Another warrior or what? I know that it will be embarrassing not for me, but for those who were crazy about RDNA 1 vs RDNA 2 and listed over and over again a set of console features, having no idea what they were writing about. It is annoying when people just fix themselves on some absolute nonsense, on some one thing or another, or are led by marketing and new loud-sounding features. And they just throw all the most important nuances into junk, for example, such as the SoC design, how it works, what is important in rendering and computing in general, what place does i/o now occupy in rendering, what bottlenecks it has. Spare me of your marketing noodles please.
You think both companies didn't think about this?
The question is not what they thought, but how they actually did it. The implementation of data management in their systems is quite different, for example. Whose approach will be better - we will see over time.
What those tweets go into is how only texturing active regions of memory saves you on memory bandwidth (it does) and storage (less resident data) and then it ends by saying people could already only upload select memory regions and not stream in the whole mipmap using PRT/Tiled Resources, but that SF (MS refers to as PRT+) and SFS extensions are the new hotness (which they are).
You really cannot praise the HW you like without dissing down the other? Just fishing for reactions?
.What should only matter is how both machines perform in the same games. And for now, PS5 has the advantage. And don't forget that PS5 is doing all that with a quite smaller APU. PS5 should be cheaper to manufacturate for those 2 reasons: smaller APU and cheaper cooling (PS5 has actually much less copper).
You really cannot praise the HW you like without dissing down the other? Just fishing for reactions?
Basic RDNA2 setup with cache scrubbers and modified Geometry Engine... oh yeah I/O unit slapped onto it
No. The cache scrubbers are not strictly for IO.
No, the cache scrubbers are implemented on the GPU. The coherency engines drive the process, cache modifications on GPU and CPU act upon the hints given to them:
No. The cache scrubbers are a general purpose GPU customization. They are a separate thing to the IO, although it can be exploited by the IO complex.
My take on why Cerny says primitive shaders - rather than talk about meshes - is because the two can be used interchangeably in a GDC audience situation, and from various comments(around Atomview technology in videos), I'm pretty sure the long term plan for PlayStation is to move beyond triangles and use new primitive forms the new GE advancements makes possible - and meshes by their very nature are old hat groups of triangles. The UE5 demo was used to push the message that with zero LoD and 4 triangles per pixel - a 4K reconstructed image from 1440p that was flawless - selling graphics on ever more polygons was over.
PS5 heatsink is bigger but actually lighter (from memory about 30% lighter, but can't find the article now) and XSX is more complex to assemble as the whole thing is in 2 parts because of the complex cooling part while PS5 is simply using a bigger motherboard in order to have an efficient (but rather cheap) cooling. PS5 uses liquid cooling but XSX uses a costly vapor chamber.
). The coherency engines drive coherency traffic and eviction intelligently thanks to HW customisations in each unit (cache scrubbers in the GPU).
The most expensive (and clever) part of liquid cooling design was actually 2 years of R&D. The liquid itself is not that expensive (negligible difference with traditionnal paste compared to the amount of copper needed).
The point is not absolute costs, but relative.
A Closer Look at How Xbox Series X|S Integrates Full AMD RDNA 2 Architecture - Xbox Wire
We here at Team Xbox would like to congratulate and celebrate our amazing partners at AMD on today’s announcement of the Radeon RX 6000 Series of RDNA 2 GPUs. It was incredible to see AMD demonstrate the power and potential that the new AMD RDNA 2 architecture can deliver to gamers around the...news.xbox.com
Xbox Series X|S are the only next-generation consoles with full hardware support for all the RDNA 2 capabilities AMD showcased today.
If anyone forgets ps5 does NOT support all of RDNA 2 hardware capabilities. PS5 has no VRS HW, mesh shading, SFS, etc. That is known and nothing to be disputed
That comment still stands, they are separate from the I/O complex (see picture with the freaking transcript
Both consoles had their own design goals and it is quite clear that XSX was a bit more bent towards a unified DirectX ecosystem (with a console flavour as it got to premiere features that are not going to be mainstream on PC for a while... see DirectStorage and the BCPACK decoder), but GPU wise they went for more raw power (a lot more CU's, heavier CU to Shader Array and Shader Engine ratios). On the I/O front they did go for a refined approach highlighting how SFS + DirectStorage + the SSD could work together to deliver a multiplier over bare HW specs.
