There isn't an RDNA5 on the roadmap, though. It's been confirmed 4 is the last (and was really meant to be 3.5, but grew into what is assumed to be the PS5/XSX mid-gen refresh architecture).
Next is UDNA1, a converged architecture with it's older sibling, CDNA (formerly GCN).
Like, the article actually states this, but runs an RDNA 5 headline anyways.
And their half step/semi-custom work can find their way back to APUs. RDNA 3.5 (the version marketed as such) is in the Zen 5 APUs with Mobile oriented improvements. It wouldn’t surprise me if a future APU gets RDNA 5. GCN had this sort of APU/Console relationship as well.
UDNA isn't a name but instead a big shift in strategy.
CDNA was for HPC / Supercomputers and Data center. GCN always was a better architecture than RDNA for that.
RDNA itself was trying to be more NVidia like. Fewer FLOPs but better latency.
Someone is getting the axe. Only one of these architectures will win out in the long run, and the teams will also converge allowing AMD to consolidate engineers to improving the same architecture.
We won't know what the consolidated team will release yet. But it's a big organizational shift that surely will affect AMDs architectural decisions.
My understanding was that CDNA and RDNA shared much if not most of their underlying architecture, and that the fundamental differences had more to do with CDNA supporting a greater variety of numeric representations to aid in scientific computing. Whereas RDNA really only needed fp32 for games.
CDNA is 64 wide per work item. And CDNA1 I believe was even 16 lanes executed over 4 clock ticks repeatedly (ie: minimum latency of all operations, even add or xor, was 4 clock ticks). It looks like CDNA3 might not do that anymore but that's still a lot of differences...
RDNA actually executes 32-at-a-time and per clock tick. It's a grossly different architecture.
That doesn't even get to Infinity Cache, 64-bit support, AI instructions, Raytracing, or any of the other differences....
A reason was backwards compatibility, studios were already putting lots of money into PS4 and XBox One, thus PS5 and XBox X|S (two additional SKUs), were already too much.
Don't forget one reason that studios tend to favour consoles has been regular hardware, and that is no longer the case.
When middleware starts to be the option, it is relatively hard to have game features that are hardware specific.
Less effort going into optimization also plays a factor. On average games are a lot less optimized than they used to be. The expectation seems to be that hardware advances will fix deficiencies in performance.
This doesn’t affect me too much since my backlog is long and by the time I play games, they’re old enough that current hardware trivializes them, but it’s disappointing nonetheless. It almost makes me wish for a good decade or so of performance stagnation to curb this behavior. Graphical fidelity is well past the point of diminishing returns at this point anyway.
PS1: 0.03 GFLOPS (approx given it didn't really do FLOPS per se)
PS3: 230 GFLOPS
Nearly 1000x faster.
Now compare PS4 with PS5 pro (also just over 10 years apart):
PS4: ~2TFLOPS
PS5 Pro: ~33.5TFLOPS
Bit over 10x faster. So the speed of improvement has fallen dramatically.
Arguably you could say the real drop in optimization happened in that PS1 -> PS3 era - everything went from hand optimized assembly code to running (generally) higher level languages and using abstrated graphics frameworks like DirectX and OpenGL. Just noone noticed because we had 1000x the compute to make up for it :)
Consoles/games got hit hard by first crypto and now AI needing GPUs. I suspect if it wasn't for that we'd have vastly cheaper and vastly faster gaming GPUs, but when you were making boatloads of cash off crypto miners and then AI I suspect the rate of progress fell dramatically for gaming at least (most of the the innovation I suspect went more into high VRAM/memory controllers and datacentre scale interconnects).
It is not just GPU performance, it is that visually things are already very refined. A ten times leap in performance doesn't really show as ten times the visual spectical like it used to.
Like all this path tracing/ray tracing stuff, yes it is very cool and can add to a scene but most people can barely tell it is there unless you show it side by side. And that takes a lot of compute to do.
Yeah there’s been a drop off for sure. Clearly it hasn’t been steep enough for game studios to not lean on anyway, though.
One potential forcing factor may be the rise of iGPUs, which have become powerful enough to play many titles well while remaining dramatically more affordable than their discrete counterparts (and sometimes not carrying crippling VRAM limits to boot), as well as the growing sector of PC handhelds like the Steam Deck. It’s not difficult to imagine that iGPUs will come to dominate the PC gaming sphere, and if that happens it’ll be financial suicide to not make sure your game plays reasonably well on such hardware.
I get the perhaps mistaken impression the biggest problem games developers have is making & managing absolutely enormous amounts of art assets at high resolution (textures, models, etc). Each time you increase resolution from 576p, to 720p to 1080p and now 4k+ you need a huge step up in visual fidelity of all your assets, otherwise it looks poor.
And given most of these assets are human made (well, until very recently) this requires more and more artists. So I wonder if games studios are more just art studios with a bit of programming bolted on, vs before with lower res graphics where you maybe had one artist for 10 programmers, now it is more flipped the other way. I feel that at some point over the past ~decade we hit a "organisational" wall with this and very very few studios can successfully manage teams of hundreds (thousands?) of artists effectively?
This hits the nail pretty close to the head. I work on an in-house AAA engine used by a number of different games. It's very expensive to produce art assets at the quality expected now.
Many AAA engine's number one focus isn't "performance at all costs", it's "how do we most efficiently let artists build their vision". And efficiency isn't runtime performance, efficiency is how much time it takes for an artist to create something. Performance is only a goal insofar as to free artists from being limited by it.
> So I wonder if games studios are more just art studios with a bit of programming bolted on.
Not quite, but the ratio is very in favor of artists compared to 'the old days'. Programming is still a huge part of what we do. It's still a deeply technical field, but often "programming workflows" are lower priority than "artist workflows" in AAA engines because art time is more expensive than programmer time from the huge number of artists working on any one project compared to programmers.
