Support NeoGAF

[BluRayHiDef]

Not only that. First of all, you cannot measure render buffer output with input data. An in-engine scene can use multiple textures, one texture for multiple objects but also none (objects with shaders only). Then you have game logic, player input data, many other factors. In other words: you can't tell if the I/O throughput is enough in theoretical terms. This can be done only on case-to-case basis, in particular applications of a given game in development.

Well, I'm not sure whether or not @Ascend was aware that the duration of render buffer output cannot be measured with input data, because his post implied that he thought so. However, I knew that was the case, which is why I calculated only the amount of time that each console's GPU would have to read the input data and render it.

 

[farmerboy]

Reading the back & forth about the SSD...

Yep, thats me.

Is that SSD nonsense is over with?

Great. So what do you guys think will be the first game Microsoft shows and the first game that Sony will show?

Bless, something to talk about. I'm gonna go with Halo Infinite the half billion dollar extravaganza and GOW Ragnarok.

And. I. Can. Not. Wait.

 

[ksdixon]

This looks great to me. I always liked the PS4 devkit, and it's even nicer in this color combo withthe blue lights on it.

 

[bitbydeath]

Dude! You're seriously grasping at straws. You've been saying a lot of half truths.

1.) Yes the PS5 eliminates bottlenecks to increase the SSD's throughput. So does the XSX and we now know it has a DMA controller as well that they haven't discussed yet.

2.) The article qualifies and says Microsoft is aiming to eliminate load times.

The PS5 SSD being twice as fast doesn't mean the XSX won't be able to eliminate load times. Both systems can push above 4GB/s into RAM. They can both eliminate load times.

It's impressive what Sony did with the PS5. But everything indicates that the XSX will be able to eliminate load times once the Velocity architecture is done.

Who's grasping? I don't even know where to begin with this post. If that's what you believe then good luck!

 

[bitbydeath]

Theres no such thing as "completely eliminating" load times, you know that right?

We'll have to see as that is the claims Mark Cerny has put forward.

 

[Games Dean]

his sources fucked him over. he shared beers with those guys.

Is that honestly what you think happened?

 

[FeiRR]

But on the other hand. As big screw up as Xbox One was from the beginning, PS4 only managed to barely outsold it in UK/US

This is not "barely", This is over 300% more for Playstation with XGP in full swing and Microsoft's ad campaign of promoting their next gen hardware, good press and support of social media warriors who are very active right now. All this for nothing. Xbox is a marginal brand. Even Switch with limited supplies sells more, even when people stay at home, even though its hardware is so much weaker.

so I think, that Microsoft have a very big chance to outsold it in that markets and that is good start. Of course price will be most important thing but with Lockhart, Microsoft will cover both performance and low price aspect of next-gen.

As I wrote in my previous post, we have exactly the same situation right now, with results as above.

If PS5 and XsX will be 499$ and Lockhart will be 299$ i can see Microsoft outselling PS5 in US and UK and making a dent to Sony stronghold in other European countries. So...1,4:1 for PS5 in terms of worldwide sales. Japan will take care of that.

You have no proof, just dreams.

 

[ethomaz]

So scales with TF. Assumed as much, thanks.

Exactly.

The only units that doesn't escale with TFs are the ones outside the CUs.

 

[John254]

This is not "barely", This is over 300% more for Playstation with XGP in full swing and Microsoft's ad campaign of promoting their next gen hardware, good press and support of social media warriors who are very active right now. All this for nothing. Xbox is a marginal brand. Even Switch with limited supplies sells more, even when people stay at home, even though its hardware is so much weaker.


As I wrote in my previous post, we have exactly the same situation right now, with results as above.


You have no proof, just dreams.

I guess you have a proof how consoles will sell

I'm sure you also created sales analysis at the end of WiiU lifecycle in terms of how Switch will sell, because it is that easy. Just take sales from one generation and paste it to another and voila.

I never said i have proof, as you also don't have any proof. It's just my observation and guess.

And I really don't know what "ad campaign of promoting next-gen hardware" have with sales of current gen. At best it can suppress them, because many people are just waiting for next-gen.
Also...since when we are taking VGchartz as relevant source?

 

[ethomaz]

RT performance scales with CUs not clockspeed. That's according to multiple sources including Digital Foundry.

Wrong again.
It scale with both.

