Current Generation Hardware Speculation with a Technical Spin [post launch 2021] [XBSX, PS5]

Status
Not open for further replies.
ML is going to be very important for AI, arguably in same way RT can aid in AI calculations, but I think ML will play a much bigger role out of the two overall in that aspect. To that end I wonder how much of MS's custom support for extended ML low-precision math took cues from CDNA design-wise but that's almost impossible to speculate on since IIRC there's no schematics, diagrams or even whitepapers for CDNA available.

From what I've been reading, it seems PS5's support for ML extends to FP16 and INT16, which makes sense considering PS4 Pro already supported FP16. So MS would have, essentially, added support for INT8 and INT4, and of course they have more CUs therefore they can dedicate more in parallel for grouped ML calculations. I'm still curious if there's anything else they've done on the GPU side specifically for DirectML; if their general tech (ish?) event hasn't already happened this month then hopefully some more details emerge around that time.

Some of the mesh shader test results have been...massive, to say the least. 1000% increases on some GPUs. In terms of actual practicality running real games though, that will probably come way down. Still though, even a 10-fold reduction would give 2x FPS in practice, that can stack with other things like what MS are already doing with their FPS enhancement tech (if this can work on older BC games, can it also theoretically be done for newer games, even if it just helps increase FPS a bit rather than 2x/4x multipliers?), that all starts to add up.

I'm sure that ML is going to be mega important going forwards, but because I know so little about the mechanics of how you make it work, I'm trying not to get myself too hyped too early. Like you, I've not seen anything for XSX beyond the reduced precision data formats.

With mesh shaders, those super impressive benchmark results are probably worst case type scenarios to stress this particular area of performance (for benching purposes, naturally!). In a real game you'd probably be looking at much smaller gains as you're so frequently bottlenecked in other areas. Googling around I found this Nvidia page on mesh shaders (I didn't read it all), but there is this paragraph near the start:

https://developer.nvidia.com/blog/using-mesh-shaders-for-professional-graphics/

"Keep in mind that mesh shaders are deliberately designed to expand existing capabilities and allow you to optimize certain use cases. They are not meant to replace the existing geometry pipeline completely. Due to the existing optimizations in the traditional VTG (vertex-, tessellation-, and geometry-shader) pipeline, do not expect mesh shader pipelines to always provide substantial wins in performance."

Anyway, hopefully PS5 has something similar. Primitive shaders on their own wouldn't seem to me to be able to do everything described with mesh shaders, but that's most definitely my "not a pro" opinion. And even if that's true, PS5 might go beyond anything we've seen described by AMD.

I guess it could depend on what kind of shader driven geometry you can push into the hardware rasteriser....? :???:
 
Mesh shaders to me is one of those things that should make current gen awesome once engines are upgraded to take advantage of the hardware. Similarly I believe ray tracing can be a game changer this gen once the best ways to build all new engines are figured out. Heavy reuse of rays, very smart distribution of ray budget and much improved noise filtering would be the things I'm looking forward to this gen. And of course the unreal5 could be current gen game changer as well. I have high hopes metro exodus enhanced edition would be one of the first games really using ray tracing to max.Temporal algorithms to improve perceived quality is another thing I'm expecting to happen this gen.

It would be kind of sad if next gen is just faster ray tracing hw. Much faster ray tracing hw is something I think can happen if it's focus of next gen instead of ray tracing being an additional cheap feature tagged along existing hw. I'm hoping neural nets or some other new algorithms would produce giant jump in rendering quality. Of course it's so unproven that I wouldn't bet anything on it. Maybe it's unreal5 that is the giant jump. In that case we will see the jump during this gen already :)
 
But new games are already using the hardware as much as they can, given the nature of their code at present.
Are you sure you're not thinking of interpolated frame inserts? Example: https://www.eurogamer.net/articles/digitalfoundry-frame-rate-upscaler-tech-interview
I don't think that's what Microsoft is doing. I could be mistaken, though.

