Tim Sweeney explains his comments with respect to IO pef on PS5

[hishnash]

Tim Sweeney has expanded on this comments about the IO perf on the PS5 to explain in more detail why he says even modern PCs with (on paper much faster SSDs) cant provide they same perf as the PS5.
 

 

you can read more on https://wccftech.com/tim-sweeney-explains-exactly-why-the-ps5s-ssd-and-i-o-architecture-is-way-more-efficient-than-pcs/ 

 

In summary he makes very good points, something the LTT notices when there had high performance storage server as well that (for latency in particular) how the OS operates with the SSD is vital to get the most out of your drive. Add to that the hardware decompression (so that you don't need to jump to user-space cpu code) his comments make a lot of sense and are correct. If on a PS5 you can directly stream (compressed) data from the SSD into the GPU without needed to go through the kernel.

For supporting features like `nanite` low latency seems to be very important since you cant load the full mesh into VRAM so you need to stream it on demand to the GPU as and when it is needed, any hick-up (due to windows update running on the cpu) will result in massive frame rate drops.

 

EDIT:

 

For some clarity I should explain how the normal flow is when a game needs to load compressed data from your SSD (on a PC)

Legend:

[GC] game code running on the cpu
[DC] driver code running on the cpu in user space (eg dx12.dll)
[K] kernel code
[KD] kernel side driver code


[GC] Detects that it needs data -- (sends request to load data from disk) -> [K]
[K] this program as permission check has (might need to check in ram)
[k] -> [SSD] send command to SSD to load data to RAM
(there are a few ways this can be done but the most common is to wait for the SSD to response)
[SSD] -- "I have data" -> [k] 
[k] Set up data copy to [RAM]
[K] ---data is streaming in it will be here...- > [GC]
[GC] - im going to need some place to put this decompressed data--> [k]
[k] - ok, put it here on ram-> [CG]
[GC] -- saves decompressed data to address --> [RAM]
[GC] ---dx12 please copy this data to the GPU when you have time, and let me know when your done-> [DC]
[DC] --please share this data with the GPU--> [K] 
[KD] Configure memory address to be assessable to GPU
[KD] --copy this data to your VRAM please-> [GPU]
[GPU] - done -> [KD]
[KD] - done > [DC]
[DC] - done > [GC]
[GC] - cool lets render that thing please -> [DC]
[DC] - please tell the gpu to use that data in this way..> [KD]
[KD] - please use this data you have in you VRM in this way..> [GPU]
[GPU] render frame

--

We don't know the exact pipeline that the PS5 will use but from these comments we can assume that it does not require all of these jumps through the cpu to do the decompression etc. 

[minibois]

I have not read much into this, other than Sweeney's original comments, a thread on Reddit going over what the speeds could mean and now these tweets (and your post), but I have thought a bit about this.

What I thought about, is that there file formats that are:

a. small, highly compressed

b. large, not very much compressed (like what is used in video editing, so speed up the editing process)

(and of course anything inbetween).

 

Perhaps the PS5 will be working with option b., files that are not compressed much, that are easier to play; but need more space on the drive.

That would make the games larger, but would make them easier to play. Would also make porting to the console harder (or more likely; would make it so the PC ports are not conventionally 'good', because it will rely on other tech).

So I will remain neutral, until the PS5 comes out and we can see its performance, but that tech I mentioned could be in use.

 

Oh, but of course I thought about how a 1TB drive (which I think the PS5 will have?) might be kind of small for the games in a much more uncompressed form. So maybe they will also (need to?) implement some form of game streaming.

[FloRolf]

2 minutes ago, minibois said:

So maybe they will also (need to?) implement some form of game streaming.

Wouldn't that completely defy the purpose of all that low latency high bandwidth storage stuff? Most people have 50/10 connections or even worse. 

 

I agree though, Sweeney and Sony can tell me what they want, unless I see real world results it doesn't matter to me. In all honesty, I don't expect much. 

[WereCatf]

10 minutes ago, minibois said:

Perhaps the PS5 will be working with option b., files that are not compressed much, that are easier to play; but need more space on the drive

Your reasoning doesn't make any sense. There is a hardware-decompressor for any compressed data on the PS5, there is no "easier to play." The compressed data doesn't have to go through the CPU to get decompressed and the hardware-decompressor can do it at 9GB/s, as mentioned in the technical live-stream some weeks back.

[minibois]

10 minutes ago, FloRolf said:

Wouldn't that completely defy the purpose of all that low latency high bandwidth storage stuff? Most people have 50/10 connections or even worse. 

Hmm, conventional game streaming wouldn't really work for that..

What I was thinking, maybe the game can download like the first half (or quarter or whatever, depends on how large the game is) and once you almost reach that point, it downloads new content in the background (and deletes the old content). That way you only have (for example) 50GB of the game on your system at a time.

 

I surely hope they won't do this, for multiple reasons, but with the storage claims being made, it almost makes me wonder how large games will become over time.

