PC system impacts from tech like UE5? [Storage, RAM] *spawn*

Yes, and hopefully that UE5 demo will be shown tomorrow on the keynote.

 

If it already looks great on 9TF hw then imagine a tech demo on a 3080Ti/optane system.

 

Optane doesn't have the raw read speed of sony solution. Advantage optane has is in random access which is unlikely to be bottle neck on these types of streaming systems that load large assets through handtuned priority queues. Also optane doesn't have efficient compression algorithm in it which makes the raw speed matter even more.

 

It doesn't sound like the demo requires a PS5 level storage solution anyway, so a faster solution (if it existed) would likely add nothing.

 

In this video Epic is claiming they're fetching assets from the SSD every frame and more than once on some frames, it does look like it's very dependent on PS5's unique storage.

Oh and no, 4x super expensive Optanes in RAID0 still won't help. I've explained this many times but I've installed entire games on a ~8GB/s ramdrive and the loading time difference wasn't large. The games still need to be compressed (or they'd occupy 20TB of data) so the bottleneck is on asset decryption and decompression which rests on CPU cores (usually just one core even).
Until there's something like a 8x PCIe 4.0 / 16x PCIe 3.0 dedicated card that takes in a PCIe 4.0 NVMe SSD and has an ASIC that does the same as PS5's dedicated hardware, the PC has nothing remotely similar to the PS5's I/O performance.

I guess maybe on a threadripper that can dedicate something like 8+ Zen2 cores to decompression from the SSD it could be done (if it doesn't hit a bandwidth bottleneck somewhere), but I doubt any developer is going to bother with making a brute-force software solution for such a small audience.

 

Just because they're streaming assets constantly doesn't mean you need the PS5's storage solution to do it. In fact you can bet your ass it won't, because it's one of UE5's new main features, it needs to work great on every platform (possibly ignoring HDDs)

 

Theres more to it then transfer speeds, Optane is having other advantages, together with velocity architecture it probably is faster then ps5. SC warp demo was atleast as inpressive if not more.

 

If sony achieves those transfer speeds they advertise + the compression works like advertised there is a lot to be said about that transfer speed. Optane is nowhere near. You can of course claim sony will not reach those advertised speeds and compression and priority queues don't work but that is an argument only people under nda would know answer to. However we know what tim sweeney said about god tier and sony solution being better than what is available on pc though.

 

At some point, they're fetching geometry out of 33 Million Polygon models every 33ms or less. It sure sounds like they're pushing the I/O effective speeds to their limit, and the only thing that comes close to the PS5 in that is the SeriesX.
There's nothing in the PC with that decompression performance, and I doubt the PC will get it until dedicated ASICs come up.

Perhaps the GPU makers will embed M.2 NVMe slots into their graphics cards at some point, and include the decompression hardware in their GPUs?

Unreal Engine 5 will obviously work on all platforms (like Unreal Engine 4 before it).
You can bet your ass this demo won't be running on an Android / iOS smartphone anytime soon, or the Nintendo Swith, or any current-gen console.

The thing that sets this demo apart the most from everything else is Nanite, and there's even no indication that Nanite will even work on all platforms AFAIK.

 

Where are you getting the idea from that PC's will be using or need to use the same kinds of full SSD compression that the new consoles are?

They don't right now, and Kraken/BCPACK are getting you at best 2:1 compression so not using it isn't suddenly going to make a 100GB game balloon to 20TB.

5GB/s SSD's (faster than XSX compressed throughput) already exist on the PC without console style compression and 7GB/s will be the high end offering by the time the consoles launch without any need for specialised decompression hardware at all.

We already know Direct Storage reduces the IO overhead of the XSX 4.8GB/s throughput to 1/10th of a Zen 2 core so there's no concern there either.

Regarding the demo, we've had it heavily implied (possibly stated directly) that it runs on both XSX and PC so it's probably a little premature to start claiming it's only possible because of the PS5's SSD.

 

In addition to that, PC can run the demo without any SSD at all by the way, just brute force it with increased RAM + VRAM requirements.

 

You just have more pop in or stream in lower quality assets if the highest level asset is not streamable on given hw due to lacking performance. Or perhaps we hit in many games lowest common denominator and faster hw goes unused. This just made graphics quality a lot more complicated equation than flops combined with output resolution.

 

AFAIK Kraken and BCPack only work on textures that still go compressed towards the GPU, which works on compressed textures to save memory bandwidth?
I remember seeing several-terabytes being the fully uncompressed data for a typical 50-60GB game.

I never said it wouldn't run on the XBX. Despite the fact that Epic didn't say it would.

 

Yes but GPU's deal with the common compressed texture formats natively so they will remain compressed in the PC space as they are now.