Actually I don't think the new games are leveraging most Series X or PS5 features at all. We know most Series games aren't using DirectStorage, for example, because DirectStorage isn't even available yet on PC and these are mainly cross-gen games with PC ports in mind, so no need to develop them with features like DirectStorage in mind, yet.

Most games are barely using hardware features like VRS, either. Same can be said of the cache scrubbers in PS5, tho I'm not 100% sure how those actually work (i.e are they something that automate themselves or do developers have to target use of them. I'm guessing if the latter they would have to use demand scrubbing; given the scrubbers are in the caches IIRC devs would have to issue the command to scrub cache from the frontend of GPU).
I think the hope is to find alternative cheaper approximate ways to compute various things. Cloth/water and general physics simulations with neural networks is one very interesting domain. Similarly perhaps we can get away by rendering certain things at lower resolution/rate and let neural networks do their thing. It's all very new and unproven. Lot of papers and demos show promise though.



I believe the one of the dreams developers joined nvidia. I wonder if he also is working on the signed distance fields with neural nets approach/was that the reason he joined nvidia.

If there is to be next generation console I wouldn't be at all surprised if the next gen feature was neural networks and drastic change in rendering capability. Just adding more of the same is not easy road anymore. Manufacturing improvements for making better chips isn't anymore what it used to be. Cost is big issue as is power consumption.

Any ideas on what type of hardware would need to go into an embedded design for neural network support? I read somewhere that there's a memory technology that was being tested for use as array representations of neural networks. I can't recall where I first read about this, but doing a quick search I came across this research document testing MRAM-DIMA:
http://shanbhag.ece.illinois.edu/publications/ADP_ISCAS_MRAM_2019.pdf

It's meant to be a physical implementation of a DNN as opposed to a digital one, and give better performance. Technologies like MRAM show a lot of promise and companies like Everspin already produce chips of...decent capacities (if a few 4 MB is considered decent), though prices are still prohibitive. Maybe as costs come down on technologies like MRAM or FRAM, we can see more of it in embedded console designs for DNNs?

Although I was also thinking to some extent FPGA logic could be integrated into embedded console designs and you can leverage LUT RAM and BRAM there and configure the logic cells to replicate DNNs physically (among other uses). I don't know exactly how "powerful" (I guess in terms of logic cell count) of FPGA slices/cores/whatever you would need for this to be done.

I'm sure that ML is going to be mega important going forwards, but because I know so little about the mechanics of how you make it work, I'm trying not to get myself too hyped too early. Like you, I've not seen anything for XSX beyond the reduced precision data formats.

With mesh shaders, those super impressive benchmark results are probably worst case type scenarios to stress this particular area of performance (for benching purposes, naturally!). In a real game you'd probably be looking at much smaller gains as you're so frequently bottlenecked in other areas. Googling around I found this Nvidia page on mesh shaders (I didn't read it all), but there is this paragraph near the start:

https://developer.nvidia.com/blog/using-mesh-shaders-for-professional-graphics/

"Keep in mind that mesh shaders are deliberately designed to expand existing capabilities and allow you to optimize certain use cases. They are not meant to replace the existing geometry pipeline completely. Due to the existing optimizations in the traditional VTG (vertex-, tessellation-, and geometry-shader) pipeline, do not expect mesh shader pipelines to always provide substantial wins in performance."

Anyway, hopefully PS5 has something similar. Primitive shaders on their own wouldn't seem to me to be able to do everything described with mesh shaders, but that's most definitely my "not a pro" opinion. And even if that's true, PS5 might go beyond anything we've seen described by AMD.

