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

iroboto

Daft Funk
Legend
Supporter
I think there's a lot of validity in the argument that you can't scale down to HDD levels of performance (unless you have a very big pool of RAM to pre-cache a significant proportion of the game to).
Right. But resulting experience of games wouldn’t be much fun either. Instead of stalling trying to stream random assets in and out; now your loading 3-4X the amount of textures into memory to account for random access during the level load to compensate for slower seek times.

It’ll work, but not fun or desirable to sit there waiting on a slower HDD just to get the game going.

Consoles are already very convienent with suspend mode to get going into gaming and get out “20 minute sessions between breaks”. PCs still have a load time to deal with which can be reduced with SSD, but I would definitely check out of PC gaming if they decided to go the RAM drive route and slow HDDs. My time is worth more than a loading screen.
 
I think the important thing to understand is that PC games aren't going to target these super fast drives for their games as they're not common enough.

Truth be known the most common SSD on PC is of the SATA variety so games will likely use that as lowest case and a 2.5Gb/s NVME drive as the upper bound to not only allow the Series consoles to run the game but also to not make it so high that only the highest end PC's can enjoy.

Does anyone know if there's any utilization data available for Direct Storage? I'm curious as to how much GPU resource is required to achieve X amount of throughout sand how that would affect the average gaming PC.
 

iroboto

Daft Funk
Legend
Supporter
I think the important thing to understand is that PC games aren't going to target these super fast drives for their games as they're not common enough.
Not all PC games sure. There's a lot of variety type gaming out there.
But games for PS5 and Xbox Series consoles will start pushing the envelope once last generation is released. This should free them to place more pressure on taxing their nvme drives and reduce streaming pool sizes, increasing the available VRAM footprints, and leveraging RDNA2 features.

If DX12U is going to be a baseline for games as being the requirement to match this generation of console features, then having an equivalent SSD speed is ideal.
 

pjbliverpool

B3D Scallywag
Legend
Right. But resulting experience of games wouldn’t be much fun either. Instead of stalling trying to stream random assets in and out; now your loading 3-4X the amount of textures into memory to account for random access during the level load to compensate for slower seek times.

It’ll work, but not fun or desirable to sit there waiting on a slower HDD just to get the game going.

Consoles are already very convienent with suspend mode to get going into gaming and get out “20 minute sessions between breaks”. PCs still have a load time to deal with which can be reduced with SSD, but I would definitely check out of PC gaming if they decided to go the RAM drive route and slow HDDs. My time is worth more than a loading screen.

I agree, even if developers were to go to the trouble of implementing a pre-caching system capable of eliminating the need for an SDD for the small subset of systems with large RAM pools and HDD's, it would still be a fall back option rather than first choice. But that's not to say it wouldn't be worthwhile if it can expand the target market for the game, even if those users are getting a compromised experience. The ideal situation is obviously still a fast NVMe drive. Or even better, a fast NVMe drive with a big RAM pool to support it.

That said, I'm not sure the initial load times would have to be too extortionate as you've still initially just load up what you need to start the game, and then you could build and maintain the pre-cache itself though background streaming. Again, to emphasise, I'm talking about a potential fall back position here which likely wouldn't work in all situations and would be essentially a last resort option to expand the target user base. The pre-cache potential gets a lot more interesting when paired with SATA SDD's or low end NVMe's though.
 

pjbliverpool

B3D Scallywag
Legend
Does anyone know if there's any utilization data available for Direct Storage? I'm curious as to how much GPU resource is required to achieve X amount of throughout sand how that would affect the average gaming PC.

https://back2gaming.com/guides/nvidia-rtx-io-in-detail/

When asked about the performance hit of RTX IO on the GPU itself, an NVIDIA representative responded that RTX IO utilizes only a tiny fraction of the GPU, “probably not measurable”.
 
