Question Work Graphs - promising and exciting but when???

[igor_kavinski]

Microsoft releases Work Graphs API enabling advanced GPU-driven rendering - VideoCardz.com

Agility SDK 1.613: Shader Model 6.8 and Work Graphs Microsoft released a new Agility Software Development Kit that that enables new features. The Agility SDK 1.613 comes with Shader Model 6.8 and Work Graphs support. The latter is an important feature that will promote the use of GPU-driver...

videocardz.com

From this post https://developer.nvidia.com/blog/work-graphs-in-direct3d-12-a-case-study-of-deferred-shading/

On a GeForce RTX 4090 GPU, the work graph lights the scene at 1920 x 1080 within 0.8 ms to 0.95 ms, whereas the uber shader dispatch technique takes 0.98 ms.

So roughly a 20% improvement in lighting performance in this particular sample. Hopefully developers can squeeze more out of it.

GPU support as of now:

Geforce 3000 and 4000 series
RDNA3

RDNA2 support may become available in Q2 2024 (AMD is silent about that).

No word from Intel. Keeping fingers crossed!

 

[BFG10K]

We were promised better performance and lower overhead from DX12 / Vulkan ten years ago when released. Instead most modern games run like garbage and look like a Baboon's rear end. They also consume horrific storage space and have incurable shader compilation stutter.

I just replayed OG Doom 3 from 2004 and the visuals are still amazing for a 20 year old game, with none of the aforementioned issues. Also just 1.5GB for a solid ~12 hour single player campaign with no micro-transactions, "games as a service", loot boxes, or Darth Vader skins.

 

[igor_kavinski]

I just replayed OG Doom 3 from 2004

Check out the original Riddick game. With max settings, it looks beautiful and runs at 100+ fps on even a measly Quadro in my Ivy Bridge Thinkpad.

 

[MoogleW]

Apparently Epic Games pushed for this one so there is hope for a wider adoption, especially by companies like CDProjektRed who are migrating to this engine and are not shy about trying cutting edge GPU technology. Hopefully its merged into the main branch of Unreal Engine

 

[GodisanAtheist]

Keeping my fingers crossed for RDNA2 support...

 

[igor_kavinski]

AMD showcases a demo with larger perf improvement than nVidia! https://videocardz.com/newz/amd-dem...phs-on-radeon-rx-7900-xtx-with-39-improvement

Hopefully it's just a matter of enabling this functionality on RDNA2 via a driver update.

 

[marees]

developers also noted that this version also brings experimental support for D3D12 Work Graphs. They note that there are no games that use this technology yet, but based on their testing of it, it already runs faster than native implementation after converting it to normal compute shaders.

Source: VKD3D-Proton

https://videocardz.com/newz/vkd3d-proton-3-0-brings-amd-fsr4-to-proton-with-an-experimental-support-for-older-gpus

 

[Bigos]

I think this has been cooking for a year. The "better than native" was already found by then.

vkd3d-proton/docs/workgraphs.md at workgraphs · HansKristian-Work/vkd3d-proton

Fork of VKD3D. Development branches for Proton's Direct3D 12 implementation. - HansKristian-Work/vkd3d-proton

github.com

I guess this might change with new hardware that is built for the new API. Or once proprietary driver writers actually need to optimize the thing when a user is found.

 

[MrMPFR]

I think this has been cooking for a year. The "better than native" was already found by then.

vkd3d-proton/docs/workgraphs.md at workgraphs · HansKristian-Work/vkd3d-proton

Fork of VKD3D. Development branches for Proton's Direct3D 12 implementation. - HansKristian-Work/vkd3d-proton

github.com

I guess this might change with new hardware that is built for the new API. Or once proprietary driver writers actually need to optimize the thing when a user is found.

AMD did manage a massive speedup (+64%) with their GDC 2024 demo, but that with PCG content. Very odd.

Probably a combination of both.

 

[MrMPFR]

Ashes of the Singularity II Exclusive Interview – Lower APM, Three Factions, and the Nitrous Engine Evolved

Oxide Games reveals new details on Ashes of the Singularity II, from the human UEF faction and AI improvements to modding and gamepad support

wccftech.com

As for Work Graphs, we are not using them because we have implemented our own subset internally. After consultation with IHVs, it was unclear if the native support would be significantly faster than our current implementation. Additionally, we would still need our current implementation for older hardware anyway (e.g., GTX 970 or RX 580). However, we’ve given the IHVs access to our code so they can experiment with it for future architectures.