 

[Deleted member 775630]

This is not "barely", This is over 300% more for Playstation with XGP in full swing and Microsoft's ad campaign of promoting their next gen hardware, good press and support of social media warriors who are very active right now. All this for nothing. Xbox is a marginal brand. Even Switch with limited supplies sells more, even when people stay at home, even though its hardware is so much weaker.

Although I agree that PS will definitely outsold Xbox, VGChartz is not a credible source. Better to wait for NPD, if that still matters so late in this gen.

 

[FeiRR]

Although I agree that PS will definitely outsold Xbox, VGChartz is not a credible source. Better to wait for NPD, if that still matters so late in this gen.

Then you'll say that NPD data doesn't matter, you're already implying that. Like money they get right now is less important. Actually it the best time to compare sales, as console production costs are the lowest so the profit margin for the company is the highest and it also helps convince shareholders that investment in the next generation (which is a huge loss in a yearly report) is a good road to take.

 

[xool]

What type of decompression is it doing? I'm assuming some form of Zlib if it is attached to the IO for the SSD. BCpack is a form of compression, but as I already mention earlier, its main aim is to allow higher resolution textures (with loss errors) to reside in place of a loseless texure 1/4 of the size. A decompression of lossy textures at the IO doesn't make any sense (IMHO) if that's what they were referring to.

"Our second component is a high-speed hardware decompression block that can deliver over 6GB/s," reveals Andrew Goossen. "This is a dedicated silicon block that offloads decompression work from the CPU and is matched to the SSD so that decompression is never a bottleneck. The decompression hardware supports Zlib for general data and a new compression [system] called BCPack that is tailored to the GPU textures that typically comprise the vast majority of a game's package size."​

 

[xool]

Some new renders I saw today! I'd bite

Never thought I'd see a controller that was uglier than a dev kit but here we are. It's been a few days and my opinion hasn't softened - that this is atrocious.

Also now I've seen the "white girl in black string top" thing I can't unsee it. Not good.

 

[xool]

"Some Devs Say PS5 Is Superior To Xbox Series X In Several Ways & Praise "Revolutionary" Hardware"​

That title reads like something The Onion would write

 

[pasterpl]

re. the console sales;

Xbox One lifetime unit sales worldwide 2024 | Statista

As of August 2024, lifetime unit sales of Xbox One consoles in North America reached nearly 33 million, while in Europe, lifetime unit sales surpassed 12.8 million.

www.statista.com

Global sales of Xbox One consoles reached over 51.51 million units in February 2020.

For PS4

The latest figures show that sales of Sony's newest console have reached over 107.57 million units.

Based on this and various other sources it is fair to estimate that PS4 sold 2x better than XBOne.

but it is interesting to see and compare these companies services;

• XGP was rumoured to have 9.5M subscribers in May last year and then in Nov. 2019 MS CEO said that the numbers doubled in the Q4 2019, so we can estimate subscribers at 20M
• XBL had a couple of months with over 65M Monthly Active Users in Q4 as well
• PS Now had 1M subscribers in late 2019
• PS Plus had 37M subscribers in late 2019

I think this is very relevant in regards to what pricing both companies can go with. And the xCloud is not yet released

 

[xool]

Lockhart has been revealed

Thank Me Later

 

[pasterpl]

Lockhart has been revealed

Thank Me Later

thats actually pretty cool looking (got GameCube vibes)

 

[xacto]

Please for the love of god they should now show actual real gameplay of next-gen only games. Show us what these consoles are made of.....

At this point I don't even need games, we all know they're going to look awesome on BOTH consoles... right now I'd really like a look at the PS5 box and maybe some OS features? Like... "we're done with all those updates that improve stability and actually they do nothing, no more HDD locking up while opening friends' list, or sluggishly switching between apps on PS4, gone are those 'copying patch'"... that sort of thing?

How about that, Sony? Is it too much to ask for at least a countdown to reveal?

 

[Fdkenzo]

Lockhart has been revealed

Thank Me Later

No disc drive?

 

[llien]

If there are no IO bottlenecks and ram the same

Same same, but different. XSeX has 10GB of faster RAM (which GPU will likely use)

The 6GB/s is a theoretical max output from XSX hardware decompressing block just like 22GB/s is a theoretical max throughput of PS5s hardware decompression unit

I need to see the receipts of MS stating that is chip's max and not expected decompression.

I also find it curious that expectations are something will be cut in half, when compressed.

 

[TBiddy]

It's a fake render. Well made and looks really cute, but fake, unfortunately.

