level. AMD will have to ensure the existing software/dev tools remain compatible whilst furthering the cause for the the new arch's optimizations/toolchain updates/APIs/application support, etc. Can AMD pull it off.. why not! Will it be a well-polished early release, i wouldn't hold my breath. I hope AMD gets it right the first time and early performance issues, driv/soft incompatibilities, bugs/glitches, power management issues, etc are quickly resolved in a respectable time frame. If not they will have shot themselves in the foot with no big or fast enough band-id to mask/shelter the madness.

I believe the new ISA will be closer to either RDNA or CDNA, while the other will just be killed. This would make the transition smoother.

But that's just me doing wild guesses.

 

[wheresmycar]

I believe the new ISA will be closer to either RDNA or CDNA, while the other will just be killed. This would make the transition smoother.

But that's just me doing wild guesses.

That makes sense.

Or, if feasible, perhaps they will build off one of their existing architectures and integrate key features from the other one. It simplifies the conversion, channels through much of the required software/driver compatibility, etc. Maybe not the best fit with potential limitations going forward or the complex nature of unifying features from both Archs.

Actually toss that, I hope it's a fresh take with a ground-up custom design... a cleanly well-polished instruction set.

 

[Jarmer]

I’m glad that AMD is going that route.

Lets be honest, you and everyone else complaining about the “lack of competition “ really want AMD to force Ngreedia to cut prices just so all of you can buy a cheaper Ngreedia gpu.

You never had the most minimum intention in giving your money to AMD.

Uhhhhhhh what the hell are you blabbering about?

I currently have a 6800xt and love it. I just want them to keep making high end gpus so I can buy them.

 

[Zach_01]

Uhhhhhhh what the hell are you blabbering about?

I currently have a 6800xt and love it. I just want them to keep making high end gpus so I can buy them.

We all want that and I will include nVidia buyers too.
But in order to do this AMD needs a much better architecture build from ground up.
So at this point with what they have, they can't compete with nVidia on the high end.
Its clear and simple.

That's why there are rumors that they are building a unified (RNDA+CDNA) architecture for later (introduced maybe after RDNA5)
So for at least a couple of gens dont expect big flagship GPUs from AMD.

For marketing and image reasons they are saying now that they need to gain market share first from entry to middle level, where the supposed bigger share is.
But in reality they dont have anything viable to show on the high end. They could... but it would be expensive and would end up be much inferior to 5080/5090.
And there is no point to allocate fab wafers on such a product. Wafer allocation is on other parts and they are keeping GPU dies small, for now.

 

[AusWolf]

For marketing and image reasons they are saying now that they need to gain market share first from entry to middle level, where the supposed bigger share is.
But in reality they dont have anything viable to show on the high end. They could... but it would be expensive and would end up be much inferior to 5080/5090.
And there is no point to allocate fab wafers on such a product. Wafer allocation is on other parts and they are keeping GPU dies small, for now.

What they're saying, and what you marked as "reality" are both realities, imo. Two sides of the same coin, kind of.
They don't have anything to compete with in the high-end, therefore they're focusing their efforts in areas where they can score some easy wins due to bigger sales numbers, smaller dies, cheaper R&D, etc.

 

[Zach_01]

What they're saying, and what you marked as "reality" are both realities, imo. Two sides of the same coin, kind of.
They don't have anything to compete with in the high-end, therefore they're focusing their efforts in areas where they can score some easy wins due to bigger sales numbers, smaller dies, cheaper R&D, etc.

Yes that's better, actually I was thinking it the last 5min after posting that.
One does not eliminate the other

 

[oxrufiioxo]

Yes that's better, actually I was thinking it the last 5min after posting that.
One does not eliminate the other

Going after the midrange hasn't helped them though they've tried it multiple times in the past.

RDNA2 really helped with their perception that they can compete and make a good product the problem is they never stick with it.

I'm just pessimistic they can compete whatsoever with their Radeon portion of their business even when they have a clearly better product people still buy nvidia instead becuase of perception of their brand and that won't change till they are clearly as good or better for multiple generations in a row.

 

[Zach_01]

Going after the midrange hasn't helped them though they've tried it multiple times in the past.

RDNA2 really helped with their perception that they can compete and make a good product the problem is they never stick with it.

I'm just pessimistic they can compete whatsoever with their Radeon portion of their business even when they have a clearly better product people still buy nvidia instead becuase of perception of their brand and that won't change till they are clearly as good or better for multiple generations in a row.

The "Yes that's better" was aiming to the better explanation @AusWolf stated about what AMD is doing.

I can agree with you but they dont have anything else right now really.
Yes they need to be consistent for several gens just like they did with Zen1~Zen3.

