Current Generation Hardware Speculation with a Technical Spin [post GDC 2020] [XBSX, PS5]

Status
Not open for further replies.
From what dobwal is saying, the number (22 BG/s) is the performance of the data, not the processor. There's no point saying your processor can handle 25 GB/s theoretically if it'll never receive data fast enough to decompress that much.

So it could be that there's two stats:
The decompression block is capable of processing 25 GB/s
The amount of data coming out of the IO system through the decompression block is up to 22 GB/s

The former isn't as meaningful as the latter. I don't know where dobwal gets this info from though.
 
From what dobwal is saying, the number (22 BG/s) is the performance of the data, not the processor. There's no point saying your processor can handle 25 GB/s theoretically if it'll never receive data fast enough to decompress that much.

So it could be that there's two stats:
The decompression block is capable of processing 25 GB/s
The amount of data coming out of the IO system through the decompression block is up to 22 GB/s

The former isn't as meaningful as the latter. I don't know where dobwal gets this info from though.
Don’t those scenarios still miss very crucial information like what type of format it is being talked about? Texture can be compressed to like 8:1(without Alpha) audio 200:1. Your capability will appear to be much bigger if you are talking about a peak performance using less demanding format (audio) than more demanding texture. So we are really missing the context of “in some instance reached 22GB/s.” It could even be talking about textures or mixed data, we just don’t know until we see more in depth information on both consoles to be able to put it into context of what’s possible.
 
different strong points for two different approachs, UE5 and others engines can use both no needs to SSD-war on internet forums.

This place gets fucking crazy at console launch times.

0774bd460a2377316ae2b4c9644445c1--thor-avengers-loki-thor.jpg
 
They did. Cerny stated the figure when giving the tech talk.
Yeah, Mark Cerny said:

PS4 was using Zlib with typically 1.5x reduction in file size
Kraken 10% smaller files than Zlib (so 1.65x)
The drive does 5.5GB/s
Typically would become 8-9GB/s, (above ratios give 8.25 for zlib, 9.07 for kraken)
But it can go up to 22GB/s if the data compress well

So....
5.5GB/s min
22GB/s max

Full game average will never be anywhere near 4:1, but the max allows peaks of higher compression for data that "compress particularly well". If the full game data is an average of 1.65x, it would be 9GB/s average in theory, but only if no parts of the data compressed more than the max output of 22GB/s allows. If geometry does 3:1 with kraken, the average might be 16GB/s for that particular section of the streaming data. Same concept for RDO processing of BCn data. It need that headroom to be effective.

Basically if you clip the top of a graph, you reduce the average. If compression ratio vary a lot, this becomes more important. If the compression ratio was completely fixed, the 22GB/s figure would be completely meaningless.
 
Don’t those scenarios still miss very crucial information like what type of format it is being talked about? Texture can be compressed to like 8:1(without Alpha) audio 200:1. Your capability will appear to be much bigger if you are talking about a peak performance using less demanding format (audio) than more demanding texture. So we are really missing the context of “in some instance reached 22GB/s.”
He's talking about the current data games use. Audio is already compressed with MP3 or OGG or somesuch. Textures are already compressed DXTCs. These are the files being transferred that are seeing up to 4:1 compression with Kraken. Raw text files likely get well over, and are meaningless examples in a talk designed to describe the system. That's why Cerny stuck with ball-parks for real-world data - 8-9x typically, up to 22 GB/s. I guess conceptually there's a real hardware limit 4:1 on the decompressor, so perhaps, 22 GB/s is the literal hardware limit and a plain text file compresses down only 4:1, but I think that very unlikely.
 