In terms of the most impactful, it’s hard to say. Unfortunately, Tiled Resource Tier 4 is difficult to use properly on a PC architecture, so we don’t think that it will have a huge impact. Work Graphs definitely have the potential to be most impactful, but will likely need a generation or two of hardware advancements to really see the benefit.

It's interesting to hear devs talk about this stuff even if it's very rarely a question being asked. Just a shame so many people are stuck on old HW due to financial concerns and HW stagnation. Maybe Ashes of the Singularity III will use Work Graphs when it releases in the late 2030s xD

But it seems seems increasingly likely that RDNA 5 a complete architectural reset for the Work Graphs paradigm. Mantle + Async compute 2.0. This time on a different level because Work Graphs is the biggest API shift since programmable shaders in the early 2000s.

Also can we please get some more new EI vs Work Graphs perf comparisons. Haven't seen anything from the IHVs for nearly two years since the GDC 2024 numbers that @igor_kavinski linked to awhile back.

 

[CINEBENCH_ACCELERATOR]

AMD did manage a massive speedup (+64%) with their GDC 2024 demo, but that with PCG content. Very odd.

Probably a combination of both.

+64% is crazy.

Game devs need to utilize Work Graphs and new hardware (RDNA 5) needs to support it? Integrated support for Unreal Engine 6 or 7?

Looks like games will be more efficient in near future.

 

[MrMPFR]

+64% is crazy.

What's crazy is it achieving those speeds with the barebones RDNA 3 support. RDNA 5 is when they actually bother to design architecture around new paradigm.

The HPG 2025 Tree paper is mental and will look forward to see how much faster RDNA 5 is here on a per CU basis.

Game devs need to utilize Work Graphs and new hardware (RDNA 5) needs to support it? Integrated support for Unreal Engine 6 or 7?

Looks like games will be more efficient in near future.

Some work on Linux side (Proton) to emulate it on older HW, because Vulkan doesn't have native support. But otherwise Ampere+ and RDNA 3+.
Yes games need to be rewritten from scratch. Complete rewrite + can't run on current gen consoles so it'll be a while.

Yeah that's very likely. Work graphs is something Epic requested for a long time. I'm sure UE6 will use it.

It's not so much about efficiency but about doing wild stuff that just isn't possible today. This and neural shading will unlock a new tier of immersion and realism.

 

[MrMPFR]

Looks like games will be more efficient in near future.

I was wrong about it mostly being forward looking. There's potential for significant optimization in the short term. For example it looks like they could achieve similar or greater speedups for ray shading. Material shader divergence is a big problem. I've explained in a related thread how Work Graphs can overcome this issue. Now if they can get anywhere near close to pixel shader occupancy then it's a game changer for path tracing.

Benefits extend to existing architectures but again it needs RDNA 5 to shine.

+64% is crazy.

Nothing compared to this insanity: https://dl.acm.org/doi/10.1145/3695053.3731060
SpMV really likes Work Graphs. Mean = 3.35x and up to 7.19x.

Work Graphs mem alloc is the equivalent of Nanite (triangles). Fixed cost regardless of input complexity. Work graph complexity decides mem alloc (scratchpad), not worst case static allocation.

This, PCG, ML asset compression, and nextgen BVH compression paradigms will be a memory massive multiplier for the nextgen consoles. >10X.

 

[MrMPFR]

Microsoft prob needs to start over from scratch with DXR 1.3 with a focus on work graphs. Having traversal + shading in one big graph is not feasible under DXR 1.2. AMD has the opportunity to make SER irrelevant long term. Work graph for entire RT pipeline (traversal and shading) tackles all problems, not a band aid fix like SER. Improved occupancy, no barriers, bubbles or empty launches, massively improved coherency etc... Def better than SER coherence when executing ray tracing workloads.
Compared this to SER is a band aid.

All NVIDIA cards from Ampere and newer + AMD RDNA 3-4 cards have RT fine wine waiting to be tapped. Cards without SER stand to benefit the most, Ampere (if we ignore OMM) on NVIDIA side should be much closer to 40-50 series. RDNA 3 will benefit the most but RDNA 4 a ton as well. If devs bother to rewrite RTGI pipelines on PC there's potential for massive perf gains in the short term, otherwise this is firmly post-crossgen territory unfortunately.

Regular gaming workloads can benefit too, but will unlock new possibilities for GPU driven procedural content generation, complex systems on GPUs (AI and physics), neural shading, and path tracing.

As previously said extreme HW+SW co-design with will fully unlock this for GFX13. RDNA 5 is going to be in a different league here, especially for weird stuff like PCG. It'll be interesting to see how synergies between this and the reworked RT traversal logic will play out in path tracing, but safe to say it'll embarass 50 series. As for regular Execute Indirect + PT hard to say but if compiler is clever then it should easily win here at iso-raster as well.