 

[xacto]

And? PS3 is not in the same position as PS4. Last gen was a hole different story. They can't hype PS5 until PS4 software still in place. When I say 'software' I mean games that have potential to push hardware.

While I do agree with you, in my particular case (and I have a feeling there are some others out there just like me) it's like this: whatever games Sony will release this year for PS4, even though I will want to play them like mad, I will not buy them anymore. I will get them directly for PS5, when it comes out.

Why, you say? For one, I have a huge backlog of PS4 games already; then, even if 1st party games will look awesome on PS4, I am quite sick of long loading times, HDD struggling to keep up the pace, frame rate drops just to name a few. Because let's be honest they do happen, even if best games of a generation come out right at the end of the said generation.

So right now, I am way more hyped by PS5 than Ghost of Tsushima or The Last of Us 2, really. While they will look awesome, there will be time to play those in a much better way on the PS5. That is why I think that not hyping PS5 just because they have PS4 games still left to launch is not exactly a good policy at Sony.

But as I've said, that's just the way I see things, probably execs at Sony might think differently. But if they think that people aren't a bit more interested in a new console than another game for PS4, they might be a little surprised...

Just give everyone what they're interested in, end-generation awesome games, and a new generation console for people looking forward to it.

 

[VanSiriusGabriel]

Not sure if you are joking but i don't see what's funny about this. Why mock him for correcting another member?

It's the way he did it.
And that's really it.

 

[Agnostic2020]

Lockhart has been revealed

Thank Me Later

Why would there be an eject button when there is no disc tray ? Haha fake

 

[John254]

Why would there be an eject button when there is no disc tray ? Haha fake

I think that button is supposed to be for controller sync, not eject button (but i agree that it is fake)

 

[xacto]

You're only looking at the SSD and excluding all the other bottlenecks removals Sony performed to achieve 0 load screens.

"As game creators, we go from trying to distract the player from how long fast travel is taking, like those Spider Man subway rides, to being so blindingly fast that we might even have to slow that transition down." - Mark Cerny, The Road to PS5.

Also Cerny stated that they had to address "a lot of other bottlenecks" and they did just that.

 

[xool]

Did anyone ever get a breakdown of the "total 76MB cache" (DF : "76MB of SRAM") in the Series X

From what I can read:
Zen2 has

• 64kB L1 x8 = 0.5MB
• 512kB L2 x8 = 4MB
• 32MB L3
• Total 36.5MB

RDNA has

• 32kB L0 instruction cache shared between dual CUs = 32*52/2 = 832kB (~0.8MB)
• 16kB L0 scalar cache shared between dual CUs = 16*52/2 = 416kB (~0.4MB)
• 16kB L0 vector cache - two in a dual CU (i think) = 16*52 = 832kB (~0.8MB)
• 128kB L1 graphics cache share between dual CUs = 128*52/2 = 3328kB (=3.25MB)
• sub-Total 5408kB (~5.28MB)
• Global L2 cache 4MB
• Grand Total ~9.28MB

APU total would be 45.28MB .. way short. (there could be some extra cache for audio engine etc but not 30MB lol ..)

Does RDNA2 have more cache? What did I miss ??

(I hope that 76MB figure isn't another BS exaggeration to waste my time on...)

 

[chilichote]

I got the 7-9 GB/s from Cerny's presentation road to PS5. Maybe I misheard. The 4.8 -6Gb/s was from MSFT. And it is Sony that mentioned the theoretical high of 22GB/s for specific data. We have heard no such thing from MSFT.

Okay so thank you for the clarification but the point still stands. Because I was using the 9GB/s rate not the lower one. Otherwise the figures to compare are:
PS5:
8-9 GB/s
XSX:
4.8-6GB/s before XBTC is finalized.

It is 4.8GB/s, up to 6GB/s under certain conditions, at Microsoft and 8 - 9GB/s, up to 22GB/s under certain conditions, at Sony. Just to restore reality.

Edit: sorry for being late^^

 

[TBiddy]

Does RDNA2 have more cache? What did I miss ??

(I hope that 76MB figure isn't another BS exaggeration to waste my time on...)

The return of the 32 MB ESRAM of course. This time running at 560 GB/s, essentially bringing the total bandwidth in the XSX up to 1,120 GB/sec.

 

[Felessan]

Do we have confirmation that BCPack is lossy algorithm? Because I find it hardly believable how it will maintain data integrity of already compressed textures, without introducing some nasty artifacts. Also 50+% for additional compression of BCn encoded textures are hard to believe, as compressed textures already heavily optimized to save space.
And if it's the whole pack (texture compression for GPU purposes and some extra to save IO bandwidth), it's hardly comparable to kraken alone, as kraken is only half of compression, and PS5 will have it's own texture-level compression in place.
I will wait and see some more details of the secret BCPack sauce.