That is a roundabout way to both admit the GPU has an additional elegant optimisation added on top of RDNA2 specs and at the same time discard it. "They are there to be used by the I/O IP"... and? That is a bit of a meaningless rebuttal sorry.
A well-explained post from
TheThreadsThatBindUs
talking about how more CUs/parallelism doesn't help that much with real-world performance when triangles/polygons are at the micro level.
You make a good point, but I wanna see how some Xbox fans, fanatics feel about that Twitter post.
I suppose they will not like it very much due to the last part but that was not my point as you surely know. I didn't like the posts contradictory narrative and the way he presents his subjective conclusions as facts with such authority. I think this leads to further misinformation.
i dont see the issue?
all i said earlier was ps5 seems less graceful since it seems to be base rdna2 + their i/o IP.
The cache scrubbers being located in the rnda2 gpu, you could say it is now more than 'base', but my point still stands, while they are logically placed there but they are there because of being part of the I/O IP.
no rdna2 game/base clock is closer to SX than ps5. look it up.
the materials to make liquid metal may not be more expensive, but the operating costs to inject and keep it safe in ps5, is logically more expensive, on top of the R&D you said.
Im still adamant that the $399 is an aggressive marketing push that is bleeding sony a bit more than they liked. something something about concerns with SS and they did not have a real cheaper sku to match.
Aggressive maybe - but highly unlikely to be competition related. They'd be working off of market data around 399/499 price points, and wanted to ramp up adoption fast.
Dude varying level of detail has nothing to do with VRS in that context... that's literally just about changing the geometry LOD levels when cerny spoke about the geometry engine. VRS isn't even in the geometry engine, it's in the ROPs... He was referring to geometry when he made those statements. VRS is about shading rate, not about geometry performance or manipulation.
Primitive Shaders are not Mesh Shaders because Mesh Shaders, unlike Primitive Shaders, does not need to rely on fixed function tesselation units. Primitive Shaders still relies on the hardware tesselation unit, Mesh Shaders does not need to and has the ability to replace the fixed function tesselator altogether using Amplification Shaders.
The PS5 I/O setup, while most impressive, is nothing close to being the exact same as Sampler Feedback Streaming. That only goes to show you don't understand what it actually is. Series X itself has a fast SSD, not as fast as PS5's on paper and it's own hardware decompression unit and advanced compression for textures, but Sampler Feedback Streaming is something else entirely after all those I mentioned already played their role, or better yet before some have done their job. PS5's I/O SSD setup is, according to Cerny, meant to move data into the RAM just in time as it's needed, and it can apparently do so really fast, but it's still doing it the more old fashion way, just faster.
Series X's Sampler Feedback Streaming is capable of a level of granularity that isn't supported on PS5, significantly cutting down on the amount of texture data that even needs to be inside VRAM or copied in the first place. Sampler Feedback Streaming is making sure texture data that was never supposed to even be in RAM in the first place for what's on screen, never goes there, leading to an effective RAM efficiency of up to 2.5x on Series X. What Sampler Feedback Streaming does is work to determine EXACTLY what's needed for the scene in such a precise, fast and accurate way that unnecessary data that would have never needed to have gone through decompression and thus end up in system RAM in the first place never actually ends up occupying system RAM. Thus your SSD transfer needs are significantly cutdown, your RAM usage is significantly cutdown, dramatically boosting effective usage. it's an absolute game changer that, believe it or not, goes much further than the PS5's SSD I/O setup.
PS5 is doing things the more old fashion way really, really fast. Series X's Sampler Feedback Streaming changes the very way texture data works for a GPU, and giving an effective RAM improvement that devs have wanted for years. Cache Scrubbers is a feature much less necessary for Series X with Sampler Feedback Streaming because often times it will do a better job of keeping useless texture data out of RAM and thus the GPU caches that shouldn't have been there in the first place if the streaming method was simply more intelligent, which is what Sampler Feedback is. It's one of the biggest and most slept on game changers on either console. Devs commonly ask for more RAM over anything else. This feature is essentially giving them that. It has the potential to do for Series X a version of what Infinity Cache does for PC RDNA 2 by significantly cutting down the bandwidth requirements due to less unnecessary reads/writes and smaller chunks of information.
You tried very hard with this one, but nearly everything you said is wrong.
And before someone goes saying it's just PRT. It's actually beyond that, though people do tend to interchange the terms.
What does any of that has to do with the price of rice?
No, you beat your chest after you deliver the goods... Some specs sheet is nice and all, but what counts is the output.
