Next-Generation NVMe SSD and I/O Technology [PC, PS5, XBSX|S]

[Albuquerque]

So what you're saying is most PCs - as the Steam hardware survey shows - is way more modest and not in any way representative of recent modern technologies standards?

I mean, I'm just answering the question you posited.:

Why is every PC in every forum technical debate, always equipped with a good amount of RAM, the latest bus technologies and the fastest I/O technologies available?

You could also ask why every person in a project car building enthusiast forum has at least 500 wheel horsepower, when the majority of cars on the road are podunking around with one third or less of this power.

Yes, the steam survey is correct. Why even bring it up? Are we angling for a console-vs-PC war post fest? And for what purpose, exactly?

 

[Deleted member 11852]

Not only this but system memory latency is several orders of magnitude lower than NVMe memory. Hence why that extra system memory copy has a negligible overall impact on the end to end latency.

Out of interest - because I googled it and couldn't find a reliable answer - what is the latency for a sustained copy from main RAM to VRAM over PCI? I would definitely not expect it to be slower than a reading from solid state read but what is the actual difference?

edit: faster not slower!

 

[davis.anthony]

I can't get through when you keep comparing PS5 ssd and system memory as if process isn't more streamlined on PS5.

That's got nothing to do with it, the question is:

1. Is there enough extra on PC to just brute force it
2. Is chucking everything in (faster) system RAM enough

PC streaming for spiderman tells us a lot about PS5 when you have high res textures failing to ever load in completely on the best PC hardware.

Which is none existent in the game now it's been patched and sorted out.

My little quad-core CPU can deliver a locked 60fps in Spiderman PC with higher ray tracing settings that PS5 (so more CPU load) while streaming the game world without issue.

 

[Albuquerque]

Out of interest - because I googled it and couldn't find a reliable answer - what is the latency for a sustained copy from main RAM to VRAM over PCI? I would definitely expect it to be slower than a reading from solid state read but what is the actual difference?

EDIT: Yeah this makes more sense after your edit. Cheers!!

 

[pjbliverpool]

I can't get through when you keep comparing PS5 ssd and system memory as if process isn't more streamlined on PS5.

More streamined yes. But that doesn't mean faster. It means requires more resources for the same result.

Decompression is i/o related.

Yes, and it impacts the level of work placed on the CPU during those decompression tasks, not necessarily the actual data throughput rate provided your CPU is able to decompress in real time up to the limits of your SSD.

Abstraction is i/o related.

Yes, and again it relates to CPU overhead.

Pc streaming for spiderman tells us a lot about PS5 when you have high res textures failing to ever load in completely on the best PC hardware.

Which is obviously a bug or else why would they fail to ever load if it's purely a performance issue?

All I ask is patience. Let us see. Random generation can stress streaming performance.

Sure it can. And if you're streaming from system memory rather than a much, much slower NVMe drive then you're obviously going to be better equipped to deal with that stress.

Out of interest - because I googled it and couldn't find a reliable answer - what is the latency for a sustained copy from main RAM to VRAM over PCI? I would definitely expect it to be slower than a reading from solid state read but what is the actual difference?

I can't find a definitive answer without doing extensive searches but the initial links I have found suggest it's in the nanosecond range (vs microsecond range for NVMe reads). e.g.

Retimers to the Rescue Webinar: PCI-SIG® Q&A | PCI-SIG

pcisig.com

 

[CD33]

Which is none existent in the game now it's been patched and sorted out.

Erroneous again. This is still an open issue. Alex of DF thought it "strange" that the 4090 couldn't stream in high res textures in timely manner. In reality, there is no strangeness here.

 

[Deleted member 11852]

Just curious, why do you think it's slower?

Apologies, I meant faster. I dropped a word. Will edit.

 

[chris1515]

That's got nothing to do with it, the question is:

1. Is there enough extra on PC to just brute force it
2. Is chucking everything in (faster) system RAM enough


Which is none existent in the game now it's been patched and sorted out.

My little quad-core CPU can deliver a locked 60fps in Spiderman PC with higher ray tracing settings that PS5 (so more CPU load) while streaming ghe game world without issue.

Spiderman have bad multithreading, it scales much better with frequency than core count. Two cores are saturated and the other cores are underused. They want to improve this and if it is ready for Spiderman 2 then I expect your CPU to suffer a lot and I expect streaming average speed to go higher and depending what they are doing if for example the Doctor Strange rumor to be true peak speed to be much higher.