Data Transfer Rate of PS5's SSD: 5.5GB/s = 5500 MB/s = 5500MB/1000ms = 5.5MB/ms
Size of 4K Frame with 32Bit Color Depth = 33.2MB

There are a lot of wrongs here.
First - you don't read than render. If you need to read, you read data in advance.
Second - typical frame will use much more data, as many different "textures" applied to the single surface.
Third - textures already heavily compressed, and operated as such, final decompression done in the GPU itself, so you need to adjust for texture compression rate in your calculation.

 

[xacto]

Sony's gpu slower, CPU slower, Bandwidth slower (Gpu 560GB/s to 443GB/s), bit rate slower(320bits to 256bits), Retrace slower(52 to36cu), compression of data slower (BCPack),
Teraflops slower(12. to 10 ish), and you think a SSD will over come all of that?

Just say no to drugs.

We have seen this before, o yes , Sony is using the MS talking points. "PS4 is only .4 or .5 more than us, don't worry you won't notice a thing" said 900p.
Ms "We have things in are system that we can't even talk about yet" said 1-4 years later.
Ms "we have the CPU to push the GPU" said the same CPU as PS4.

You guys of PS5 power are in for a long gen. With good games at least.

Power doesn't sell a generation.

In Microsoft's case, actually power never did anything, especially sell its console. So thinking that now that logic will apply... well, it could, but history tells otherwise.

 

[BluRayHiDef]

Interesting question. But games (and thus consoles) don't work on the 'second' level, but on the millisecond level. What this means?

Let's assume the games will be 4K 60 FPS. 60 FPS means you have to have a new 4K frame at the latest every 16.7 ms. How large is a 4K frame? If we assume 32 bit color depth, a non-compressed 4K frame would be around 33.2 MB.

So by that metric, the transfer rate of the PS5's SSD amounts to 91.7MB per frame (raw), or 133.3MB - 150 MB per frame (compressed).
What about the XSX? 2.4 GB/s raw would be 40 MB per frame (raw). As for the compressed value, ignoring BCPack and using only the given compressed number by MS, it would net you 80 MB per frame.

So what do we get from this? The raw SSD transfer speed of the XSX is already enough to transfer a full uncompressed 4K image from the SSD to your display in less than 16.7ms, allowing you to have 60 fps. Add compression, and you can transfer two frames per 16.7 ms, which basically means it is fast enough to transfer a 4K image directly from your SSD to your display at 120 FPS. The PS5 is beyond that.
But what does this all mean? It means that the SSD of the XSX is theoretically fast enough to do what Sony is advertising, which is, not requiring loading times. The speed of the SSD in the XSX allows faster loading than the max time that a frame has available for construction.

Another simple way of looking at it is, if a 4K frame is 33.2 MB (raw), and you need 60 every second, you need a transfer of 33.2 x 60 = 1992 MB = 1.992GB every second. The XSX has 2.4 GB. Note that this is for formed frames only though. You'll need to load multiple textures and other assets to form the full image. And for the Xbox, that's where SFS comes into play.

----------Data Transfer Rates--------- (More data per unit time is better)​

PlayStation 5​

Uncompressed Data Transfer Rate of PS5's SSD: 5.5GB/s = 5500 MB/s = 5500MB/1000ms = 5.5MB/ms

Compressed Data Transfer Rate of PS5's SSD: 8GB/s = 8000 MB/s = 8000MB/1000ms = 8MB/ms

Compressed Data Transfer Rate of PS5's SSD: 9GB/s = 9000 MB/s = 9000MB/1000ms = 9MB/ms

Xbox Series X​

Uncompressed: Data Transfer Rate of XSX's SSD: 2.4GB/s = 2400MB/s = 2400MB/1000ms = 2.4MB/ms

Compressed: Data Transfer Rate of XSX's SSD: 4.8GB/s = 4800MB/s = 4800MB/1000ms = 4.8MB/ms

----------Time that it takes each console's I/O system to transfer a full 4K frame's worth of uncompressed data to RAM---------- (Less time is better)​