If NVIDIA doesn't prepare for this on 60 series and later gens and later they'll get annihilated in benchmarks when work graphs adoption starts to pick up.

Since it's programmable shaders 2.0 it might warrant its own thread TBH. Or @igor_kavinski maybe consider renaming this thread to something more general?

Am I the only one that's excited about GPU work graphs? What do you think @basix and @soresu?

 

[MrMPFR]

Work graphs is something Epic requested for a long time. I'm sure UE6 will use it.

It'll expand on this a bit. UE6 will likely use work graphs extensively. They tried wild stuff with UE5 why not more with next update like @MoogleW said almost two years ago.

Just untangle GPU from CPU + let Work Graphs do its magic. Remember people were blown away by UE5, but unlike UE5, UE6 will deliver because workgraphs "just work", no more low level API ressource management nightmare, just let GPU do its thing.

Here's the short version: Work graphs has the potential to make SER completely redundant and revolutionize gaming. Even regular game rendering workloads, but especially for procedural content generation, running complex systems on GPUs (AI and physics), neural shading and path tracing.

If I were to guess outside of limited AMD demos, UE6 is prob be the first to properly demo just how impressive work graphs will be.

 

[NTMBK]

If I were to guess outside of limited AMD demos, UE6 is prob be the first to properly demo just how impressive work graphs will be.

I sincerely doubt it to be honest.

Unreal has to cater to a huge range of targets. They need a mobile-suitable renderer, they need a renderer that can target current gen hardware effectively, they need to utilise the next gen hardware, they need to scale all the way down to integrated graphics in laptops and up to 500W monster GPUs. They fundamentally can't go "all in" on a huge rewrite of the render pipeline to maximise the effectiveness of work graphs, because doing so would lock them out of a ton of places they need to be. Even if they roll out work graph support in some places, I expect it will still be bolted onto a render pipeline which is fundamentally not designed around it.

Given the current catastrophe in hardware pricing and availability, it would be foolish for any studio to stake their future on work graphs. If the next console generation even launches, it will be either heavily delayed, insanely highly priced, or in vanishingly small quantities. (Or a combination of all three!) Developers would be better suited targeting lower end systems that users might actually be able to acquire, either old systems still clinging onto life or low end new systems with limited RAM and a small or integrated GPU. I hope Epic are putting more focus into getting their engine to run better on today's hardware, because that might be all that is available to the majority of players for a long time.

 

[MrMPFR]

They can ship legacy broader market version + upgraded full version for PC with all the bells and whistles alongside each other. Seen that before with DX11/DX12 in late 2010s and early 2020s.

Hope it's not bolted on that would indeed limit the benefits.

Work graphs is supported on 30-50 series and RDNA 3 and newer. For bleeding edge AAA and AA post-crossgen games launching in >2032 or later that sounds good enough.

In the mean time they can keep refining UE5 and I too hope they continue to work towards optimizing it further. Rn it's still a mess.

Your outlook is very grim. How long do you guess it'll be before we see market normalize?

 

[NTMBK]

Your outlook is very grim. How long do you guess it'll be before we see market normalize?

With how chaotic the world is right now it's very hard to say. It depends on how long it takes this AI bubble to pop, whether China invades Taiwan, whether the USA stops imposing arbitrary tariffs at 10 minutes' notice, etc etc. But I'll save the in depth grumbling for P&N.

At best, I'm hopeful the games industry stabilises again in 5 years or so.

 

[Win2012R2]

it would be foolish for any studio to stake their future on work graphs

Surely most of game dev is to do with stuff that got nothing to do with work graphs? If the core rendering supports that then everybody else can just do their stuff (models, scripts, textures) and the engine will apply all that much more effectively using work graphs (when they are hardware supported)?

 

[MrMPFR]

They need a mobile-suitable renderer, they need a renderer that can target current gen hardware effectively, they need to utilise the next gen hardware, they need to scale all the way down to integrated graphics in laptops and up to 500W monster GPUs.

Emulation is possible as shown with vkd3d-proton. If the HW supports newer Vulkan it's prob gonna happen. The downsides of EI are too big, devs and everyone else will move to new paradigm ASAP. IHVs, MS and Khronos will fill in the gaps for HW that doesn't have native support.

Early on it should make existing code run better and selective implementations of new stuff.
Long term completely re-envisioned pipelines and heavy focus on GPU driven simulation systems and procedural generation.

UE6 will be built with this in mind.