I guess it could depend on what kind of shader driven geometry you can push into the hardware rasteriser....? :???:

I think PS5 should be capable of things vaguely similar to Mesh Shaders with its Primitive Shaders, but clear and away better? No. I try keeping in mind that in all likelihood, Sony started development on PS5 earlier than Microsoft did the Series systems, which means things like deciding between the Primitive Shaders (which IIRC AMD could not get functioning properly in Vega, hence why they retooled with Mesh Shaders, that way they could get certified as being DX12-compliant) or Mesh Shaders would've came up likely early on.

Basically, if Sony came up with some crazy stuff for Primitive Shaders, would AMD not have had enough time to look at that and try bringing that into RDNA 2 GPUs via the Mesh Shaders? I get that Sony and Microsoft had their own separate AMD teams helping to design their systems, but AMD had been working on RDNA 2 for a while by the time core aspects of PS5's design were being wrapped up. Or maybe it'd of been too late to go back in on RDNA 2 design blueprints and work in customizations Sony did with Primitive Shaders into the batch of GPUs which've since been launched?

In any case, I think whatever Sony have done with their Primitive Shaders should be good enough to bring general performance parity in terms of feature support with the Mesh Shaders, tho some of it might have to come in through the Geometry Engine as well. But I'm not expecting customizations on the Primitive Shaders front a generation ahead of things already on the market; when Cerny spoke of GPUs bearing witness to fruitful collaborative efforts coming to market, his timeline seemed focused on RDNA 2 GPUs, not necessarily RDNA 3 ones. Though since RDNA 3 will repurpose a lot of features designed for RDNA 2, on a technical level they would be correct ;)
 
Last edited:
Any ideas on what type of hardware would need to go into an embedded design for neural network support? I read somewhere that there's a memory technology that was being tested for use as array representations of neural networks.

Very fast large matrix multiplications + memory bandwidth to feed the beast. If one believes Keller then add graph support on top of that. I find jim keller's current work fascinating: https://www.tenstorrent.com/
 
Actually I don't think the new games are leveraging most Series X or PS5 features at all. We know most Series games aren't using DirectStorage, for example, because DirectStorage isn't even available yet on PC and these are mainly cross-gen games with PC ports in mind, so no need to develop them with features like DirectStorage in mind, yet.

Most games are barely using hardware features like VRS, either. Same can be said of the cache scrubbers in PS5, tho I'm not 100% sure how those actually work (i.e are they something that automate themselves or do developers have to target use of them. I'm guessing if the latter they would have to use demand scrubbing; given the scrubbers are in the caches IIRC devs would have to issue the command to scrub cache from the frontend of GPU).


Any ideas on what type of hardware would need to go into an embedded design for neural network support? I read somewhere that there's a memory technology that was being tested for use as array representations of neural networks. I can't recall where I first read about this, but doing a quick search I came across this research document testing MRAM-DIMA:
http://shanbhag.ece.illinois.edu/publications/ADP_ISCAS_MRAM_2019.pdf

It's meant to be a physical implementation of a DNN as opposed to a digital one, and give better performance. Technologies like MRAM show a lot of promise and companies like Everspin already produce chips of...decent capacities (if a few 4 MB is considered decent), though prices are still prohibitive. Maybe as costs come down on technologies like MRAM or FRAM, we can see more of it in embedded console designs for DNNs?

Although I was also thinking to some extent FPGA logic could be integrated into embedded console designs and you can leverage LUT RAM and BRAM there and configure the logic cells to replicate DNNs physically (among other uses). I don't know exactly how "powerful" (I guess in terms of logic cell count) of FPGA slices/cores/whatever you would need for this to be done.



I think PS5 should be capable of things vaguely similar to Mesh Shaders with its Primitive Shaders, but clear and away better? No. I try keeping in mind that in all likelihood, Sony started development on PS5 earlier than Microsoft did the Series systems, which means things like deciding between the Primitive Shaders (which IIRC AMD could not get functioning properly in Vega, hence why they retooled with Mesh Shaders, that way they could get certified as being DX12-compliant) or Mesh Shaders would've came up likely early on.