He's talking about the current data games use. Audio is already compressed with MP3 or OGG or somesuch. Textures are already compressed DXTCs. These are the files being transferred that are seeing up to 4:1 compression with Kraken. Raw text files likely get well over, and are meaningless examples in a talk designed to describe the system. That's why Cerny stuck with ball-parks for real-world data - 8-9x typically, up to 22 GB/s. I guess conceptually there's a real hardware limit 4:1 on the decompressor, so perhaps, 22 GB/s is the literal hardware limit and a plain text file compresses down only 4:1, but I think that very unlikely.
I was thinking it could be the ASIC having a 4:1 local buffer size ratio. So if the data goes above 4:1 ratio it's the block input cycle that stalls, because it needs two passes to decode more than 4 times the input block size. So it creates a cap of 4x the data input rate, and they would have sized the input rate of the ASIC exactly what they need, which their design goals was 5GB, and the final product was 5.5GB.
 
He's talking about the current data games use. Audio is already compressed with MP3 or OGG or somesuch. Textures are already compressed DXTCs. These are the files being transferred that are seeing up to 4:1 compression with Kraken. Raw text files likely get well over, and are meaningless examples in a talk designed to describe the system. That's why Cerny stuck with ball-parks for real-world data - 8-9x typically, up to 22 GB/s. I guess conceptually there's a real hardware limit 4:1 on the decompressor, so perhaps, 22 GB/s is the literal hardware limit and a plain text file compresses down only 4:1, but I think that very unlikely.
So you are saying that Audio is already sitting on Disc in MP3 format or something like that and gets further compressed 4x by Kraken? I didn’t think of it like that because I read somewhere how over the years audio and textures have grown and their proportions. They were saying that sometimes up to 60% of entire game data is uncompressed audio files. My thought was that if it’s raw audio file it would be compressed something like 8x, texture would be compressed 1.5x using Kraken for PS5 game disc and using Zlib/BCPack for XSX with plus/minus set compression profiles. Then you have different game sizes due to different compression algorithms for each and everything on SSD sits in compressed state as opposed to uncompressed/lowcompressed current gen. When file is needed it is hwdecompressed into ram using the dedicated engine, amount of uncompressed data reaching ram is your bandwidth. When wanting to suspend a game and launch the next one your entire game data loaded in ram is compressed again into the temporary manner on SSD.


So I guess I might be under the wrong impression in a manner how the data is handled
 
I'm not sure if my post was moved due to the extra bit I tacked on at the end or if with multiple tabs open I accidentally posted it in the wrong thread; but the main body of it was directly relevant to recent conversation in this thread and seems totally out of place elsewhere, so I'll just quote it in here:

It was my understanding the XSX SSD is 2.4GB/s (Raw), 4.8GB/s (Compressed) and the "over 6GB/s" referred to the speed at which the decompression block ran.

What is unclear to me at this point is whether or not the 4.8GB/s Compressed figure is a hard limit somewhere in the I/O pathway and the decompression block's >6GB/s simply provides overhead. Or if it is an average and the entire pathway can actually move >6GB/s end to end in ideal circumstances. I'd hope and assume it's the latter for the sake of Xbox devs and end users.

If it is the latter; I'm surprised MS didn't capitalise on it and state "2.4-6GB/s", or more honestly "2.4GB/s (RAW), 4.8GB/s (Average Compressed), 6.0GB/s (Peak Compressed)"..... saying that, their official spec sheet only says 12TF when they have 12.15TF, so perhaps they're playing it a little more reserved.

Regarding Sony and PS5, the way it was worded makes me pretty sure it's 5.5GB/s (Raw), 8-9GB/s (Average Compressed) & 22GB/s (Peak Compressed); that 22GB/s being peak end to end throughput in what are likely very rare circumstances and as a result of a limit in the decompression block. Cerny said "..the unit itself [kraken decompression block] is capable of outputting as much as 22GB/s if the data happened to compresses particularly well".

-------

Now for a bit of levity, according to youtube closed captions -- while checking for the quote -- it turns out the unit can move over 5GB of crack a second; not to mention "format input data"..! =P

cerncrk1.jpg
cerncrk2.jpg
 
the unit itself [kraken decompression block] is capable of outputting as much as 22GB/s if the data

If that is the actual verbatim quote I read it. Completely differently and cannot be compares to 5.5gbs.

5.5gbs I assumed was input data from the ssd that can be decompressed live.