Just go look at the credits for any recent AAA game. Look at how many artists positions there are compared to programmer positions and it becomes pretty clear.
That depends a lot on art direction and stylization. Highly stylized games scale up to high resolutions shockingly well even with less detailed, lower resolution models and textures. Breath of the Wild is one good example that looks great by modern standards at high resolutions, and there’s many others that manage to look a lot less dated than they are with similarly cartoony styles.
If “realistic” graphics are the objective though, then yes, better displays pose serious problems. Personally I think it’s probably better to avoid art styles that age like milk, though, or to go for a pseudo-realistic direction that is reasonably true to life while mixing in just enough stylization to scale well and not look dated at record speeds. Japanese studios seem pretty good at this.
Yeah, its flipped. Overall, it has meant studios are more and more dependent on third party software (and thus license fees), it led to game engine consolidation, and serious attrition when attempting to make something those game engines werent built for (non-pbr pipelines come to mind).
It's no wonder nothing comes out in a playable state.
You divided 230 by .03 wrong, which would be 10000-ish, but you underestimated the PS1 by a lot anyway. The CPU does 30 MIPS, but also the geometry engine does another 60 MIPS and the GPU fills 30 or 60 million pixels per second with multiple calculations each.
Not to mention that few developers were doing hand optimized assembly by the time of PSX. They were certainly hand optimizing models and the 3D pipeline (with some assembler tuning), but C and SDKs were well in use by that point.
Even Naughty Dog went with their own LISP engine for optimization versus ASM.
Naughty Dog's GOAL was PS2 specific and essentially chock full of what would be called intrinsics these days that let you interleave individual assembly instructions particularly for the crazy coprocessor setup of Emotion Engine.
My understanding is that the mental model of programming in PS2 era was originally still very assembly like outside of few places (like Naughty Dog) and that GTA3 on PS2 made possibly its biggest impact by showing it's not necessary.
If by "mental model" you mean "low-level" programming, sure. But you might as well conflate "religion" with "Southern Baptist protestantism" then. You're working with the same building blocks, but the programming style is drastically different.
The vast majority of PSX games were done completely in C, period. Some had small bits of asm here and there, but so do the occasional modern C/C++ apps.
To your last point, before there was GOAL there was GOOL (from the horse's mouth itself):
Because optimized games aren’t completely extinct and there’s titles with similar levels of size, fidelity, and feature utilization with dramatically differing performance profiles.
Given the N64-PS1 era is filled with first party games that run at like 20 fps, I'm having a hard time saying things are worse now.
I am a bit uncomfortable with the performance/quality stuff that people have set up but I personally feel that the quality floor for perf is way higher than it used to be. Though there seem to be less people parking themselves at "60fps locked", which felt like a thing for a while
The current generation has a massive leap in storage speed but games need to be architected to stream that much data into RAM.
Cyberpunk is a good example of a game that straddled the in between, many of it's performance problems on the PS4 were due to constrained serialization speed.
Nanite and games like FF16 and Death Stranding 2 do a good job of drawing complex geometry and textures that wouldn't be possible on the previous generation
Nanite has a performance overhead for simple scenes but will render large, complex scenes with high-quality models much more efficiently, providing a faster and more stable framerate.
It’s also completely optional in Unreal 5. You use it if it’s better. Many published UE5 games don’t use it.
And loading times. I think people already forgot how long you had to wait on loading screens or how many faked loading (moving through a brush while the next area loads) there was on PS4
PS4 wasnt too terrible but jumping back to PS3... wow I completely forgot how memory starved that machine was. Working on it, we knew at the time but in retro spect it was just horrible.
Small RAM space with the hard CPU/GPU split (so no reallocation) feeding off a slow HDD which is being fed by an even slower Bluray disc, you are sitting around for a while.
If only we could just ship a 256GB NVMe SSD with every game and memory map the entire drive like you could with cartridges back then. Never have loading times again.
Seems like an overgeneralization. I get it when FPS players want the best performance: players have FOMO of the best reaction time and the games are more built for fast action than contemplative scenery watching.
I wonder if players of single player action/adventure games make the same choice. Those games are played less (can be finished in 10-30 hours instead of endlessly) so the statistics might be skewed to favor performance mode.
This is the result of an industry wide problem where technology just is not moving forward as quickly as it used to move. Dennard scaling is dead. Moore’s law is also dead for SRAM and IO logic. It is barely clinging to life for compute logic, but the costs are skyrocketing as each die shrink happens. The result is that we are getting anemic improvements. This issue is visible in Nvidia’s graphics offerings too. They are not improving from generation to generation like they did in the past, despite Nvidia turning as many knobs as they could to higher values to keep the party going (e.g. power, die area, price, etcetera).
AFAIK, this generation has been widely slammed as a failure due to lack of new blockbuster games. Most things that came out were either for PS4, or remasters of said games.
There have been a few decent sized games, but nothing at grand scale I can think of, until GTA6 next year.
"Bloated" might be the wrong word to describe it, but there's some reason to believe that the dominance of Unreal is holding performance back. I've seen several discussions about Unreal's default rendering pipeline being optimized for dynamic realtime photorealistic-ish lighting with complex moving scenes, since that's much of what Epic needs for Fortnite. But most games are not that and don't make remotely effective use of the compute available to them because Unreal hasn't been designed around those goals.
TAA (temporal anti-aliasing) is an example of the kind of postprocessing effect that gamedevs are relying on to recover performance lost in unoptimized rendering pipelines, at the cost of introducing ghosting and loss of visual fidelity.