Size of 4K Frame with 32Bit Color Depth = 33.2MB

PlayStation 5​

@5.5GB/s (uncompressed): 33.2MB/(5.5MB/ms) = 6.036ms

@8GB/s (compressed): 33.2MB/(8MB/ms) = 4.15ms

@9GB/s (compressed): 33.2MB/(9MB/ms) = 3.69ms

Xbox Series X​

@2.4GB/s (uncompressed): 33.2MB/(2.4MB/ms) = 13.83ms

@4.8GB/s (compressed): 33.2MB/(4.8MB/ms) = 6.92ms

----------Time that each console's GPU has to read the data from RAM and render it while maintaining a render rate of 60 frames per second---------- (More time is better, because it's more time to get the work done)​

1 second = 1000ms
1000ms/60 frames = 16.67ms

PlayStation 5​

16.67ms - 6.036ms = 10.634ms (for uncompressed data)

16.67ms - 4.15ms = 12.52ms (for compressed data)

16.67ms - 3.69ms = 12.98ms (for compressed data)

Xbox Series X​

16.67ms - 13.83ms = 2.84ms (for uncompressed data)

16.67ms - 6.92ms = 9.75ms (for compressed data)

----------Number of pixels that each console's GPU can render in the amount of time that it has to read the data from RAM and render it while maintaining 60 frames per second---------- (More pixels are better)​

PlayStation 5​

64 ROPs x 2,230 x 1000 = 142,720,000 pixels per second = 142,720 pixels/ms

(142,720 pixels/ms) x 10.634ms = 1,517,684.48 pixels (after processing uncompressed data at 5.5GB/s)

(142,720 pixels/ms) x 12.52ms = 1,786,854.4 pixels (after processing compressed data at 8GB/s)

(142,720 pixels/ms) x 12.98ms = 1,852,505.6 pixels (after processing compressed data at 9GB/s)

Xbox Series X​

64 ROPs x 1,825 x 1000 = 116,800,000 pixels per second = 116,800 pixels/ms

(116,800 pixels/ms) x 2.84ms = 331,712 pixels (after processing uncompressed data at 2.4GB/s)

(116,800 pixels/ms) x 9.75ms = 1,138,800 pixels (after processing compressed data at 4.8GB/s)

----------Number of texels that each console's GPU can render in the amount of time that it has to read the data from RAM and render it while maintaining 60 frames per second---------- (More texels are better)​

PlayStation 5​

144 TMUs x 2,230 x 1000 = 321,120,000 texels per second = 321,120 texels/ms

(321,120 texels/ms) x 10.634ms = 3,414,790.08 texels (after processing uncompressed data at 5.5GB/s)

(321,120 texels/ms) x 12.52ms = 4,020,422.4 texels (after processing compressed data at 8GB/s)

(321,120 texels/ms) x 12.98ms = 4,168,137.6 texels (after processing compressed data at 9GB/s)

Xbox Series X​

208 TMUs x 1,825 x 1000 = 379,600,000 texels per second = 379,600 texels/ms

(379,600 texels/ms) x 2.84ms = 1,078,064 texels (after processing uncompressed data at 2.4GB/s)

(379,600 texels/ms) x 9.75ms = 3,701,100 texels (after processing compressed data at 4.8GB/s)

 

[xool]

Do we have confirmation that BCPack is lossy algorithm?

afaik - no

Also 50+% for additional compression of BCn encoded textures are hard to believe, as compressed textures already heavily optimized to save space.

They're not really "heavily optimized" at all - BCn is a very simple data reorder and bit length truncation that is mostly designed to be really easy and fast to get the data back out.

Like you said - decompression takes place on GPU - so it's likely that they still used the old BCn format - thus BCPack is just compression for BCn - so it's either lossless, or designed so that errors introduced are not cumulatively doubly bad..

..but here's the thing - if tweek the BCn data the right way (or more speficially tweek the input bitmap), then encode with lossy compressor the resultant unpacked data can be closer (less noise error) to the original if you tweek right .. it's an iterative process - chose the wrong input data and the output is worse, choose well and the output error is less ..

..did that make sense ?

And if it's the whole pack (texture compression for GPU purposes and some extra to save IO bandwidth), it's hardly comparable to kraken alone, as kraken is only half of compression, and PS5 will have it's own texture-level compression in place.
I will wait and see some more details of the secret BCPack sauce.