1 minute ago, WereCatf said:

Your reasoning doesn't make any sense. There is a hardware-decompressor for any compressed data on the PS5, there is no "easier to play." The compressed data doesn't have to go through the CPU to get decompressed and the hardware-decompressor can do it at 9GB/s, as mentioned in the technical live-stream some weeks back.

Okay, good to know

[3rrant]

I opened a thread in general about this right after the wan show for how poor and extremely misinformed LMG as a whole an Linus in particular have shown to be regarding this whole issue. Embarrassing to say the least.

 

Also it's 5.5 to 9gbs UP TO 22.5gbs for "data that compress particularly well", taken directly from Cerny during the talk.

[porina]

1 minute ago, WereCatf said:

The compressed data doesn't have to go through the CPU to get decompressed and the hardware-decompressor can do it at 9GB/s, as mentioned in the technical live-stream some weeks back.

A claimed "typical" rate of 9GB/s, with peak more than double that.

 

I do agree, we have to see how this really works in practice. It kinda frustrates me that while hard disk to SSD can give multiple times faster loading speed reduction, the difference between SSD and ramdisk isn't much at all as bottlenecks lie elsewhere. It would be interesting to see how low you can go if you're not held up by the CPU.

 

Linus' comments on this topic during the Wan show I feel were really off the mark, as he focused almost purely on the raw storage speed and nothing else about the rest of the system, only giving the slightest nod near the end when Luke tried to expand on it.

[WereCatf]

5 minutes ago, 3rrant said:

I opened a thread in general about this right after the wan show for how poor and extremely misinformed LMG as a whole an Linus in particular have shown to be regarding this whole issue. Embarrassing to say the least.

One thing that irks me is how people here constantly go "PCs can also have 4GB/s NVME SSDs!!!!oneoneleven", completely ignoring the hardware-decoder and the fact that it can directly decompress stuff to RAM. It's obviously not some magic bullet, but if definitely does reduce the latency caused by certain kinds of tasks a huge amount. With PCs, all the data has to be decompressed by the CPU, doing this silly round-trip of storage->RAM->CPU->(V)RAM.

 

There's a lot more to the whole thing than just the raw speed of the SSD itself.

[Zodiark1593]

Wouldn’t this be pretty much DMA on roids?

[Sauron]

They had bullshit claims like this last gen too, don't bet your house on it.

[3rrant]

Just now, Sauron said:

They had bullshit claims like this last gen too, don't bet your house on it.

No. They did not have hardware diagrams and full detailed talks about it  look at the road to ps4 and it's an hour how "what did I do to became hardware architect at PlayStation" by Cerny. The new talk is something really really different.

[WereCatf]

6 minutes ago, Sauron said:

They had bullshit claims like this last gen too, don't bet your house on it.

What part, exactly, are you claiming to be bullshit? Do tell, since you're the one making the claim.

[WereCatf]

9 minutes ago, Zodiark1593 said:

Wouldn’t this be pretty much DMA on roids?

Yes, basically. Instead of just shuffling data around, the controller can also manipulate it. We also don't know yet all the different ways it can manipulate the data on the fly, we just know that the decompression is one of the big features of it.

[hishnash]

38 minutes ago, minibois said:

Perhaps the PS5 will be working with option b.,

I think rather option `a` since he talks about having a hardware accelerated decompression system and this being why the IO is so much faster. Doing this in the cpu (user space) is both slow, but also massively damages latency and means you are using up CPUs compute you could be using for the games runtime.
 

16 minutes ago, Zodiark1593 said:

Wouldn’t this be pretty much DMA on roids?

Normally there is very little point in DMA to the SSD from the GPU if the data on the SSD is compressed, then you end up using up most of you GPU compute de-compressing (compression algorithms are very hard to optimise for GPU). Sounds like there is a dedicated decompression silicon to offload this.

 

the other side of it seems to be a lot of OS level optimisations, we all know this is something that windows is very bad at.

17 minutes ago, WereCatf said:

There's a lot more to the whole thing than just the raw speed of the SSD itself.

When you consider latency being very very important for this use-case the raw `speed` of your SSD measured in sequential read is tangental to what is important.

 

[Astroflash]

What will this mean for PC gaming going forward? Will we see architectural changes?

[Sauron]

5 minutes ago, WereCatf said:

What part, exactly, are you claiming to be bullshit? Do tell, since you're the one making the claim.

"The upcoming console is going to be somehow faster than computers with better specs because magic optimization" has been done before and it was nonsense. I won't believe it until I see a benchmark.

[hishnash]

2 minutes ago, Astroflash said:

What will this mean for PC gaming going forward? Will we see architectural changes?

Need large OS kernel changes first. 

It is the sort of thing that apple could do easily but MS would struggle massively. Possibly we could see stuff like the Radeon SSG were SDDs dies are directly attached to the GPU. 

 

1 minute ago, Sauron said:

"The upcoming console is going to be somehow faster than computers with better specs because magic optimization" has been done before and it was nonsense. I won't believe it until I see a benchmark.