This is a very one-sided perspective on things. Any precomputed solution to lighting comes with enormous drawbacks across the board. The game needs to ship the precomputed data when storage is usually already tight. The iteration cycle for artists and level designers suchs when lighting is precomputed - they almost never see accurate graphics for their work while they are iterating because rebaking takes time away from their work. Game design become restricted to those limitations, too. Can't even think of having the player randomly rearranging big things in a level (e.g. building or tearing down a house) because the engine can't do it. Who knows what clever game mechanics are never thought of because of these types of limitations?
Fully dynamic interactive environments are liberating. Pursuing them in is the right thing to do.
In principle, Epic's priorities for Unreal should be aligned to a lot of what we've seen in the PS3/4/5 generation as far as over-the-shoulder 3rd person action adventure games.
I mean, look at Uncharted, Tomb Raider, Spider-Man, God of War, TLOU, HZD, Ghost of Tsushima, Control, Assassins Creed, Jedi Fallen Order / Survivor. Many of those games were not made in Unreal, but they're all stylistically well suited to what Unreal is doing.
TAA isn't a crutch being used to hold up poor performance, it's an optimization to give games anti-aliasing that doesn't suck.
Your other options for AA are
* Supersampling. Rendering the game at a higher resolution than the display and downscaling it. This is incredibly expensive.
* MSAA. This samples ~~vertices~~surfaces more than once per pixel, smoothing over jaggies. This worked really well back before we started covering every surface with pixel shaders. Nowadays it just makes pushing triangles more expensive with very little visual benefit, because the pixel shaders are still done at 1x scale and thus still aliased.
* Post-process AA (FXAA,SMAA, etc). These are a post-process shader applied to the whole screen after the scene has been fully rendered. They often just use a cheap edge detection algorithm and try to blur them. I've never seen one that was actually effective at producing a clean image, as they rarely catch all the edges and do almost nothing to alleviate shimmering.
I've seen a lot of "tech" YouTubers try to claim TAA is a product of lazy developers, but not one of them has been able to demonstrate a viable alternative antialiasing solution that solves the same problem set with the same or better performance. Meanwhile TAA and its various derivatives like DLAA have only gotten better in the last 5 years, alleviating many of the problems TAA became notorious for in the latter '10s.
Erm your description of MSAA isn't quite correct, it has nothing to do with vertices and doesn't increase vertex processing cost..
It's more similar to supersampling, but without the higher pixel shader cost (the pixel shader still only runs once per "display pixel", not once per "sample" like in supersampling).
A pixel shader's output is written to multiple (typically 2, 4 or 8) samples, with a coverage mask deciding which samples are written (this coverage mask is all 1s inside a triangle and a combo of 1s and 0s along triangle edges). After rendering to the MSAA render target is complete, an MSAA resolve operation is performed which merges samples into pixels (and this gives you the smoothed triangle edges).
> solves the same problem set with the same or better performance
The games industry has spent the last decade adopting techniques that misleadingly inflate the simple, easily-quantified metrics of FPS and resolution, by sacrificing quality in ways that are harder to quantify. Until you have good metrics for quantifying the motion artifacts and blurring introduced by post-processing AA, upscaling, and temporal AA or frame generation, it's dishonest to claim that those techniques solve the same problem with better performance. They're giving you a worse image, and pointing to the FPS numbers as evidence that they're adequate is focusing on entirely the wrong side of the problem.
That's not to say those techniques aren't sometimes the best available tradeoff, but it's wrong to straight-up ignore the downsides because they're hard to measure.
Yeah. Only problem is that overly aggressive TAA implementations blur the whole frame during camera rotation. The thing that is even better than standard TAA is a combination of TAA and temporal upscaling, called TSR in Unreal. Still better is the same system but performed by an ML model, e.g. DLSS. Though this requires special inference hardware inside the GPU.
In the past, MSAA worked reasonably well, but it was relatively expensive, doesn't apply to all forms of high frequency aliasing, and it doesn't work anymore with the modern rendering paradigm anyway.
If the Playstation contributions are good enough, maybe RDNA4 -> RDNA5 will be just as good as RDNA3 -> RDNA4. As long as they get the pricing right, anyway.
We've known this for a while, it's an extension of the upscaling and frame generation AMD already worked on in conjunction with Sony for FSR 3 and to a much greater extent FSR 4. Previous articles also have highlighted their shared focus on BVH optimizations
Excited to see how the software support for UDNA1 works out. Very hopeful we'll see some real competition to Nvidia soon in the datacenter. Unfortunately I think the risk is quite high: if AMD burns developers again with poor drivers and poor support, it's hard to see how they'll be able to shake the current stigma.
Take this with a pinch of salt, but the most recent ROCm release installed out of the box on my WSL2 machine and worked first time with llama.cpp. I even compiled llama.cpp from source with 0 issues. That has never happened ever in my 5+ years of having AMD GPUs. Every other time I've tried this it's either failed and required arcane workarounds, or just not worked entirely (including running on 'real' Linux).
I feel like finally they are turning the corner on software and drivers.
Sony's low level APIs for PS4 and PS5 (it doesn't use GNM) is almost a direct mapping to the hardware with very little abstraction compared to Vulkan/DX12. Vulkan is still very high level compared to what's going on inside the driver. There's no point paying the cost of Vulkan's abstractions when half the point of a game console is to have a fixed hardware target, hence GNM.
Because if they write their own they get to own the political/bureaucratic portion of the problem. For better or worse, they don't have to deal with the Kronos Group. They get to optimize their APIs directly against their research with AMD.
Why would Vulkan, as opposed to a custom solution designed to target that hardware and games specifically, be a better solution?
If you’re making a PS game you’re already doing tons of bespoke PS stuff. If you don’t want to deal with it there are plenty of pieces of middleware out there to help.
Honestly these “where’s Vulkan” posts on every bit of 3D capable hardware feel like a stupid meme at this point as opposed to a rational question.
Maybe they should just ship DX12. That’s multi-platform too.