Excellent. So you both acknowledge that it's possible to scale the graphics of a game downwards to reduce the IO requirement and therefore it would not require a 5.5GB/s SSD in a PC to enable a port of a PS5 game which takes full advantage of the PS5's IO.

You seem to be confusing bandwidth and IOPS.

Reducing asset quality in R&C may reduce the bandwidth requirements to be low enough for a PC to handle but the game may still be pulling more IOPS from PS5's I/O complex then said PC can do without tanking.

In which case reducing asset quality will do very little.
 

pjbliverpool

B3D Scallywag
Legend
Not really helped....what's 'Only a tiny fraction' to Nvidia? What GPU is that comment based on?

It's a useless quote.

I would imagine a "tiny fraction" to Nvidia is something that is “probably not measurable”. Seems clear enough to me. And even if they're basing that on a 3090, "probably not measurable" isn't exactly going to turn into "GPU crippling" on something 1/3rd of the speed is it?

Besides, during actual gameplay, the streaming (and thus decompression) requirements are going to be very low. It's only at load screens that significant amounts of data are going to be pulled from disk and at those points you can afford to dedicate near 100% of the GPU to it if needs be.

You seem to be confusing bandwidth and IOPS.

Reducing asset quality in R&C may reduce the bandwidth requirements to be low enough for a PC to handle but the game may still be pulling more IOPS from PS5's I/O complex then said PC can do without tanking.

In which case reducing asset quality will do very little.

That depends very much on the page size. Say for a given 4k texture the PS5 is streaming that in 16KB pages (pretty common for games as I understand it), then if the 2k version of that texture is also streamed in 16KB pages, you need 1/4 of the pages to transfer the same texture and thus 1/4 the IOPS.
 
I would imagine a "tiny fraction" to Nvidia is something that is “probably not measurable”. Seems clear enough to me. And even if they're basing that on a 3090, "probably not measurable" isn't exactly going to turn into "GPU crippling" on something 1/3rd of the speed is it?

Besides, during actual gameplay, the streaming (and thus decompression) requirements are going to be very low. It's only at load screens that significant amounts of data are going to be pulled from disk and at those points you can afford to dedicate near 100% of the GPU to it if needs be.

All speculation on your part with no data to back your theory up, and I don't trust Nvidia as far as I can throw them as their whole RTX I/O presentation was bullshit.

That depends very much on the page size. Say for a given 4k texture the PS5 is streaming that in 16KB pages (pretty common for games as I understand it), then if the 2k version of that texture is also streamed in 16KB pages, you need 1/4 of the pages to transfer the same texture and thus 1/4 the IOPS.

Never mind, just seen another thread where someone tried to educate you on bandwidth vs IOPS and they got no where so I'm not even going to bother go there with you.
 

pjbliverpool

B3D Scallywag
Legend
All speculation on your part with no data to back your theory up, and I don't trust Nvidia as far as I can throw them as their whole RTX I/O presentation was bullshit.

Speculation? You asked for information on the GPU utilization of Direct Storage decompression and I provided you with a direct quote from Nvidia about the utilization on their GPU's (you're welcome by the way). But it seems that because the answer doesn't suite the narrative you were clearly hoping to build that you're willing to dismiss as "bullshit" the only source on the internet that begins to answer your query.
 
Speculation? You asked for information on the GPU utilization of Direct Storage decompression and I provided you with a direct quote from Nvidia about the utilization on their GPU's (you're welcome by the way). But it seems that because the answer doesn't suite the narrative you were clearly hoping to build that you're willing to dismiss as "bullshit" the only source on the internet that begins to answer your query.

Yes, I asked about GPU utilization for Direct Storage and you provide a quote from Nvidia about RTX I/O which may have a completely different set of performance characteristics and overheads then pure Direct Storage.