BCPack has been stated to be specifically for textures. Series X compression unit supports zlib in hardware. (Kraken is better version of zlib)

 

[Ass of Can Whooping]

No... His exact words were "the unit itself is capable of outputting as much as 22 GB a second, if the data happened to compress particularly well"

So, the hardware itself is capable of up to 22GB/s

That's the limit

 

[ZywyPL]

Just caught up with the past 50 pages of this thread, this is basically all I got:

Seriously, who's paying you guys, I wanna jump on that cash train as well, sign me in please!

This thread has taken a turn for the worse. A long time ago to be honest.

Let's just put 'er down like Old Yeller and wash our hands of it?

The thread started horribly to begin with to be honest - 16C CPU, 32GB HMB3, 20TF GPU, 4TB SSD, and so on, the thread was basically a clownfiesta from the get-go and never held back. And now you have the same delusional kids/clowns spamming their baseless theories, trying to convince everyone that black is white and white is black.

Both camps fail miserably with their arguments before they even start, firmly believing that ALL the RAM will be flushed in and out 24/7, every single second, while that's not going to happen, not even close. Some part of those ~13GB will be exchanged back and forth, sure, but that will still hugely depend on the game type, just as it is now, while some, most, if not even all part of it will be static data and kept in RAM the whole time.

Take Fifa/PES/Madden/MLB/NBA for example - all the players are there on the screen, the playfield is there the whole time, the whole stadium/tribune, the whole audience, what data is there than can be streamed in and out? None. Same deal for every fighting game. There will be many many games, if not entire genres, where in order to play them they will simply need to be fully loaded in into RAM first, just as it is right now, for some titles those 13GB will be even more than they'll actually ever need to begin with, yet no one is even taking it into account (or don't want to)... Racing games like GT/FM/PC/AC, another example where you have all the cars on the track, the track itself might be streamable to some degree, but I personally don't think it will be the case, just because of replays where you can instantly switch between the cars and the part they are current at (which can be miles apart) needs to be there as well, instantly available after every press of a button, a.k.a. in RAM.

Streaming from the drive, also just as it takes place now, will mostly benefit open-world games even further, that's for sure, and maybe some tight corridor games, that have limited field of view and there's really no need to store assets that won't be reachable/visible in next couple of seconds/minutes no matter what, and/or there's no backtracking so there's no reason to keep assets of already finished parts of the level.

"Can" is rather different from "Will".

This. That's exactly why I take Cerny's whole presentation with a full truck of salt - "might", "should", "we expect", the lack of certainty was astonishing, it left the taste than PS5 is still in early-mid development stage, rather than a product that's about to launch in half a year... They should've let Jim Ryan do the job, he's a much, much better spokesman than Cerny ever will.

 

[xool]

This thread on 22GB/s

At most it's one person who won't accept reality, doesn't understand the meaning of "typical". Not sure they're even here anymore and people are explaining to others who already agree with them.

Move on. One idiot. 10 pages. Please.

 

[Thirty7ven]

Same same, but different. XSeX has 10GB of faster RAM (which GPU will likely use)

Without extra bandwidth the more powerful gpu would be bottlenecked so the only thing that extra bandwidth is doing is allowing the devs to leverage that extra power.

Maybe on your mind these things stack on each other, but the reality is that they need each other.

 

[xool]

Take Fifa/PES/Madden/MLB/NBA for example - all the players are there on the screen, the playfield is there the whole time, the whole stadium/tribune, the whole audience, what data is there than can be streamed in and out?

Well .. one idea was that a more detailed audience could be streamed in as the camera pans/rotates along the pitch .. because .. that's literally what Cerny told people in the talk .. (can find the timestamp if you need it)..

Don't shoot the messenger -- I stated that all this will achieve is to replace "forced walking" and "elevator chat" with "slow turning" in next gen games .. about 100 pages back ..

 

[llien]

Without extra bandwidth the more powerful gpu would be bottlenecked so the only thing that extra bandwidth is doing is allowing the devs to leverage that extra power.

In other words, more CUs bring you nothing, if you are on the same BW RAM.

That one bold and unsubstantiated claim, one of the many in this thread.

 

[BluRayHiDef]

There are a lot of wrongs here.
First - you don't read than render. If you need to read, you read data in advance.
Second - typical frame will use much more data, as many different "textures" applied to the single surface.
Third - textures already heavily compressed, and operated as such, final decompression done in the GPU itself, so you need to adjust for texture compression rate in your calculation.

I was simply applying the logic of the person to whom I was responding. Also, I am sure that other factors affect rendering, such as the number of elements that have to be procedurally generated rather than streamed from storage; however, such elements are variable and cannot be taken into account in these theoretical calculations.

