That RDNA 1.8 Consoles Rumor *spawn*

We do not have a single public technical information to talk about. So it's gossip. So it's meaningless.

And it's perfectly reasonable that the ISA of the upcoming consoles are not 100% equal to the ISA of the future AMD RDNA2 PC GPUs. As any other past console.

EDIT: by a developer point of view, a possible lack of a new geometry pipeline in a mesh+amplification shader stage-like form and the lack of a programmable sampler/sampelr feedback would be a big lack. And yes, I do not care about ray-tracing as I do not care if we can make a colony on Mars.
 
Last edited:
VRS (MS Tier) and SFS (not simple Sampler Feedback!) will have significant performance enhancement features. These two features together can significantly improve performance, so it’s not worth underestimating them.
 
The tldr; as I see it,a Neogaf user who is apparently a long time dev tweeted a DM to architecture engineer for PS5 and got back a return statement.
But because this guy was blasted by everyone for being wrong, he posted the DM and pulled it down immediately but everyone screenshot it and it’s been circulating.

I feel pretty bad for the dude who's DMs got shared. Sucks to be in the spotlight at such an important time. If your friends tell you stuff that might be sensitive, best to keep it to yourself for their sake.

I had given it some thought and there might be some truth. He was not even the first person to bring it up, back in 2018 this was also discussed.

Between the Github leak naming, and the radio silence on VRS and Sampler feedback, I've been considering it a distinct possibility for a while now. So much so that this doesn't actually surprise me.

In the end it doesn’t matter if there is feature parity. It’s just a label, unless the PC RDNA2 is completely outperforming both consoles by a sizeable amount, which some news outlets and Odium also have committed to.

We will just have to wait and see if the statements are true. As Jay wrote, if RDNA 2 fails to see light this year and the longer it takes to release may indicate it was never ready in time for production for both consoles.

That's a good point. MS may not have the full suite of "PC RDNA2" technology either, for one reason or another. Though so far, they seemed to have ticked all the publicly revealed boxes. Hotchips in a couple of weeks might offer more insight...

Dr David Kanter said RDNA1 architecture retains elements of GCN, so I wonder if Sony chose to forgo certain RDNA2 features in order to retain the necessary GCN arch to help facilitate a more hardware approach to backwards compatibility?

Sony's software stack is not as good as MS, coupled with PS4's lower level GNM API and the fact they'd need to recompile shader binaries packaged into PS4 games/downloads were a hill they chose not to climb with a more software based BC approach and so went with a hardware based approach to BC with a GPU which retains some GCN elements (present in RDNA1 arch)?

This is a really interesting point. If they chose 36 CUs for BC reasons it figures that there may be other aspects of the PS5 hardware determined by it. Even MS have made mention of having to do work at the hardware level for BC, despite the extra hardware abstraction that they've built into their systems.

Other possible reasons might be:
- Was planned for 2019, but slipped
- Extensive customisation meant having to branch from available RDNA technologies sooner

... but I have to say I like the possibility you've raised here the most. Actually ... you know those rumours about "AMD building RDNA for Sony"? Maybe some aspects of RDNA 1 actually were determined by Sony's needs.
 
The big issue that bothers most people raising these questions of RDNA 2.0 vs 1.9 (or maybe even RDNA 1.9.2.85.3 for that matter) is: Does it support Mesh/Geometry Shaders. Does it support Sampler Feedback. Does it support VRS.

Best case scenario: Yes, 100% to all of the above.
Worst case scenario: Pretty much yes, 90% to all of the above.

Let me explain.

There is no way PS5 does not have something akin to mesh shaders. AMD has been advertising this "unlocking" of the geometry pipeline for half a decade already. It was one of their main game changing features for their (then) future architectures. Then it was advertised as Next Gen Geometry Pipeline.

Now, does the implementation in PS5 work and perform exactly the same way as it does in RDNA 2.0 for pc under DX12 or vulkan? Best case scenario: Yes, and then some. Worst case scenario: yes for 90% of use cases, with slight workarounds for the others. That is it, that is the worst it gets, and I state this completely pulled out of my ass, but trust me on that one.

Sampler Feedback & VRS: those are not as much cornerstones of AMD vision for GPU arch as Geometry was, so I think those have a more solid chance of actually not making it to PS5. Yet, it really matters very little. The things they achieve can be done (and with good performance) with other aproaches.