Vulkan is a mess as bad as OpenGL, to the point Khronos has been forced to public admit it didn't went out as planned, is yet again an extension spaghetti, and they are now trying to clean the house with the Vulkan Roadmap introduced at Vulkanised 2025.
That is how good Khronos "design by commitee APIs" end up being.
Vulkan's extension story is a mess. Not surprising from the people that brought you OpenGL.
However please don't undersell what they got right. Because what they got right, they got _very_ right.
Barriers. Vulkan got barriers so absolutely right that every competing API has now adopted a clone of Vulkan's barrier API. The API is cumbersome but brutally unambiguous compared to DirectX's resource states or Metal's hazard tracking. DirectX has a rats nest of special cases in the resource state API because of how inexpressive it is, and just straight up forgot to consider COPY->COPY barriers.
We also have SPIR-V. Again, D3D12 plans to dump DXIL and adopt SPIR-V.
The fundamentals Vulkan got very right IMO. It's a shame it gets shackled to the extension mess.
SPIR-V only exists to reaction of Khronos losing to CUDA and PTX, it is the evolution of SPIR, created when they realised industry didn't really want to write OpenCL in a textual form of C99 dialect, rather bytecode formats targeted by multiple compiler backends, including Fortran ignored by Khronos.
Microsoft adopting SPIR-V as DXIL replacement is most likely a matter of convenience, the format got messy to maintain, is tied to an old fork of LLVM, and HLSL has gotten the industy weight, even favoured over GLSL for Vulkan (which Khronos acknowledge at Vulkanised 2024 not doing any work at all, zero, nada), so why redo DXIL from scratch, when they could capitalize on existing work to target SPIR-V from HLSL.
DXIL has other issues beyond being an orphaned fork of LLVM-IR. LLVM is ill-equipped to represents shaders, and there have been real bugs in the wild because of how not-a-shader-language LLVM-IR is. See this [0] post under the 'The classic subgroup breakage – maximal convergence gone awry' section. Read the whole series really if you haven't, it's a fantastic dive into how awful this stuff is.
DXIL is also a useless intermediate representation because parsing it is so bloody difficult nobody could actually do anything with it. DXIL was, and still is, functionally opaque bytes. You can't introspect it, you can't modify it or do anything of use with it. SPIR-V is dead-simple to parse and has an array of tools built around it because it's so easy to work with.
I don't really see how the OpenCL history is relevant to Vulkan either. Khronos losing the OpenCL game to Nvidia, certainly no thanks to Nvidia's sabotage either, doesn't change that SPIR-V is a much more successful format.
OpenCL history is more than relevant, without SPIR, there would not exist SPIR-V.
People love to blame NVidia for AMD and Intel failures pushing OpenCL into the industry, and Google completly ignoring it for mobile devices, pushing their own RenderScript dialect instead, even if Apple would have behave differently regarding their platforms, the other 80% market has completly ignored it.
This comment is just a bunch of bunk, but regardless, Vulkan is the only thing that's a collaborative effort, whether it's a mess or not. NIH proponents have nothing to offer as an alternative.
And most of the standard we have now starts with something similar to NIH. Vulkan itself is an offshoot of mantel from AMD.
There are valid reason to have a custom api. Especially in domain like game console with hardware with long release cycle, tight performance requirement and legacy (ps4) code to support.
So you're right, though I would never have guessed— in the PS5 hype cycle he gave that deep dive architecture presentation that for all the world looked like he was a Sony spokesperson.
Why not link to the original article here:
https://www.tomsguide.com/gaming/playstation/sonys-mark-cern...
There isn't an RDNA5 on the roadmap, though. It's been confirmed 4 is the last (and was really meant to be 3.5, but grew into what is assumed to be the PS5/XSX mid-gen refresh architecture).
Next is UDNA1, a converged architecture with it's older sibling, CDNA (formerly GCN).
Like, the article actually states this, but runs an RDNA 5 headline anyways.
Maybe read the article before commenting on it, it's not that long.
"Big chunks of RDNA 5, or whatever AMD ends up calling it, are coming out of engineering I am doing on the project"
AMD does do semi-custom work.
Whats to stop sony being like we dont want UDNA 1, we want a iteration of RDNA 4.
For all we know, it IS RDNA 5... it just wont be available to the public.
And their half step/semi-custom work can find their way back to APUs. RDNA 3.5 (the version marketed as such) is in the Zen 5 APUs with Mobile oriented improvements. It wouldn’t surprise me if a future APU gets RDNA 5. GCN had this sort of APU/Console relationship as well.
Also steamdeck before the OLED version and magic leap 2 shared a custom chip, with some vision processing parts fused off for steamdeck.
It's just a name. I'm sure this is all pretty iterative work.
UDNA isn't a name but instead a big shift in strategy.
CDNA was for HPC / Supercomputers and Data center. GCN always was a better architecture than RDNA for that.
RDNA itself was trying to be more NVidia like. Fewer FLOPs but better latency.
Someone is getting the axe. Only one of these architectures will win out in the long run, and the teams will also converge allowing AMD to consolidate engineers to improving the same architecture.
We won't know what the consolidated team will release yet. But it's a big organizational shift that surely will affect AMDs architectural decisions.
My understanding was that CDNA and RDNA shared much if not most of their underlying architecture, and that the fundamental differences had more to do with CDNA supporting a greater variety of numeric representations to aid in scientific computing. Whereas RDNA really only needed fp32 for games.
Who told ya that??
CDNA is 64 wide per work item. And CDNA1 I believe was even 16 lanes executed over 4 clock ticks repeatedly (ie: minimum latency of all operations, even add or xor, was 4 clock ticks). It looks like CDNA3 might not do that anymore but that's still a lot of differences...