Basically, if Sony came up with some crazy stuff for Primitive Shaders, would AMD not have had enough time to look at that and try bringing that into RDNA 2 GPUs via the Mesh Shaders? I get that Sony and Microsoft had their own separate AMD teams helping to design their systems, but AMD had been working on RDNA 2 for a while by the time core aspects of PS5's design were being wrapped up. Or maybe it'd of been too late to go back in on RDNA 2 design blueprints and work in customizations Sony did with Primitive Shaders into the batch of GPUs which've since been launched?

In any case, I think whatever Sony have done with their Primitive Shaders should be good enough to bring general performance parity in terms of feature support with the Mesh Shaders, tho some of it might have to come in through the Geometry Engine as well. But I'm not expecting customizations on the Primitive Shaders front a generation ahead of things already on the market; when Cerny spoke of GPUs bearing witness to fruitful collaborative efforts coming to market, his timeline seemed focused on RDNA 2 GPUs, not necessarily RDNA 3 ones. Though since RDNA 3 will repurpose a lot of features designed for RDNA 2, on a technical level they would be correct ;)
Regarding PS5 I/O and cache scrubbers, I'm pretty certain Cerny was saying it’s all automatic and the devs don’t even need to understand, just say ‘I want that data put there’ and it just works.

This is why I was thinking PS5 is showing advantages in some scenarios vs XSX where devs need to utilise the functionality.
 
Last edited:
https://eetimes.jp/ee/articles/2104/05/news012.html

All the same year


dp9RnnE.jpg
 
They both only advertised that their solutions are RDNA 2 based. Not that they are 100% RDNA 2 carbon copies.
Yes, if you want to be technical about it, it does look like XSX has more RDNA 2 blocks. And, I guess possibly Zen 2 blocks with some extra customizations of their own.

Not being RDNA 2 in itself doesn't make the system deficient however, it just makes it not RDNA 2. They both set out to solve the same problems, but their approaches are different.
a) Solving the memory cost problem. At a maximum of 16GB to keep costs down, they needed to find a way to increase fidelity while keeping within the limits of 16GB.
Sony:
5.5 GB/s SSD + Kraken Compression
This gets them 80% of the way there. The remaining 20%, the memory has to get into cache for processing. So what to do? Cache scrubbers to remove that last bit of random latency to go from textures of SSD directly into frame.

MS:
2.5 GB/ssd + BCPack Compression
This gets them 50% of the way there. The remaining 50%? Sampler Feedback Streaming, only stream in exactly what you need, further reducing what needs to be sent and when. That gets you 30% of the remaining way. You still got to get it through cache, instead of cache scrubbers their SFS will use an algorithm to fill in values for textures being called by SFS. If for whatever reason the texture has not arrived in frame, it will fill in the texture with a calculated value while it waits for the data to actually arrive within 1-2 frames later.

b) Approaching how we redistribute rendering power on the screen?
MS exposes VRS hardware
PS5 in this case will have to rely on a software variant

C) Improving the front end Geometry Throughput, Culling etc.
MS aligns features sets around Mesh Shaders, and XSX also supports (hopefully by now) Primitive shaders
PS5: Primitive shaders and whatever customizations they put towards their GE
* they are not the same, but comparable.

From a hardware perspective, feature wise while they aren't the same at a technical level, strategically they have things in place to solve the same problems.
However from a software stack perspective, they are dramatically different. Xbox does everything through VM containers, and DX12 is likely significantly less efficient compared to Vulkan and GNM. If you are to compare GNM to DX12, you can see how much more register usage there is on DX12 over Vulkan for instance to do the same things on the GPU. That alone probably has a very significant impact on the performance of some games over others.
Direct X 12 U on the XSX has an ever lower level of access than Direct X 12 on the PC. You can't compare DX 12 U on PC with Vulkan and then put that onto the XSX.
 
Status
Not open for further replies.
Back
Top