If the "new" stuff comes after RDNA5 and its the starting point of Zen1 then imagine a few more gens to be up for serious competition.
Its grim... for now

 

[oxrufiioxo]

The "Yes that's better" was aiming to the better explanation @AusWolf stated about what AMD is doing.

I can agree with you but they dont have anything else right now really.
Yes they need to be consistent for several gens just like they did with Zen1~Zen3.

If the "new" stuff comes after RDNA5 and its the starting point of Zen1 then imagine a few more gens to be up for serious competition.
Its grim... for now

I'm still less optimistic Zen worked becuase they went after marketshare and it paid off big time for them. I'd argue the 8700k and 9900k were better than the 1700X/2700X but it didn't matter AMD priced their products significantly better especially over time.

RTX 4000 series left a titanic sized door for them to offer way better products from a price to performance perspective and they just did what they always do offer a slightly worse product for slightly less money. For some depending on what's important significantly worse.

Everything I'm hearing Zen5 is almost non existent sales wise and even with much higher ram/mobo prices Zen 4 was way more successful at launch it just shows how much they're sputtering in the Desktop space on both sides.

We need the AMD Ryzen team of 2017/18 not the current one honestly. They've honestly forgotten what's got them to where they're at.

I hope I'm wrong because overall AMD is one of the great stories over the last Decade.

 

[chrcoluk]

Yeah, Ada got the better optical flow thingy for FG, but I don't see any other advancement, honestly. Enabling RT results in the same performance drop as it did on Ampere or Turing. The efficiency and clock speeds gains are due to the node shrink, not the architecture.

This is the difference between the Nvidia and AMD mindset. If Nvidia says they've got some cool new stuff, everybody praises them for it without checking out how (or if) it works in real life first. If AMD says the same, everybody mocks them for not delivering 20% better everything than Nvidia for a 50% lower price.

Probably irrelevant for most people, but they did upgrade the encoder/decoder chips, my 3080 couldnt do 1440p or above in youtube without stuttering, monitoring the video chip on gpu-z at fastest possible polling showed on every stutter the chip was saturated, on my 4080 those issues are gone. So I can hardware decode youtube now lol.

Not yet tested encoding but they now officially support HEVC so I expect thats a big improvement. That might be appreciated by a fair amount of people, as NVENC on 3000 series was pretty bad, ended up with massive files compared to software encoding for anything even remotely comparable quality.

Will probably start gaming regularly again (including recording) by end of this month, I slow down my gaming in summer as the heat puts me off too much.

 

[igormp]

Probably irrelevant for most people, but they did upgrade the encoder/decoder chips, my 3080 couldnt do 1440p or above in youtube without stuttering, monitoring the video chip on gpu-z at fastest possible polling showed on every stutter the chip was saturated, on my 4080 those issues are gone. So I can hardware decode youtube now lol.

They did not. The decoder in both the 3000 and 4000 series is the exact same, 5th gen NVDEC. Your issues may have been due to some other weird software thing in your system.

Not yet tested encoding but they now officially support HEVC so I expect thats a big improvement.

For the encoder, it's pretty similar to the Ampere one, with the addition of AV1 encoding support. The 4080 and 4090 also have two encoders now instead of just a single one.
HEVC was already supported in Ampere. The quality is the same in Ada.

That might be appreciated by a fair amount of people, as NVENC on 3000 series was pretty bad, ended up with massive files compared to software encoding for anything even remotely comparable quality.

HW encoding will always give you larger files since it's a tradeoff for speed vs quality vs file size.

 

[chrcoluk]

They did not. The decoder in both the 3000 and 4000 series is the exact same, 5th gen NVDEC. Your issues may have been due to some other weird software thing in your system.

For the encoder, it's pretty similar to the Ampere one, with the addition of AV1 encoding support. The 4080 and 4090 also have two encoders now instead of just a single one.
HEVC was already supported in Ampere. The quality is the same in Ada.

HW encoding will always give you larger files since it's a tradeoff for speed vs quality vs file size.

The only thing that changed was the drivers and the GPU swap.

HEVC I could only use it in OBSS with a unofficial plugin and it was overloading the chip quite badly. Would get frame drops when was overloaded.

Improving Video Quality and Performance with AV1 and NVIDIA Ada Lovelace Architecture | NVIDIA Technical Blog

NVIDIA NVENC AV1 offers substantial compression efficiency with respect to H.264 and HEVC at better performance.

developer.nvidia.com

I am probably thinking of AV1 as the new supported codec for encoding, but regardless I could only use HEVC unofficially in OBSS.

This page does support what you saying its 5th gen on both generations, but end of the day I can only report what I observe, I couldnt run hardware decode at high resolutions and 60fps on my 3080 and now can, and when I monitored it in GPUz the video chip was hitting high utilisation briefly when it stuttered. Willing to accept may have been a driver problem.