RDNA actually executes 32-at-a-time and per clock tick. It's a grossly different architecture.
That doesn't even get to Infinity Cache, 64-bit support, AI instructions, Raytracing, or any of the other differences....
CDNA is based on the older gcn arch so they share the same as pre RDNA ones and RDNA ones.
PS5 was almost twice as fast as the PS4 pro, yet we did not see the generational leap we saw with the previous major releases.
It seems that we are the stage where incremental improvements in graphics will require exponentially more computing capability.
Or the game engines have become super bloated.
Edit: I stand corrected in previous cycles we had orders of magnitude improvement in FLOPS.
A reason was backwards compatibility, studios were already putting lots of money into PS4 and XBox One, thus PS5 and XBox X|S (two additional SKUs), were already too much.
Don't forget one reason that studios tend to favour consoles has been regular hardware, and that is no longer the case.
When middleware starts to be the option, it is relatively hard to have game features that are hardware specific.
Less effort going into optimization also plays a factor. On average games are a lot less optimized than they used to be. The expectation seems to be that hardware advances will fix deficiencies in performance.
This doesn’t affect me too much since my backlog is long and by the time I play games, they’re old enough that current hardware trivializes them, but it’s disappointing nonetheless. It almost makes me wish for a good decade or so of performance stagnation to curb this behavior. Graphical fidelity is well past the point of diminishing returns at this point anyway.
We have had a decade of performance stagnation.
Compare PS1 with PS3 (just over 10 years apart).
PS1: 0.03 GFLOPS (approx given it didn't really do FLOPS per se) PS3: 230 GFLOPS
Nearly 1000x faster.
Now compare PS4 with PS5 pro (also just over 10 years apart):
PS4: ~2TFLOPS PS5 Pro: ~33.5TFLOPS
Bit over 10x faster. So the speed of improvement has fallen dramatically.
Arguably you could say the real drop in optimization happened in that PS1 -> PS3 era - everything went from hand optimized assembly code to running (generally) higher level languages and using abstrated graphics frameworks like DirectX and OpenGL. Just noone noticed because we had 1000x the compute to make up for it :)
Consoles/games got hit hard by first crypto and now AI needing GPUs. I suspect if it wasn't for that we'd have vastly cheaper and vastly faster gaming GPUs, but when you were making boatloads of cash off crypto miners and then AI I suspect the rate of progress fell dramatically for gaming at least (most of the the innovation I suspect went more into high VRAM/memory controllers and datacentre scale interconnects).
It is not just GPU performance, it is that visually things are already very refined. A ten times leap in performance doesn't really show as ten times the visual spectical like it used to.
Like all this path tracing/ray tracing stuff, yes it is very cool and can add to a scene but most people can barely tell it is there unless you show it side by side. And that takes a lot of compute to do.
We are polishing an already very polished rock.
Yes but in the PS1 days we were doing a 1000x compute performance a decade.
I agree that 10x doesn't move much, but that's sort of my point - what could be done with 1000x?
Yeah there’s been a drop off for sure. Clearly it hasn’t been steep enough for game studios to not lean on anyway, though.
One potential forcing factor may be the rise of iGPUs, which have become powerful enough to play many titles well while remaining dramatically more affordable than their discrete counterparts (and sometimes not carrying crippling VRAM limits to boot), as well as the growing sector of PC handhelds like the Steam Deck. It’s not difficult to imagine that iGPUs will come to dominate the PC gaming sphere, and if that happens it’ll be financial suicide to not make sure your game plays reasonably well on such hardware.
I get the perhaps mistaken impression the biggest problem games developers have is making & managing absolutely enormous amounts of art assets at high resolution (textures, models, etc). Each time you increase resolution from 576p, to 720p to 1080p and now 4k+ you need a huge step up in visual fidelity of all your assets, otherwise it looks poor.
And given most of these assets are human made (well, until very recently) this requires more and more artists. So I wonder if games studios are more just art studios with a bit of programming bolted on, vs before with lower res graphics where you maybe had one artist for 10 programmers, now it is more flipped the other way. I feel that at some point over the past ~decade we hit a "organisational" wall with this and very very few studios can successfully manage teams of hundreds (thousands?) of artists effectively?
This hits the nail pretty close to the head. I work on an in-house AAA engine used by a number of different games. It's very expensive to produce art assets at the quality expected now.
Many AAA engine's number one focus isn't "performance at all costs", it's "how do we most efficiently let artists build their vision". And efficiency isn't runtime performance, efficiency is how much time it takes for an artist to create something. Performance is only a goal insofar as to free artists from being limited by it.
> So I wonder if games studios are more just art studios with a bit of programming bolted on.
Not quite, but the ratio is very in favor of artists compared to 'the old days'. Programming is still a huge part of what we do. It's still a deeply technical field, but often "programming workflows" are lower priority than "artist workflows" in AAA engines because art time is more expensive than programmer time from the huge number of artists working on any one project compared to programmers.
Just go look at the credits for any recent AAA game. Look at how many artists positions there are compared to programmer positions and it becomes pretty clear.
That depends a lot on art direction and stylization. Highly stylized games scale up to high resolutions shockingly well even with less detailed, lower resolution models and textures. Breath of the Wild is one good example that looks great by modern standards at high resolutions, and there’s many others that manage to look a lot less dated than they are with similarly cartoony styles.
If “realistic” graphics are the objective though, then yes, better displays pose serious problems. Personally I think it’s probably better to avoid art styles that age like milk, though, or to go for a pseudo-realistic direction that is reasonably true to life while mixing in just enough stylization to scale well and not look dated at record speeds. Japanese studios seem pretty good at this.
Yeah, its flipped. Overall, it has meant studios are more and more dependent on third party software (and thus license fees), it led to game engine consolidation, and serious attrition when attempting to make something those game engines werent built for (non-pbr pipelines come to mind).