The difference here is the full data compression that the consoles are offering on top of that in the form of Kraken/BCPACK which is on average no more than 2:1.

 

I/O overhead for the SeriesX, which is a solution that already has dedicated hardware for decompression. You seem to be implying that Direct Storage makes every decompression+decryption done on 1/10th of a Zen2 core, which it's clearly not the case:

It doesn't matter. Unless you think games on the PC aren't going to be compressed for the sake of pure performance throughput and occupy >200GB on the PC just for those people with 5GB/s+ NVMe drives, all the assets in the game are still coming inside zlibs, most probably with encription to the mix.
You still need hardware for decompressing assets at 6GB/s, and that's not something you have on the PC.

Like I said, it doesn't matter that you have 4x Optane drives in RAID0 capable of 20GB/s. The content is still coming compressed and you still need custom hardware for it, and that doesn't exist in the PC space.

 

Nick Penwarden, UE VP of engineering about demo rendering resolution on PS5:

Tim Sweeney

The truth is that consoles looks nice now, 6 months from release, but will be behind a gaming PC in CPU (cripple Zen2 vs full Zen3), GPU (crippled RDNA2 vs Ampere) and I/O speed (7GB/s NVME SSD are coming this summer) when they will be available. Only I/O looks a bit better, but nothing to brag about.
2 years from now, at 1/3 life of this console gen ? Well business as usual... Consoles will be in the low range of what a gaming PC can do and I/O will be solved with HW decompression on GPU or SSD controllers if they are any need of it (which I doubt, the requirements doesn'y look anything special)
Rince and repeat with next gen...

 

There's not one roadmap from AMD, nVidia, Intel, etc. mentioning hardware for high-speed decompression.
Sure the PC gets awesome performance with a 4x NVMe PCIe 4.0 SSD, but you still need to decompress the data.

I'll be happily proven wrong if/when devs come forward with a high-speed and highly-parallel decompression algorithm that works on a number of Zen2 / Skylake / IceLake cores, that decompresses data at 6GB/s or more. Or GPU makers putting NVMe slots in consumer graphics cards.
Right now what we have are single-threaded decompression algorithms, which is why people (like I experimented myself) running entire games from RAMdrives with 60GB/s theoretical bandwidth aren't getting instant loading times.


Also, Tim Sweeney is in the business of selling an engine. Don't expect him to say "the PC crowd is in deep shit and this tech will take years to appear there" when he's promoting a solution scales down towards a Nintendo Switch.

 

I don't think you're interpreting that article correctly. The IO requirements and the decompression requirements are two different things. They state that decompression is the equivalent of 3 Zen2 cores while IO would be taking up another 2 without Direct Storage - down to 1/10th of one core with it. Direct Storage is coming to the PC too so the assumption at least, until we know more is that it will have a similar impact on IO there too. Thus taking the CPU overhead for IO out of the equation.

Then you've got the 3 Zen 2 cores for BCPACK decompression which aren't applicable in the PC space because it won't be using BCPACK (or zlib as far as I'm aware) compression on the SSD. And nor does it need it in comparison to the XSX given that NVMe Gen4 drives have more raw uncompressed bandwidth than the XSX SSD does combined with BCPACK compression.

I do think they aren't going to be compressed in that way but not for any reasons relating to high speed SSD's, but rather just because they're not at the moment as far as I'm aware. Otherwise even today an NVMe 4.0 drive would be bringing most quad core systems to their knees.

I think you're mistaken on this. If data were currently compressed on PC drives in this fashion it would essentially make all modern high speed SSD's useless since it would kill CPU performance in anything outside of an 8 core monster.

If it was compressed on the PC drive in zlib or similar format and needed to be decompressed after it's left the SSD in real time prior to use then I'd agree. But I don;t think that is the case.

Incidentally some high end PC SSD's do use on drive encryption but the encryption and decryption is handled within the SSD itself and thus saves storage space, but not off drive bandwidth like the consoles.

 

I did interpret it correctly and you're confirming that the decompression alone is taking the equivalent of 3 Zen2 cores. Which neither the SeriesX or the majority of PC gamers can afford to lose.
And this is even assuming the decompression performance would scale linearly with just throwing more CPU cores working in parallel - which it may not.

AFAIK the majority of PC games use zlib. Why would devs choose not to compress the files using the same methods as the console counterparts?

They don't kill CPU performance because up until this point they were designed for storage units capable of ~40MB/s which is what you find in current consoles with HDDs. Therefore, all current and older games - console or PC - just use one CPU thread to decompress the data.
On PCs using high-speed SSDs, the loading times are limited by single-core performance.

It's not like devs could design a higher-performance decompression method for PCs with NVMe drives. There are no games demanding SSDs so far, unless you consider Star Citizen a game...

 

That'd be nice, yeah.