 

[igormp]

HEVC isnt supported in Ampere either, they added it as a new thing to the 4000 series. I could only use it in OBSS with a unofficial plugin and it was overloading the chip quite badly. Would get frame drops when was overloaded.

It is. HEVC encoding has been a thing since Pascal (1000 series).
Take a look here:

Video Encode and Decode GPU Support Matrix

Find the related video encoding and decoding support for all NVIDIA GPU products.

developer.nvidia.com

Also on the ADA whitepaper, on pages 24-25:

https://images.nvidia.com/aem-dam/Solutions/Data-Center/l4/nvidia-ada-gpu-architecture-whitepaper-v2.1.pdf

Ada GPUs take streaming and video content to the next level, incorporating support for AV1 video encoding in the Ada eighth generation dedicated hardware encoder (known as NVENC). Prior generation Ampere GPUs supported AV1 decoding, but not encoding. Ada’s AV1 encoder is 40% more efficient than the H.264 encoder used in GeForce RTX 30 Series GPUs. AV1 will enable users who are streaming at 1080p today to increase their stream resolution to 1440p while running at the same bitrate and quality, or for users with 1080p displays, streams will look similar to 1440p, providing better quality.
NVIDIA collaborated with OBS Studio to add AV1 — on top of the recently released HEVC and HDR support — within an upcoming software release, expected later this year. OBS is also optimizing encoding pipelines to reduce overhead by 35% for all NVIDIA GPUs. The new release will additionally feature updated NVIDIA Broadcast effects, including noise and room echo removal, as well as improvements to virtual background.
We’ve also worked with Discord to enable end-to-end livestreams with AV1. In an update releasing later this year, Discord will enable its users to use AV1 to dramatically improve screen sharing, be it for game play, schoolwork, or hangouts with friends.
To further aid encoding performance, GeForce RTX 4090 and RTX 4080 are equipped with dual NVENC encoders. This enables video encoding at 8K/60 for professional video editing or four 4K/60. (Game streaming services can also take advantage of this to enable more simultaneous sessions, for instance.) Blackmagic Design’s DaVinci Resolve, the popular Voukoder plugin for Adobe Premiere Pro, and Jianying — the top video editing app in China — are all enabling AV1 support, as well as a dual encoder through encode presets. Dual encoder and AV1 availability for these apps will be available in October. NVIDIA is also working with the popular video-effects app Notch to enable AV1, as well as Topaz to enable support for AV1 and the dual encoders.
In addition to NVENC, Ada GPUs also include the fifth-generation hardware decoder that was first launched with Ampere (known as NVDEC). NVDEC supports hardware-accelerated video decoding of MPEG-2, VC-1, H.264 (AVCHD), H.265 (HEVC), VP8, VP9, and the AV1 video formats. 8K/60 decoding is also fully supported

The issue you may have faced is that HEVC support was not added on OBS at the time, and the plugin you used was not that good. Or you were just hitting the encoder's limit, the 4090 improved on that by having two encoders than can be used simultaneously now.

I am probably thinking of AV1 as the new codec, but regardless I could only use HEVC unofficially in OBSS.

It now has official support, and has had it for quite some years for recording. For streaming, AFAIK only youtube accepts it.

but end of the day I can only report what I observe, I couldnt run hardware decode at high resolutions and 60fps on my 3080 and now can, and when I monitored it in GPUz the video chip was hitting high utilisation briefly when it stuttered.

Sure, and I do believe you, but it may have been a problem with your system or driver for some unkown reason. The whole 3000 series has the same decode capability as the 4000 series in the end of the day.
Maybe your driver reinstall solved the issue, go figure.

 

[RJARRRPCGP]

Or they have to repeat what they've done a decade ago with Maxwell. Same node, much more bang per watt compared to Kepler.

Maxwell, was what Kepler should have been!

Nvidia can raise prices all it wants, many cards are already priced at what the market will bear.

People were already extending how long they keep their cards to help accommodate for increased pricing but $1,000 for a 5070 Ti that will have a tiny memory bus and little VRAM doesn't sound anywhere remotely appealing.

Sounds more like the '21 "video-card-mageddon"!

Outlaws uses software rt at all times.

But, does it easily make the CPU requirement a Ryzen 9 5900X or equivalent?

 

[Krit]

Let’s hope that rx 8800 xt will cost the same amount as rx 7800 xt. Rx 7800 xt is a good decent card but a bit boring becouse of it’s 6800xtish performance level. If rx 8800 xt manages to be at least 25-40% faster at the same price it’s already a win. Personally not interested in nvidia becouse it will be overpriced.