 

[PaintTinJr]

"Our second component is a high-speed hardware decompression block that can deliver over 6GB/s," reveals Andrew Goossen. "This is a dedicated silicon block that offloads decompression work from the CPU and is matched to the SSD so that decompression is never a bottleneck. The decompression hardware supports Zlib for general data and a new compression [system] called BCPack that is tailored to the GPU textures that typically comprise the vast majority of a game's package size."​

Yeah, after seeing a post by ethomaz, where I previously thought Bcpack was part of Direct3D's version of DXT compression - because of them talking about lossy textures and block compression - it now appears to be a variant of Zlib like kraken is, but if using lossy encoding (maybe reusing an endpoint from a previous block), hence the comparison tweet that Xbox is doing 50% compression on DXT textures (vs 20-30% using Kraken algorithm).

That is looking like how they go from 2.4GB/s to 4.8GB/s because 50% is a 2 -1 ratio. The 6GB/s claim surely comes from them hoping to refine the algorithm to get compression to 75%. Which is *maybe* possible, depending three things: a) What is the IO hardware block compute and bandwidth limit? b) Within those limits, is that enough compute to realistically get Bcpack compression an extra 25%? c) can texture quality endure any more loss to be more aggressive with the algorithm? eg store every other index and statistically extrapolate back the missing ones. The technique being NDA'ed rather than patented, suggests it can be applied on other systems, and with the IO complex having two co-processors, it can probably be accelerated on the PS5 (if any good).

 

[Thirty7ven]

In other words, more CUs bring you nothing, if you are on the same BW RAM.

That one bold and unsubstantiated claim, one of the many in this thread.

I'm sorry you don't understand how these things work. Or maybe you understand it better than everybody and MS, Sony or Nintendo needs you on the phone ASAP so that they come out with a console with 80 CU and 16 GB of DDr3, and the slowest SSD on the market. Those teraflops work alone baby, unshackle them!

 

[Xyphie]

Did anyone ever get a breakdown of the "total 76MB cache" (DF : "76MB of SRAM") in the Series X

From what I can read:
Zen2 has

• 64kB L1 x8 = 0.5MB
• 512kB L2 x8 = 4MB
• 32MB L3
• Total 36.5MB

RDNA has

• 32kB L0 instruction cache shared between dual CUs = 32*52/2 = 832kB (~0.8MB)
• 16kB L0 scalar cache shared between dual CUs = 16*52/2 = 416kB (~0.4MB)
• 16kB L0 vector cache - two in a dual CU (i think) = 16*52 = 832kB (~0.8MB)
• 128kB L1 graphics cache share between dual CUs = 128*52/2 = 3328kB (=3.25MB)
• sub-Total 5408kB (~5.28MB)
• Global L2 cache 4MB
• Grand Total ~9.28MB

APU total would be 45.28MB .. way short. (there could be some extra cache for audio engine etc but not 30MB lol ..)

Does RDNA2 have more cache? What did I miss ??

(I hope that 76MB figure isn't another BS exaggeration to waste my time on...)

CPU will likely only have 2x4MB L3 just like Renoir / PS5 has. There's a bunch of stuff you're not including, a bunch of uncore stuff will have caches and all of the registers etc in the CPU/GPU will be made out of SRAM and contribute to the 76MB total.

 

[Kusarigama]

----------Data Transfer Rates---------

PlayStation 5​

Uncompressed Data Transfer Rate of PS5's SSD: 5.5GB/s = 5500 MB/s = 5500MB/1000ms = 5.5MB/ms

Compressed Data Transfer Rate of PS5's SSD: 8GB/s = 8000 MB/s = 8000MB/1000ms = 8MB/ms

Compressed Data Transfer Rate of PS5's SSD: 9GB/s = 9000 MB/s = 9000MB/1000ms = 9MB/ms

Xbox Series X​

Uncompressed: Data Transfer Rate of XSX's SSD: 2.4GB/s = 2400MB/s = 2400MB/1000ms = 2.4MB/ms

Compressed: Data Transfer Rate of XSX's SSD: 4.8GB/s = 4800MB/s = 4800MB/1000ms = 4.8MB/ms

Size of 4K Frame with 32Bit Color Depth = 33.2MB

----------Time that it takes each console's I/O system to transfer a full 4K frame's worth of uncompressed data to RAM----------