Output sounds like that is the link to apu and that is the maximum the apu can be fed with and not specifically tied to any decompression speed, just what it can sustain as output?

22 out is very close to. 5.5 in with a 4:1 compression ratio no?

It's late so I may be reading it wrongly but it tickles my inner sceptic.
 
He's talking about the current data games use. Audio is already compressed with MP3 or OGG or somesuch. Textures are already compressed DXTCs. These are the files being transferred that are seeing up to 4:1 compression with Kraken. Raw text files likely get well over, and are meaningless examples in a talk designed to describe the system. That's why Cerny stuck with ball-parks for real-world data - 8-9x typically, up to 22 GB/s. I guess conceptually there's a real hardware limit 4:1 on the decompressor, so perhaps, 22 GB/s is the literal hardware limit and a plain text file compresses down only 4:1, but I think that very unlikely.

DXTC compression ratio is low for being able to be decompressed fast in GPU, you can compress more on SSD.

http://www.jacobstrom.com/publications/StromWennerstenHPG2011.pdf
 
@RDGoodla They actually say over 6 GB/s, but I imagine if it were much more than that they'd said 7 GB/s or whatever.
That's why Cerny stuck with ball-parks for real-world data - 8-9x typically, up to 22 GB/s. I guess conceptually there's a real hardware limit 4:1 on the decompressor, so perhaps, 22 GB/s is the literal hardware limit and a plain text file compresses down only 4:1, but I think that very unlikely.

7GB/s and 22GB/s are both rather weird numbers for theoretical bus limits.
From the numbers being shown, it seems the SeriesX decompressor has a 8GB/s IF bus to the main RAM whereas the PS5's has a 24GB/s IF bus.


Regardless, the PS5's 22GB/s limit also means if we connect a M.2 7GB/s PCIe 4.0 NVMe to it, then there might be headroom to increase its effective storage bandwidth by 7/5.5 = 27%, meaning the average ~8.5GB/s output would increase towards ~10.8GB/s. Assuming there's no RAID0 capabilities (which most probably there isn't).
Which of course is probably useless for most gameplay features because devs will be targeting the lowest common denominator, but games depending on storage to reduce pop-in may actually get a bit less pop-in if people upgrade their storage with one of these fastest NVMe drives.
 
So you could have an "accelerator cartridge" where you could multiply the onboard SSD +cartridge throughput for select games. But, as if 5~9GB/s wasn't already enough! (famous last words?)
 
New consoles may render 4k for cross-gen games, but they struggle to make 4k next-gen games.
If PS5 can only render 1440p for UE5 demo, Xsx may only render 1550p~1600p images. This
is an indication that we may not see native resolution for next-gen games on 4K TV.

Previous generations developers often use the most powerful console to render "native" resolution
for main stream TVs at that time.


2013 PS4: 1080p
2005 XB360: 720p


This generation TVs improve so fast (4X resolution) that none of the new consoles can render
native resolution using true next-gen engine. In that UE5 demo new consoles may only render
50% of 4k resolution


In previous generations a more powerful console often benefit from "native resolution" of the TVs
while other consoles render blurry images for TV. "Native resolution" plays an very important
rules for modern TVs, Therefore this generation we will see much less comparison between the consoles (especially resolution).

Because it is not as meaningful as previous consoles.

4k TV: we see 45~55% of native resolution, maybe 65% with temporal projection. A more powerful
console just means a little less blurry, but not benefit from native display.


1080p TV: Both of them will render more than 1080p, with some super sampling effect. A more powerful
consoles means a little better supersampling. But it is very hard to identify unless one console
can render 1.5~2x more pixels for super sampling but it is not possible between PS5 and Xbsx.
 
Last edited:
XSX might go for higher settings instead of higher resolution, or a combination of both.

It will be resolution 90% of the time in my opinion or frame rate. It's just the simplest thing for multiplatform devs to incorporate. First party it's where the real innovation might happen.
 
Status
Not open for further replies.
Back
Top