PS4 PRO has been using checkerboarding and reconstruction extensively which adresses the same things as VRS. I really think VRS is way more usefull for DX12 scenarios, where you cant cant build algos that are optimised for the exact particularities of one GPU. It is very likely that hand crafted compute shaders and rasterization tricks targetting PS5 can perform better and achieve better results than VRS ever will on PC.

Sampler feedback also does things that can, and in fact have been, adressed in software before. Aka. RAGE. God, I keep mentioning this title, but honestly, anybody who wants to discuss and speculate about Virtual Texturing HAS TO re-read their papers and re-watch their presentations. How did the Rage Engine know which texture pages to load into their cache and at which mips? That is an interesting problem which saw a lot of experimentation from Carmak, which he describes by the way. His ultimate solution was quite simple. Rage renders a "proxy camera" encoding UV's only and discovers the needed pages from that. There are thousands of ways to optmize that by the way. Render at lowe res, render only section of the frustrum each frame and get the coverage temporally spread across multiple frames. One can also improve results by having a wider FOV on that proxy, or randomily render things behind the camera every so often to cache the full 360... Here, Sampler Feedback really is a WAY more ellegant way to solve the problem, and a great loss if not present for that reason. But there is still a viable workaround, if not more. The workaround performance is also clearly less performant, but probably it is a drop in the ocean for modern engines.

1. There are zero scenarios where a naive solution like checkerboarding will yield better results than VRS. MS's solution in particular enables the calling of a shader per primitive with an almost infinite amount of ways of calculating the shading rate. Lack of VRS is a huge miss.
2. Virtual texture streaming is old news and, as the SIE engineer said, was already being done using shaders on the PS4. However, the old way of doing things cannot determine with certainty what level of a given mip was being sampled. Also, sampler feedback streaming, or at least MS implementation, will provide gains in latency, granularity and efficiency which are crucial when targeting textures on a per frame basis in a JIT system. The PS5 does away with this possible setback by having an incredibly fast throughput such that it can well afford to be more inefficient.
3. I am not familiar with the differences (if any) between the geometry engines of RDNA 1 and the putative mesh shaders of RDNA 2. So won't comment on this.
 
My take is it's all complete bullshit and was attention seeking by certain people and places.

In the end, the version labeling is a matter of pedantry that won't make any difference. The consoles have whatever hardware features they have. The devs will make exceptional use of both platforms, regardless of what they're marketed as having. The games will be amazing on both. Gamers will truly enjoy the NextGen consoles.
Featuresets are not pedantry. Will indicate whether performance of the two platorms are comparable.
 
Featuresets are not pedantry. Will indicate whether performance of the two platorms are comparable.
Nothing is released yet we don’t know if they haven’t announced thing because they don’t want to. Or they SDK is not ready. Or they do not have it. Etc

only benchmarking can compare the performance of the two systems. This is speculation of have and have nots and relaying that to game performance before either has been released.
 
Of course... BC
Yeah, I think @Pixel has come up with a really sensible possibility there.

BC can put real constraints on the evolution of systems.

I don't really buy the 2019 release theory. Can a one year delay cause such paucity of features? VRS and mesh shaders were not developed in one year time.

I mean, if all goes to plan for AMD you're looking at less than 18 months between the release of RDNA 1 and RDNA 2 for PC. GPUs are highly modular and there's always going to be a time when a feature isn't available, or it is.

That said, AMD have to serve a number of markets and customers. Bringing VRS and Sampler Feedback (useful for more than just efficient, low overhead texture streaming!) to their latest architecture was essential. Simultaneously backporting it to a single-customer design based off an older RDNA derivative may not have been viable even if Sony was prepared to pay for it

Delays, BC, or even a combination of both could be an entirely realistic reason for such a feature disparity. And if BC helps Sony gain more momentum early on than SF or VRS, can you honestly say it was a bad idea?

only benchmarking can compare the performance of the two systems. This is speculation of have and have nots and relaying that to game performance before either has been released.

True, but speculation is a blast. It lets the mind run free! I still remember fondly being one of the very first people ON EARTH to suggest that the Wii had 192 shaders. Woop woop!*

*I've also suggested many things that turned out to be oh so wrong.
 
Between the Github leak naming, and the radio silence on VRS and Sampler feedback, I've been considering it a distinct possibility for a while now. So much so that this doesn't actually surprise me.
.
Since when was a radio ever silent, except when switched off/down? :cool:
 
Back
Top