That is fanboy talk.
I kind of agree with you on that front.
Machine LearningYes, him, and I believe either he or another person also made a statement saying there was no Machine Learning on the PS5 GPU. That tweet I think is the follow-up to what he said prior.
found it.
Ehh - that would be implementation specific, it's not a given.
No - it happens at the rasterization stage. There's no two ways about it.
1.) The PS5 likely doesn't support the hardware based VRS solution which the Xbox Series and RDNA2 include.
You can precisely tell that the Floating-Point Register File was cut in half, that's enough to know that the PS5 can't execute FP256-Instructions as fast as vanilla Zen2 cores.
Pretty sure RDNA2 is a
Good job! GOT EMMMMMMMMMM!1.) The PS5 likely doesn't support the hardware based VRS solution which the Xbox Series and RDNA2 include.
It's possible to implement similar techhniques in software, even quite efficient or with advantages but it's not universally better and is already achieved on the last gen consoles.
Further one can use and combine both solutions, the Xbox Series has this option the PS5 likely hasn't.
How large is the advantage? Nobody knows without benchmarks, so claiming it's in general much better, the same or worse is pure fantasy without good data and arguments backing it up.
2.) I said it already on this forum but you really need to know the specifications and features from "Primitive Shaders" on the PS5 to know what's the difference, if there is any, in comparison to "Mesh Shaders".
Those terms are just arbitrary names otherwise.
Personally I don't think there is a significant difference.
3.) You can support Machine Learning on nearly every advanced processor but you can't train or execute ML networks fast enough for real time rendering on every hardware.
There are many ML models which don't need high precision mathematics, so one can use FP16, INT8, INT4 or even INT1.
What's important for real time rendering is that the throughput is high enough.
It wouldn't be surprising to me if the PS5 only has packed math for FP16 and INT16, that's quite a bite worse than the mixed precision dot-product instructions which the Xbox Series and RDNA2 GPUs can offer.
They can multiply 2xFP16, 4xINT8 or 8xINT4 data elements and add the results to a FP32 or INT32 accumulator in one execution step.
This is more precise and/or a lot faster than just working with single precision and packed math operations.
You can precisely tell that the Floating-Point Register File was cut in half, that's enough to know that the PS5 can't execute FP256-Instructions as fast as vanilla Zen2 cores.
What's harder to tell or impossible is, if and in what way the digital logic on the execution paths was cut.
I think Nemez's fifth tweet is talking about theoretical TFLOPs and that in real world terms more factors are important, putting both closer together in practise.
The raster pipeline on the PS5 is from a high level view structured as on RDNA1 GPUs.
This may be even a performance advantage, since AMD's own RDNA2 GPUs rebalanced execution resources and made some cut downs.
They have less Primitive Units and Depth ROPs per Shader Engine.
From a Compute Unit perspective, for most operations, there isn't a real difference between RDNA1 or RDNA2.
Smart Shift is basically "firmware magic".
Since multiple generations AMD has the necessary hardware built in to control and set voltage levels, clocks, thermal limits, etc.
It's all very programmable and can be set as desired.
Sony went for a clock budget, which is also shared between CPU and GPU, based on activity counters and MS wanted fixed clock rates.
AFAIK the way Sony is doing it, is very similar to how AMD first implemented variable frequencies under Cayman (6970/50, 2010/11 GPUs), where AMD opted for no clock differences between the same GPU SKUs.
That's different now, AMD is using real time measurements and let's each chip optimatly behave under his unique attributes.
Every chip is different in quality, they all behave a bit differently, something you certainly don't want for a gaming console which is why every PS5 is modeled after a reference behaviour based on activity counters.
That's a high level diagram from GCN2, showing how the power/thermal/clock control is laid down, with programmable software layers.
It's not a gotcha posting, just my commentary you may agree or disagree with.
This is how I got something described - please note that I am an interested layman here and not an expert - nor do I sit on insider knowledge but I have had talks with someone that works in the PS5 development environment. The functionalities of the GE in PS5 (what is transistor, i.e. hardware driven, and what is software driven is unknown to me) allows for culling as well as merging of geometries. These two function as such create a similar functionality as mesh shaders in terms of culling but it happens earlier in the pipeline as well as VRS by allowing for control of the number of pixels that enters the final shader step through merging of geometries (i.e. the same functionality that VRS plays in Nvidia and current AMD cards).