But you have time to buy a better CPU when it will arrive on PC probably two or three year after PS5 release and the game will probably use Direct Storage.

 

[Albuquerque]

Apologies, I meant faster. I dropped a word. Will edit.

Ahhh yeah, this totally makes more sense now! Not that I've ever dropped an important word before in any of my own posts of course!

 

[Deleted member 11852]

Spiderman have bad multithreading, it scales much better with frequency than core count. Two cores are saturated and the other cores are underused.

This feels like a massive challenge given the combinations and variety of configurations of cores, threads, clocks, cache arrangements, bus and RAM speeds. How TF do PC developers even get close to optimising code for 200 common configurations and 5,000 uncommon ones compared a handful of consoles?

 

[CD33]

Sure it can. And if you're streaming from system memory rather than a much, much slower NVMe drive then you're obviously going to be better equipped to deal with that stress

PS5 streams from disk directly into unified gddr6 with game ready data for cpu and gpu. System memory much slower but you talk as if the slow steps doesn't exist. What is pcie transfer rate for ddr4 between cpu and gpu?

 

[Deleted member 11852]

PS5 streams from disk directly into unified gddr6 with game ready data for cpu and gpu. System memory much slower but you talk as if the slow steps doesn't exist. What is pcie transfer rate for ddr4 between cpu and gpu?

Whilst I don't have exact figures here, I'm sure you wrong on this. Solid state storage is orders of magnitude slower than reading from RAM and even allowing from favourable compression (then decompression) writes to RAM, there is no way PS5 reading from SSD is faster than a PC moving data from main RAM to elsewhere in main RAM or VRAM.

What PS5 has going for it is its SSD capacity (850Gb) is vastly larger than most PCs surplus RAM that can be used as a cache. But that's a different situation.

 

[Shifty Geezer]

Bad approach to think about it. PS5 i/o can swap RAM data on per frame basis. PC can't. System RAM is not a suitable replacement. It is a mediocre bandage.

This isn't true. I/O isn't just SSD speed. We are discussing hardware software solution and integration. PS5 is unmatched so far. Let us be patient to see DirectStorage. Hopefully it is a big winner.

This line of discussion is completely moot. We have the hardware requirements for this game, Returnal, which state 32 GBs, 8 GB VRAM, 60 GB storage, no specification for that to be fast storage. Ergo, for this game, according to the developers (to be confirmed in testing when released), 32 GBs RAM is plenty enough of a replacement for PS5's IO system. Then we have the other minimum requirements which specify 16 GBs RAM, 6 GBs VRAM, and ideally an SSD. No requirement set for that to be a super-fast NVMe, only that it's relatively low access latency versus HDD. PS5's super-awesome-mega-ultra-IO++ clearly doesn't come into executing Returnal on Windows (pending evaluation where maybe we'll see a heavily cut-down game even on recommended specs...).

 

[chris1515]

This feels like a massive challenge given the combinations and varity of configures of cores, threads, clocks, cache arrangements, bus and RAM speeds. How TF to PC developers even get close to optimising code for 200 common configurations and 5,000 uncommon ones compared a handful of consoles?

I suppose CPU recommandation will go higher when Insomniac will have done the job on the game engine. 6 cores will probably be the minimum configuration.

 

[Remij]

Bad approach to think about it. PS5 i/o can swap RAM data on per frame basis. PC can't. System RAM is not a suitable replacement. It is a mediocre bandage.

LOL... you know Sony put that super fast SSD in the PS5 in lieu of more costly RAM, right? That's LITERALLY the replacement for it.

For Returnal... it's a case of:
Oh you have a fast SSD? We can stream in data no problem. 16GBs is enough.
Oh, you have a HDD or SATA SSD? We will buffer more level data into memory to ensure there's no problems during gameplay. 32GBs recommended (16x2)

That's it. That's literally what some of us here have been saying since the very start. That RAM requirements would go up as developers used more RAM as a buffer.

 

[pjbliverpool]

PS5 streams from disk directly into unified gddr6 with game ready data for cpu and gpu. System memory much slower but you talk as if the slow steps doesn't exist. What is pcie transfer rate for ddr4 between cpu and gpu?

This post aptly demonstrates why this is a topic you really shouldn't be debating.