When the PS3 came out if you needed to do large sale compute it was massively cheaper to buy large numbers of PS3s and install linux on them. This is very common, in general desktop PC's lags behind in inovation for a lot of reasons but mainly due to the chicken and egg issue,

 

if there were a motherboard vendor who did the work Sony is doing they would end up putting in billions on R&D and then to ship it at a usable price they would need to ship more motherboards than the entier current PC market consumers. (Sony makes money on games, and sells consoles at cost/loss). They would then also need to write thier own OS (and somehow get users to use it)...

 

[3rrant]

2 minutes ago, Sauron said:

"The upcoming console is going to be somehow faster than computers with better specs because magic optimization" has been done before and it was nonsense. I won't believe it until I see a benchmark.

Leaving aside the fact that there is absolutely no 1.8tf card on PC that can come even remotely close to the graphic settings of a PS4 in games like RDR2 etc? Don't be blind.

[vetali]

Throughput and raw numbers != real world performance. Shocker.

[WereCatf]

6 minutes ago, Sauron said:

"The upcoming console is going to be somehow faster than computers with better specs because magic optimization" has been done before and it was nonsense. I won't believe it until I see a benchmark.

Translation: "I have zero fucking clue what the discussion is actually about, so I'll just try to throw some random shade and hope I'll look cool"

[Sauron]

4 minutes ago, 3rrant said:

Leaving aside the fact that there is absolutely no 1.8tf card on PC that can come even remotely close to the graphic settings of a PS4 in games like RDR2 etc? Don't be blind.

Says who...? Based on what information detailing the settings RDR2 runs on a ps4...? Based on what benchmark performance measured on a ps4...? Not to mention that gpu performance can vary over time depending on gpu generation etc, flops are not a particularly good indicator of absolute performance.

 

Also here we're talking about hardware that, on paper, is faster, not equally fast. Hey, maybe it's true - but this is something that has been blatantly lied about in the past and mr.Sweeney has a vested interest in hyping the console.

6 minutes ago, hishnash said:

When the PS3 came out if you needed to do large sale compute it was massively cheaper to buy large numbers of PS3s and install linux on them. This is very common, in general desktop PC's lags behind in inovation for a lot of reasons but mainly due to the chicken and egg issue,

The PS3 had radically different hardware from PCs, it wasn't directly comparable. Current consoles just have slightly modified AMD chips.

2 minutes ago, WereCatf said:

Translation: "I have zero fucking clue what the discussion is actually about, so I'll just try to throw some random shade and hope I'll look cool"

If you say so. I feel like I've made my case just fine.

[RejZoR]

It's not even about raw speed. It's about out of the box expectations. Developers know for a fact what storage speed they'll have on every single PS5 and they can code games to seamlessly stream assets during gameplay to a point they can load them directly through streaming system. Streaming basically means loading only content in visual field of the player and a bit more outside of field of view in a certain area and nothing else outside of that. Meaning they can create huge worlds without a single load screen and no limitations for the size of world itself and never experience a problem with it on any PS5.

 

On PC however, they can't know what kind of storage will be in use and what will be its speed, meaning they can't code the game that way and need to take slowest method (HDD) as the design point. Unless they can add a mechanism which would detect game residing on a stupid fast SSD and turn the streaming logic same way as it is on PS5. Question at that point is, is it even worth their time designing it this way or just doing it old way of loading all of it to memory with partial streaming?

[hishnash]

1 minute ago, Sauron said:

The PS3 had radically different hardware from PCs, it wasn't directly comparable. Current consoles just have slightly modified AMD chips.

 

 

they (at least the PlayStation) have radially different operating systems. If you'r thinking about optimisation that is in the OS for these things.

 

1 minute ago, RejZoR said:

Unless they can add a mechanism which would detect game residing on a stupid fast SSD and turn the streaming logic same way as it is on PS5.

Its not even that simple, on a PC you have all sort of other shit running, (maybe windows is downloading an update in the background). the Software optimisations in the operating system are just as important (if not more) for having a predictable throughput. 

 

[Sauron]

1 minute ago, hishnash said:

they (at least the PlayStation) have radially different operating systems. If you'r thinking about optimisation that is in the OS for these things.

I'm talking hardware.

[Commodus]

13 minutes ago, Astroflash said:

What will this mean for PC gaming going forward? Will we see architectural changes?

What I suspect: gaming on PCs will have a partial handicap for a while.  While the gap in real-world performance might not be as cavernous as the raw specs would suggest, PCs will have a lot of catching up to do on storage technology.

 

Remember, most PC games are still based around spinning hard drives.  It's going to be a long time before game devs can assume that PCs have NVMe SSDs inside, and it's still unclear how the necessary overhead of a standard PC will limit those drives' potential.  One thing's for sure: if you're hoping to play games that behave like they will on the PS5, you should make the leap to all-NVMe storage when you can.