An extreme fail of reading comprehension on your part.
 

iroboto

Daft Funk
Legend
Supporter
Again, to emphasise, I'm talking about a potential fall back position here which likely wouldn't work in all situations and would be essentially a last resort option to expand the target user base. The pre-cache potential gets a lot more interesting when paired with SATA SDD's or low end NVMe's though.
Just thinking out loud, the user base would easier support pcie3/4 nvme drives than they would upgrading their ram pool much higher.

looking at laptops specifically if we’re talking about expanding the user base, nvme support is there, getting memory up to 32/48/64 GB of memory is harder. Most of the population isn’t interesting in investing in so much memory to just play games.

It takes some serious jobs to drive that much memory, offers no additional performance improvement outside of games if you already aren’t RAM limited and using it as a large scaled cache seems less functionally desirable than nvme drive improving all I/O aspects for all of PC functionality.

technically ram drives would work, but it’s a very tiny niche of PC owners to have a huge ram pool and a slow HDD. Also it takes some additional programming to support this type of setup. Most games are not willing to fill up 64GB of memory.

I think the publican and devs are funneling themselves towards nvme. You’re more than likely to just get nvme drives if money is no object And laptop owners have less challenges getting nvme drives.
 

pjbliverpool

B3D Scallywag
Legend
Just thinking out loud, the user base would easier support pcie3/4 nvme drives than they would upgrading their ram pool much higher.

looking at laptops specifically if we’re talking about expanding the user base, nvme support is there, getting memory up to 32/48/64 GB of memory is harder. Most of the population isn’t interesting in investing in so much memory to just play games.

It takes some serious jobs to drive that much memory, offers no additional performance improvement outside of games if you already aren’t RAM limited and using it as a large scaled cache seems less functionally desirable than nvme drive improving all I/O aspects for all of PC functionality.

technically ram drives would work, but it’s a very tiny niche of PC owners to have a huge ram pool and a slow HDD. Also it takes some additional programming to support this type of setup. Most games are not willing to fill up 64GB of memory.

I think the publican and devs are funneling themselves towards nvme. You’re more than likely to just get nvme drives if money is no object And laptop owners have less challenges getting nvme drives.

Oh yes I absolutely agree. However in PC's where there is generally additional RAM available over consoles, I think there's benefit to be taken from making the most of pre-caching to ease the pressure on storage. Put another way, I think a gaming PC is more likely to have an excess of RAM over the PS5 than it is to have an equivalent SSD and so that should be taken advantage of to give a wider range of PC's access to a PS5 equivalent IO experience, even though in an ideal world, it would be better if everyone just had a 5.5GB/s or faster SSD.
 

pjbliverpool

B3D Scallywag
Legend
Yes, I asked about GPU utilization for Direct Storage and you provide a quote from Nvidia about RTX I/O which may have a completely different set of performance characteristics and overheads then pure Direct Storage.

While the relationship/crossover between RTX-IO and Direct Storage isn't entirely clear at this stage, in the context of the GPU overhead, I would wager that Nvidia are talking about the same GPU based decompression that we know DirectStorage brings. What other GPU overhead do you think they might be referring to? Or are you suggestion that RTX-IO has an entirely different GPU based decompression solution to DirectStorage that also has entirely different performance characteristics? There is actually another data point for this which I think @BRiT may have highlighted earlier in the thread which is around a RADGame Tools quote about the performance impact of one of their decompression algorithms when running on a GPU (it was very low).

An extreme fail of reading comprehension on your part.

:rolleyes:
 
While the relationship/crossover between RTX-IO and Direct Storage isn't entirely clear at this stage, in the context of the GPU overhead, I would wager that Nvidia are talking about the same GPU based decompression that we know DirectStorage brings. What other GPU overhead do you think they might be referring to? Or are you suggestion that RTX-IO has an entirely different GPU based decompression solution to DirectStorage that also has entirely different performance characteristics? There is actually another data point for this which I think @BRiT may have highlighted earlier in the thread which is around a RADGame Tools quote about the performance impact of one of their decompression algorithms when running on a GPU (it was very low).