It's no wonder nothing comes out in a playable state.
You divided 230 by .03 wrong, which would be 10000-ish, but you underestimated the PS1 by a lot anyway. The CPU does 30 MIPS, but also the geometry engine does another 60 MIPS and the GPU fills 30 or 60 million pixels per second with multiple calculations each.
Not to mention that few developers were doing hand optimized assembly by the time of PSX. They were certainly hand optimizing models and the 3D pipeline (with some assembler tuning), but C and SDKs were well in use by that point.
Even Naughty Dog went with their own LISP engine for optimization versus ASM.
Naughty Dog's GOAL was PS2 specific and essentially chock full of what would be called intrinsics these days that let you interleave individual assembly instructions particularly for the crazy coprocessor setup of Emotion Engine.
My understanding is that the mental model of programming in PS2 era was originally still very assembly like outside of few places (like Naughty Dog) and that GTA3 on PS2 made possibly its biggest impact by showing it's not necessary.
If by "mental model" you mean "low-level" programming, sure. But you might as well conflate "religion" with "Southern Baptist protestantism" then. You're working with the same building blocks, but the programming style is drastically different.
The vast majority of PSX games were done completely in C, period. Some had small bits of asm here and there, but so do the occasional modern C/C++ apps.
To your last point, before there was GOAL there was GOOL (from the horse's mouth itself):
https://all-things-andy-gavin.com/tag/lisp-programming/
And it was used in all of Naughty Dog's PSX library.
By what metric can you say this with any confidence when game scope and fidelity has ballooned?
Because optimized games aren’t completely extinct and there’s titles with similar levels of size, fidelity, and feature utilization with dramatically differing performance profiles.
Given the N64-PS1 era is filled with first party games that run at like 20 fps, I'm having a hard time saying things are worse now.
I am a bit uncomfortable with the performance/quality stuff that people have set up but I personally feel that the quality floor for perf is way higher than it used to be. Though there seem to be less people parking themselves at "60fps locked", which felt like a thing for a while
The current generation has a massive leap in storage speed but games need to be architected to stream that much data into RAM.
Cyberpunk is a good example of a game that straddled the in between, many of it's performance problems on the PS4 were due to constrained serialization speed.
Nanite and games like FF16 and Death Stranding 2 do a good job of drawing complex geometry and textures that wouldn't be possible on the previous generation
Nanite is actively hurting performance
Nanite has a performance overhead for simple scenes but will render large, complex scenes with high-quality models much more efficiently, providing a faster and more stable framerate.
It’s also completely optional in Unreal 5. You use it if it’s better. Many published UE5 games don’t use it.
A lot of the difference went into FPS rather than improved graphics.
And loading times. I think people already forgot how long you had to wait on loading screens or how many faked loading (moving through a brush while the next area loads) there was on PS4
PS4 wasnt too terrible but jumping back to PS3... wow I completely forgot how memory starved that machine was. Working on it, we knew at the time but in retro spect it was just horrible.
Small RAM space with the hard CPU/GPU split (so no reallocation) feeding off a slow HDD which is being fed by an even slower Bluray disc, you are sitting around for a while.
Did you forget that on the N64, load times were near instantaneous?
The N64 was cartridge based.
If only we could just ship a 256GB NVMe SSD with every game and memory map the entire drive like you could with cartridges back then. Never have loading times again.
This is correct. Also, it speaks to what players actually value.
I have played through CP2077 with 40, 30 and 25 fps. A child doesn't care if Zelda runs with low FPS.
The only thing I value is a consistent stream of frames on a console.
When given a choice, most users prefer performance over higher fidelity
I would like to see the stats for that.
> "When asked to decide on a mode, players typically choose performance mode about three-quarters of the time,
From PS5 Pro reveal https://youtu.be/X24BzyzQQ-8?t=172
Seems like an overgeneralization. I get it when FPS players want the best performance: players have FOMO of the best reaction time and the games are more built for fast action than contemplative scenery watching.
I wonder if players of single player action/adventure games make the same choice. Those games are played less (can be finished in 10-30 hours instead of endlessly) so the statistics might be skewed to favor performance mode.
Children eat dirt. I'm not sure "children don't care" is a good benchmark.
Also FPS just requires throwing more compute at it.
Excessively high detail models require extra artist time too.
Yes PS5 can output 120hz on hdmi. A perfect linear output to direct your more compute at.
twice as fast, but asked to render 4x the pixels. Do the math
Well you see... I got nothing.
The path nowadays is to use all kinds of upscaling and temporal detail junk that is actively recreating late 90s LCD blur. Cool. :(
This is the result of an industry wide problem where technology just is not moving forward as quickly as it used to move. Dennard scaling is dead. Moore’s law is also dead for SRAM and IO logic. It is barely clinging to life for compute logic, but the costs are skyrocketing as each die shrink happens. The result is that we are getting anemic improvements. This issue is visible in Nvidia’s graphics offerings too. They are not improving from generation to generation like they did in the past, despite Nvidia turning as many knobs as they could to higher values to keep the party going (e.g. power, die area, price, etcetera).
Jim Keller disagrees: https://www.youtube.com/watch?v=oIG9ztQw2Gc
This article shows how great a leap there was between previous console generations.
https://www.gamespot.com/gallery/console-gpu-power-compared-...
AFAIK, this generation has been widely slammed as a failure due to lack of new blockbuster games. Most things that came out were either for PS4, or remasters of said games.
There have been a few decent sized games, but nothing at grand scale I can think of, until GTA6 next year.
There were the little details of a global pandemic and interest rates tearing through timelines and budgets.
GTA VI is going to be a showcase on these consoles.
> Or the game engines have become super bloated.