The PCIe transfer rate between system memory and VRAM is up to 32GB/s assuming a PCIe 4.0 16x connection.
The system memory itself can transfer at 51.2GB/s assuming DDR4 3200Mhz but it would be limited by the PCIe interconnect.

By contrast the PS5 interconnect between the SSD and it's GDDR5 memory tops out at 7.5GB/s.
And of course the transfer rate of the SSD itself tops out as 5.5GB/s

Add to that the read latency of an NVMe drive is in the dozens to hundreds of microeconds while the read latency from DDR4 is in the 10 nanosecond range, and I hope it's clear that you need to rethink some of your assumptions on this topic.

 

[Albuquerque]

Add to that the read latency of an NVMe drive is in the dozens to hundreds of milliseconds while the read latency from DDR4 is in the 10 nanosecond range, and I hope it's clear that you need to rethink some of your assumptions on this topic.

Minor correction here -- NVMe storage latency would be in the microseconds (dozens, perhaps hundreds) rather than milliseconds. A modern spinning magnetic hard disk would be measured in milliseconds for sure. Regardless, microseconds are still several orders of magnitude slower than the nanosecond latency of main memory.

Large -> Small -> Tiny
Milli -> Micro -> Nano

But yeah, your point still stands: If someone is somehow confused about how NVMe storage performance (both bandwidth and latency) stack up against main memory, then this isn't the conversation for such a person.

 

[chris1515]

This post aptly demonstrates why this is a topic you really shouldn't be debating.

The PCIe transfer rate between system memory and VRAM is up to 32GB/s assuming a PCIe 4.0 16x connection.
The system memory itself can transfer at 51.2GB/s assuming DDR4 3200Mhz but it would be limited by the PCIe interconnect.

By contrast the PS5 interconnect between the SSD and it's GDDR5 memory tops out at 7.5GB/s.
And of course the transfer rate of the SSD itself tops out as 5.5GB/s

Add to that the read latency of an NVMe drive is in the dozens to hundreds of milliseconds while the read latency from DDR4 is in the 10 nanosecond range, and I hope it's clear that you need to rethink some of your assumptions on this topic.

No latency of NVME SSD are lower than this for the best below 10 microsecond to the worst around 225 microseconds.This is 0.01 ms and 0.225 ms. Basically RAM or NVME SSD aren't a problem for load data per frame. And we agree having 32 GB or RAM and a HDD is ok, it is just meaning some longer loading time.

 

[pjbliverpool]

Minor correction here -- NVMe storage latency would be in the microseconds (dozens, perhaps hundreds) rather than milliseconds. A modern spinning magnetic hard disk would be measured in milliseconds for sure. Regardless, microseconds are still several orders of magnitude slower than the nanosecond latency of main memory.

Large -> Small -> Tiny
Milli -> Micro -> Nano

But yeah, your point still stands: If someone is somehow confused about how NVMe storage performance (both bandwidth and latency) stack up against main memory, then this isn't the conversation for such a person.

Thanks, just a mistype (my earlier post did refer to microsecond latency), I'll edit it.

 

[PSman1700]

Yes, compressed.

Games dont dont really have to be compressed on PC, most gamers looking to play higher end games arent stuck with 625GB ssd's these days.

PS5 is unmatched so far.

It was outclassed before it even released.

Why is every PC in every forum technical debate, always equipped with a good amount of RAM, the latest bus technologies and the fastest I/O technologies available? I know mine is, but the Steam hardware survey tells me most PCs are vastly behind the consoles that launched more than two years ago.

DF not too long ago shared a source stating that there are more GTX2060S or better class GPUs out there in the hands of gamers, thats just the dGPU market. Consoles where behind, they sure are now. Besides, Steam survey's dont show the whole or even accurate picture, anyways.

But you have time to buy a better CPU when it will arrive on PC probably two or three year after PS5 release and the game will probably use Direct Storage.

He has the time, but does he really need to change the CPU? Yes 8 cores is ideal, however the PS5 has to do with 6 cores (6.5?) as far i understand, hence the bad multi core usage in spiderman.

What PS5 has going for it is its SSD capacity (850Gb) is vastly larger than most PCs surplus RAM that can be used as a cache. But that's a different situation.

Obviously. However in that case a gaming PC would be sporting atleast 1TB of fast nvme drive.