PlayStation 5​

@5.5GB/s (uncompressed): 33.2MB/(5.5MB/ms) = 6.036ms

@8GB/s (compressed): 33.2MB/(8MB/ms) = 4.15ms

@9GB/s (compressed): 33.2MB/(9MB/ms) = 3.69ms

Xbox Series X​

@2.4GB/s (uncompressed): 33.2MB/(2.4MB/ms) = 13.83ms

@4.8GB/s (compressed): 33.2MB/(4.8MB/ms) = 6.92ms

----------Time that each console's GPU has to read the data from RAM and render it while maintaining a render rate of 60 frames per second----------​

1 second = 1000ms
1000ms/60 frames = 16.67ms

PlayStation 5​

16.67ms - 6.036ms = 10.634ms (for uncompressed data)

16.67ms - 4.15ms = 12.52ms (for compressed data)

16.67ms - 3.69ms = 12.98ms (for compressed data)

Xbox Series X​

16.67ms - 13.83ms = 2.84ms (for uncompressed data)

16.67ms - 6.92ms = 9.75ms (for compressed data)

----------Number of pixels that the each console's GPU can render in the amount of time that it has to read the data from RAM and render it while maintaining 60 frames per second----------

PlayStation 5​

64 ROPs x 2,230 x 1000 = 142,720,000 pixels per second = 142,720 pixels/ms

(142,720 pixels/ms) x 10.634ms = 1,517,684.48 pixels (after processing uncompressed data at 5.5GB/s)

(142,720 pixels/ms) x 12.52ms = 1,786,854.4 pixels (after processing compressed data at 8GB/s)

(142,720 pixels/ms) x 12.98ms = 1,852,505.6 pixels (after processing compressed data at 9GB/s)

Xbox Series X​

64 ROPs x 1,825 x 1000 = 116,800,000 pixels per second = 116,800 pixels/ms

(116,800 pixels/ms) x 2.84ms = 331,712 pixels (after processing uncompressed data at 2.4GB/s)

(116,800 pixels/ms) x 9.75ms = 1,138,800 pixels (after processing compressed data at 4.8GB/s)

----------Number of texels that each console's GPU can render in the amount of time that it has to read the data from RAM and render it while maintaining 60 frames per second----------

PlayStation 5​

144 TMUs x 2,230 x 1000 = 321,120,000 texels per second = 321,120 texels/ms

(321,120 texels/ms) x 10.634ms = 3,414,790.08 texels (after processing uncompressed data at 5.5GB/s)

(321,120 texels/ms) x 12.52ms = 4,020,422.4 texels (after processing compressed data at 8GB/s)

(321,120 texels/ms) x 12.98ms = 4,168,137.6 texels (after processing compressed data at 9GB/s)

Xbox Series X​

208 TMUs x 1,825 x 1000 = 379,600,000 texels per second = 379,600 texels/ms

(379,600 texels/ms) x 2.84ms = 1,078,064 texels (after processing uncompressed data at 2.4GB/s)

(379,600 texels/ms) x 9.75ms = 3,701,100 texels (after processing compressed data at 4.8GB/s)

Thanks a lot for these calculations. Can you or anybody else put where higher numbers are better and where lower numbers are better for the less technical minded folks

 

[llien]

I'm sorry you don't understand how these things work

That's ok, at least we have you, who understands.

So can I take, say, 5700, cut it's number of CUs by 33% and get the same perf, as, well, RAM bandwidth isn't changing.

I'd say, it's a revolutionary concept! Kudos!

 

[lepermessiah]

PowAh PoWAh PowAh SloWeR SloWeR

BENCHMARKS BENCHMARKS BENCHMARKS

PapEr SpEcs PaPeR SpEcS PaPer SpEcS

EFFECTIVE REAL WORLD PERFORMANCE

Hand me the nail gun this is getting maddening.

 

[Thirty7ven]

That's ok, at least we have you, who understands.

So can I take, say, 5700, cut it's number of CUs by 33% and get the same perf, as, well, RAM bandwidth isn't changing.

I'd say, it's a revolutionary concept! Kudos!

What?

I said if you increase CUs you should increase bandwidth, or you create bottleneck as you don't have enough bandwidth to feed the CUs, as ideally you want it to scale with it.

Your example makes literally no sense and you are actually trying to walk back your original point now.

 

[Hawking Radiation]

SSD SSD SSD SSD....

 

[BluRayHiDef]

Thanks a lot for these calculations. Can you or anybody else put where higher numbers are better and where lower numbers are better for the less technical minded folks

You're welcome. Also, I've done as you've asked; just go back to the post to see the changes.