"Bloated" might be the wrong word to describe it, but there's some reason to believe that the dominance of Unreal is holding performance back. I've seen several discussions about Unreal's default rendering pipeline being optimized for dynamic realtime photorealistic-ish lighting with complex moving scenes, since that's much of what Epic needs for Fortnite. But most games are not that and don't make remotely effective use of the compute available to them because Unreal hasn't been designed around those goals.
TAA (temporal anti-aliasing) is an example of the kind of postprocessing effect that gamedevs are relying on to recover performance lost in unoptimized rendering pipelines, at the cost of introducing ghosting and loss of visual fidelity.
This is a very one-sided perspective on things. Any precomputed solution to lighting comes with enormous drawbacks across the board. The game needs to ship the precomputed data when storage is usually already tight. The iteration cycle for artists and level designers suchs when lighting is precomputed - they almost never see accurate graphics for their work while they are iterating because rebaking takes time away from their work. Game design become restricted to those limitations, too. Can't even think of having the player randomly rearranging big things in a level (e.g. building or tearing down a house) because the engine can't do it. Who knows what clever game mechanics are never thought of because of these types of limitations?
Fully dynamic interactive environments are liberating. Pursuing them in is the right thing to do.
https://www.youtube.com/watch?v=Ed4vNNQwCDU
In principle, Epic's priorities for Unreal should be aligned to a lot of what we've seen in the PS3/4/5 generation as far as over-the-shoulder 3rd person action adventure games.
I mean, look at Uncharted, Tomb Raider, Spider-Man, God of War, TLOU, HZD, Ghost of Tsushima, Control, Assassins Creed, Jedi Fallen Order / Survivor. Many of those games were not made in Unreal, but they're all stylistically well suited to what Unreal is doing.
I agree. UE3 was made for Gears of War (pretty much) and as a result the components were there to make Mass Effect.
TAA isn't a crutch being used to hold up poor performance, it's an optimization to give games anti-aliasing that doesn't suck.
Your other options for AA are
* Supersampling. Rendering the game at a higher resolution than the display and downscaling it. This is incredibly expensive.
* MSAA. This samples ~~vertices~~surfaces more than once per pixel, smoothing over jaggies. This worked really well back before we started covering every surface with pixel shaders. Nowadays it just makes pushing triangles more expensive with very little visual benefit, because the pixel shaders are still done at 1x scale and thus still aliased.
* Post-process AA (FXAA,SMAA, etc). These are a post-process shader applied to the whole screen after the scene has been fully rendered. They often just use a cheap edge detection algorithm and try to blur them. I've never seen one that was actually effective at producing a clean image, as they rarely catch all the edges and do almost nothing to alleviate shimmering.
I've seen a lot of "tech" YouTubers try to claim TAA is a product of lazy developers, but not one of them has been able to demonstrate a viable alternative antialiasing solution that solves the same problem set with the same or better performance. Meanwhile TAA and its various derivatives like DLAA have only gotten better in the last 5 years, alleviating many of the problems TAA became notorious for in the latter '10s.
Erm your description of MSAA isn't quite correct, it has nothing to do with vertices and doesn't increase vertex processing cost..
It's more similar to supersampling, but without the higher pixel shader cost (the pixel shader still only runs once per "display pixel", not once per "sample" like in supersampling).
A pixel shader's output is written to multiple (typically 2, 4 or 8) samples, with a coverage mask deciding which samples are written (this coverage mask is all 1s inside a triangle and a combo of 1s and 0s along triangle edges). After rendering to the MSAA render target is complete, an MSAA resolve operation is performed which merges samples into pixels (and this gives you the smoothed triangle edges).
> solves the same problem set with the same or better performance
The games industry has spent the last decade adopting techniques that misleadingly inflate the simple, easily-quantified metrics of FPS and resolution, by sacrificing quality in ways that are harder to quantify. Until you have good metrics for quantifying the motion artifacts and blurring introduced by post-processing AA, upscaling, and temporal AA or frame generation, it's dishonest to claim that those techniques solve the same problem with better performance. They're giving you a worse image, and pointing to the FPS numbers as evidence that they're adequate is focusing on entirely the wrong side of the problem.
That's not to say those techniques aren't sometimes the best available tradeoff, but it's wrong to straight-up ignore the downsides because they're hard to measure.
Yeah. Only problem is that overly aggressive TAA implementations blur the whole frame during camera rotation. The thing that is even better than standard TAA is a combination of TAA and temporal upscaling, called TSR in Unreal. Still better is the same system but performed by an ML model, e.g. DLSS. Though this requires special inference hardware inside the GPU.
In the past, MSAA worked reasonably well, but it was relatively expensive, doesn't apply to all forms of high frequency aliasing, and it doesn't work anymore with the modern rendering paradigm anyway.
If the Playstation contributions are good enough, maybe RDNA4 -> RDNA5 will be just as good as RDNA3 -> RDNA4. As long as they get the pricing right, anyway.
We've known this for a while, it's an extension of the upscaling and frame generation AMD already worked on in conjunction with Sony for FSR 3 and to a much greater extent FSR 4. Previous articles also have highlighted their shared focus on BVH optimizations
Excited to see how the software support for UDNA1 works out. Very hopeful we'll see some real competition to Nvidia soon in the datacenter. Unfortunately I think the risk is quite high: if AMD burns developers again with poor drivers and poor support, it's hard to see how they'll be able to shake the current stigma.
Take this with a pinch of salt, but the most recent ROCm release installed out of the box on my WSL2 machine and worked first time with llama.cpp. I even compiled llama.cpp from source with 0 issues. That has never happened ever in my 5+ years of having AMD GPUs. Every other time I've tried this it's either failed and required arcane workarounds, or just not worked entirely (including running on 'real' Linux).
I feel like finally they are turning the corner on software and drivers.
Who is a better computer architect, Mark Cerny or Anand Shimpi?