Exactly, which is why using it for Direct Storage performance metrics is silly.

Especially as RTX I/O is all rubbish anyway, unless you're one of these people who can't see through their bullshit marketing slides?

In their presentation slide they claim 2x CPU cores can do 7Gb/s and 24x CPU cores to do 14Gb/s uncompressed.

In what world would a 2x bandwidth increase requires a 12x increase in CPU performance?

What the fuck are they using for the decompression? Intel Atom CPU's? :rolleyes:

And you believe their claims that the performance cost next to nothing when they can't even be honest in their own presentation slides? :-?

I think people need to lower their expectations for Direct Storage and RTX I/O.
 

t0mb3rt

Newcomer
I think the important thing to understand is that PC games aren't going to target these super fast drives for their games as they're not common enough.

Truth be known the most common SSD on PC is of the SATA variety so games will likely use that as lowest case and a 2.5Gb/s NVME drive as the upper bound to not only allow the Series consoles to run the game but also to not make it so high that only the highest end PC's can enjoy.

Does anyone know if there's any utilization data available for Direct Storage? I'm curious as to how much GPU resource is required to achieve X amount of throughout sand how that would affect the average gaming PC.
Direct Storage doesn't mean GPU decompression (or any sort of hardware decompression).
 

pjbliverpool

B3D Scallywag
Legend
Exactly, which is why using it for Direct Storage performance metrics is silly.

Especially as RTX I/O is all rubbish anyway, unless you're one of these people who can't see through their bullshit marketing slides?

In their presentation slide they claim 2x CPU cores can do 7Gb/s and 24x CPU cores to do 14Gb/s uncompressed.

In what world would a 2x bandwidth increase requires a 12x increase in CPU performance?

What the fuck are they using for the decompression? Intel Atom CPU's? :rolleyes:

And you believe their claims that the performance cost next to nothing when they can't even be honest in their own presentation slides? :-?

I think people need to lower their expectations for Direct Storage and RTX I/O.

Yeah.... you've misunderstood the slides man.

I assume you're referring to the slide below?

The 2 CPU cores at 7GB/s are referring to IO overhead only (data is not compressed, therefore no decompression necessary.) The 24 cores at 14GB/s are assuming the data is compressed and therefore realtime decompression is also required (at a 2:1 ratio), hence the significant increase in CPU requirement. It then shows RTX-IO, which in this context I take to be Nvidia's re-branding of Direct Storage functionality using only 1/2 a CPU core because of both the reduced IO overhead, and the decompression being moved to the GPU. Both of which we know Direct Storage does.

geforce-rtx-30-series-rtx-io-announcing-rtx-io.jpg
 

iroboto

Daft Funk
Legend
Supporter
Oh yes I absolutely agree. However in PC's where there is generally additional RAM available over consoles, I think there's benefit to be taken from making the most of pre-caching to ease the pressure on storage. Put another way, I think a gaming PC is more likely to have an excess of RAM over the PS5 than it is to have an equivalent SSD and so that should be taken advantage of to give a wider range of PC's access to a PS5 equivalent IO experience, even though in an ideal world, it would be better if everyone just had a 5.5GB/s or faster SSD.
Id be curious to see the steam survey of PCs with > 16GB of memory. I was very comfortable with 16GB until I hit a wall with my quality analysis.

I feel like going from 16GB to 48/64 is nearly the same cost to get to nvme
 

pjbliverpool

B3D Scallywag
Legend
Id be curious to see the steam survey of PCs with > 16GB of memory. I was very comfortable with 16GB until I hit a wall with my quality analysis.

I feel like going from 16GB to 48/64 is nearly the same cost to get to nvme

It's about 12% (of 120m) I think.
 

Davros

Legend
Funny enough I'm currently thinking about buying another 32gb not because I think I need it but just because the price has gone down by £70 since jan 2020
but the slower memory (same type) 3200mhz/3000mhz hasnt gone down.
 
Last edited:
Top