Did we ever hear what Anand does at Apple?
Yes, but what will use it when there are so few games on the platform in the current PS generation?
When will Sony support Vulkan on PS?
Why would they? They have their own (two actually) proprietary graphics APIs: GNM and GNMX.
I'd ask why wouldn't they. Not a fan of NIH and wheel reinvention proponents.
Sony's low level APIs for PS4 and PS5 (it doesn't use GNM) is almost a direct mapping to the hardware with very little abstraction compared to Vulkan/DX12. Vulkan is still very high level compared to what's going on inside the driver. There's no point paying the cost of Vulkan's abstractions when half the point of a game console is to have a fixed hardware target, hence GNM.
Because if they write their own they get to own the political/bureaucratic portion of the problem. For better or worse, they don't have to deal with the Kronos Group. They get to optimize their APIs directly against their research with AMD.
That still doesn't make NIH a better approach. NIH is dinosaur idea really when it comes to technology like this.
Why would Vulkan, as opposed to a custom solution designed to target that hardware and games specifically, be a better solution?
If you’re making a PS game you’re already doing tons of bespoke PS stuff. If you don’t want to deal with it there are plenty of pieces of middleware out there to help.
Honestly these “where’s Vulkan” posts on every bit of 3D capable hardware feel like a stupid meme at this point as opposed to a rational question.
Maybe they should just ship DX12. That’s multi-platform too.
Because it won't tax developers with the need to learn yet another NIH. Same reason any standard exists and makes things easier for those who use it.
Honestly any idea that defends NIH like this belongs with dinosaurs. NIH is a stupid meme, not the opposite of it.
Vulkan is a mess as bad as OpenGL, to the point Khronos has been forced to public admit it didn't went out as planned, is yet again an extension spaghetti, and they are now trying to clean the house with the Vulkan Roadmap introduced at Vulkanised 2025.
That is how good Khronos "design by commitee APIs" end up being.
Vulkan's extension story is a mess. Not surprising from the people that brought you OpenGL.
However please don't undersell what they got right. Because what they got right, they got _very_ right.
Barriers. Vulkan got barriers so absolutely right that every competing API has now adopted a clone of Vulkan's barrier API. The API is cumbersome but brutally unambiguous compared to DirectX's resource states or Metal's hazard tracking. DirectX has a rats nest of special cases in the resource state API because of how inexpressive it is, and just straight up forgot to consider COPY->COPY barriers.
We also have SPIR-V. Again, D3D12 plans to dump DXIL and adopt SPIR-V.
The fundamentals Vulkan got very right IMO. It's a shame it gets shackled to the extension mess.
SPIR-V only exists to reaction of Khronos losing to CUDA and PTX, it is the evolution of SPIR, created when they realised industry didn't really want to write OpenCL in a textual form of C99 dialect, rather bytecode formats targeted by multiple compiler backends, including Fortran ignored by Khronos.
Microsoft adopting SPIR-V as DXIL replacement is most likely a matter of convenience, the format got messy to maintain, is tied to an old fork of LLVM, and HLSL has gotten the industy weight, even favoured over GLSL for Vulkan (which Khronos acknowledge at Vulkanised 2024 not doing any work at all, zero, nada), so why redo DXIL from scratch, when they could capitalize on existing work to target SPIR-V from HLSL.
DXIL has other issues beyond being an orphaned fork of LLVM-IR. LLVM is ill-equipped to represents shaders, and there have been real bugs in the wild because of how not-a-shader-language LLVM-IR is. See this [0] post under the 'The classic subgroup breakage – maximal convergence gone awry' section. Read the whole series really if you haven't, it's a fantastic dive into how awful this stuff is.
DXIL is also a useless intermediate representation because parsing it is so bloody difficult nobody could actually do anything with it. DXIL was, and still is, functionally opaque bytes. You can't introspect it, you can't modify it or do anything of use with it. SPIR-V is dead-simple to parse and has an array of tools built around it because it's so easy to work with.
I don't really see how the OpenCL history is relevant to Vulkan either. Khronos losing the OpenCL game to Nvidia, certainly no thanks to Nvidia's sabotage either, doesn't change that SPIR-V is a much more successful format.
[0] https://themaister.net/blog/2022/04/24/my-personal-hell-of-t...
DXIL has the DirectX SDK tooling.
OpenCL history is more than relevant, without SPIR, there would not exist SPIR-V.
People love to blame NVidia for AMD and Intel failures pushing OpenCL into the industry, and Google completly ignoring it for mobile devices, pushing their own RenderScript dialect instead, even if Apple would have behave differently regarding their platforms, the other 80% market has completly ignored it.
This comment is just a bunch of bunk, but regardless, Vulkan is the only thing that's a collaborative effort, whether it's a mess or not. NIH proponents have nothing to offer as an alternative.
Market and tooling, industry has usually ignored Khronos APIs when multiple APIs are available to chose from, e.g. Nintendo Switch.
And most of the standard we have now starts with something similar to NIH. Vulkan itself is an offshoot of mantel from AMD. There are valid reason to have a custom api. Especially in domain like game console with hardware with long release cycle, tight performance requirement and legacy (ps4) code to support.
Good, but if it's actually a standard - that's the benefit. If it's not - it's just NIH as in lock-in.
AMD’s next-gen GPUs may have some PlayStation tech inside.
I don't think he is employed by Sony, but work as a consultant for them. So both Sony PS 4/5 and AMD GPUs have his tech inside.
So you're right, though I would never have guessed— in the PS5 hype cycle he gave that deep dive architecture presentation that for all the world looked like he was a Sony spokesperson.
It looks like all of your comments are low-effort summaries like this. What’s going on here? Or is this a bot…
They're summarizing the submissions they're making. All of the summary comments are on their own submissions.