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

We can look at Valhalla for example the min spec is quite wide.


PC games are extremely scalable these days. I don't think we need to have any concerns about PC versions holding back any console specific features.
Is Valhalla designed to max out the consoles IO, or was it designed to support PC IO and then just uses the consoles for faster loading?

It's equivalent to PC GPU;s being way, way more capable than consoles in shader abilities but games never being designed around that peak abliity set because devs target a lower spec, whether consoles or just a lower, more mainstream PC spec. It takes some time from a new feature to actually become a design consideration in a game rather than just using the more power to drive the existing older methods at higher resolutions.

And again, I'm not specoifically arguing about games being 'held back' - I'm wondering what tech level is required in PC to match consoles in functionality and when that's due. I'd rather learn about specifics. eg. How important is GPU decompression? What GPUs are capable of supporting that already and/or will there need to be a period of upgrading before a signfiicant portion of PCs can use that feature? Until then I'm sure games will be designed to support all platforms but there'll also come a time, maybe 2035 say, when the minimum spec is 'PS5 level tech' and games won't scale down lower than that. So given that point of feature parity, what are the milestones PC needs to adopt to reach that point?
 
With a high performance NVMe SSD and the proper drivers, Windows 11

The Win11 requirement was dropped. It works on W10 now as well, but requires DX12 for the full effect.

PCI3.0 drives look to be largely off the table, at least I couldn't find any that met the requirements!?!

What requirements? There are no requirements other an NMVe to take full advantage from the drives perspective. But older drives are still compatible.

There will be overheads with PC because of the need to swish data back and forth between main RAM and VRAM to achieve that is a bit easier on decompression built into I/O which benefits particularly UMA but isn't specific to UMA, it just requires a read-RAM-decompress-write block that isn't accommodated on PC currently. But to could be really easily if Intel/AMD/Nvidia could agree a common implementation - which obviously won't happen for years.

Data doesn't go back and forth between RAM and VRAM. In the GPU decompression scenario it goes to RAM where the CPU decompresses the CPU data, and any GPU data goes on from there to the GPU where the GPU decompresses it's data.

AMD's Smart Access Storage (and probably RTX-IO) take this further by cutting system memory out of the GPU path altogether.

1660771237165.png

1660771249588.png

1660771764150.png

You may both be correct, but the use of "high performance NVMe SSD" seems to be differentiating between PCI3.0 and older PCI4.0 drives. Otherwise, why not just say PCI3.0 or NVMe? It feels like Microsoft are consciously differentiating. Why undersell, or over specify the requirements, of a new API?

All NVMe are "high performance" relatively speaking. A 2.5GB/s NVMe drive (i.e. PCIe 3.0 speeds) gets you into "optimised territory" for Direct Storage.

 
The next 4/5 years are going to be crazy indeed, with those quality ports coming in from Nixxess, technically the best versions.
That's a lot of PC trumpeting and not a lot of technical discussion which misses entirely the debate point.
I think theres no reason for concern.
There is no concern. There's a technical question, or rather two. 1) What are the roadblocks to console-like IO performance reaching PC? and 2) What is market penetration goign to be like until that market becomes significant for devs not to factor in weaker IO performance in their game development decisions?
PC tech still isnt 'lagging behind', technically IO speeds are far beyond what the PS5 can deliver.
You've picked a single metric to compare. The other points of the console IOs are lower access times and latencies. To date, no game exists that maxes PS5's IO. Possibly no game ever will because it'll marginalise itself; it wouldn't be the first time a gaming platform had a feature that went mostly unused if so! The question is, again, what will it take to bring PC up to speed all round, and not what workarounds or other solutions are there. There's a whole other thread debating the console storage solutions. This is just a little subset of the discussion from the Nixxes interview and hopes there'll be some insight, extended to a question about the current rate of progress for DS and what it's aiming to accomplish and how long it might take to get there.
 
Indeed, its the other way around. Even though i doubt we are going to see games pulling 14gb/s or more from an ssd, the capability is there but held back by consoles.
Obviously there's more to it than just throughput... PS5 is a lot more efficient in how that data comes off the drive and is processed and put directly into memory which is shared across CPU and GPU... there's absolutely no getting around that.

PS5 and their reliance on this ultra efficient I/O block and SSD bandwidth is born from a lack of system RAM. On PC you don't lack RAM. Games can also scale to take less RAM for PCs with lower amounts of RAM. It's like a bus ride. Yes the trip takes a bit longer, there's a couple extra stops along the way, but your bus holds more than 2x the amount of people.. so instead of making multiple trips you might only have to make one.
 
I don’t think high performance ssd is required but recommended for best results.
Here is the link from forspoken dev presentation and they managed to run ds on old ssd and even hdd

What constitutes 'running DS'? The idea is to get a minimum performance target. Accessing the DS API to load data from an HDD isn't going to provide the low latency, high speed access of the console IOs. We need to set a clearer goal not of 'will a PC running DS play games also on consoles (it will)' but 'what are the technical hurdles to match console IO as a minimum in their completeness' where we assume the console designers had good reason to want to IO features they chose and they didn't overinvests complete overkill where a cheaper, simpler solution would have been just as serviceable. ;)
 
What constitutes 'running DS'? The idea is to get a minimum performance target. Accessing the DS API to load data from an HDD isn't going to provide the low latency, high speed access of the console IOs. We need to set a clearer goal not of 'will a PC running DS play games also on consoles (it will)' but 'what are the technical hurdles to match console IO as a minimum in their completeness' where we assume the console designers had good reason to want to IO features they chose and they didn't overinvests complete overkill where a cheaper, simpler solution would have been just as serviceable. ;)
We know the reason their designers did what they did... because RAM is costly, and it's not feasible to put 32GB or 64GB of RAM in a product that sells for $499.
 
That's a lot of PC trumpeting and not a lot of technical discussion which misses entirely the debate point.

Look at what i reacted to with that.

1) What are the roadblocks to console-like IO performance reaching PC?

There arent really many roadblocks, the hardware is there since aswhile now, deployment is/should be good too, DS is there in its early stages. Whats not there is the games taking true advantage of such systems. And i doubt its the PC as a platform to blame for.

2) What is market penetration goign to be like until that market becomes significant for devs not to factor in weaker IO performance in their game development decisions?

There is going to be weaker io performance, aswell as more capable io performance. Developers can set min requirement specs just like they do with Spiderman pc. I assume people with 2010 sata ssd hardware can play it although compromised. Mechanical hdd's and gaming dont really belong in this discussion i think. I seriously doubt pc gamers playing such games are stuffed with those. The ssd predates the last generation of consoles....

You've picked a single metric to compare. The other points of the console IOs are lower access times and latencies. To date, no game exists that maxes PS5's IO. Possibly no game ever will because it'll marginalise itself; it wouldn't be the first time a gaming platform had a feature that went mostly unused if so! The question is, again, what will it take to bring PC up to speed all round, and not what workarounds or other solutions are there. There's a whole other thread debating the console storage solutions. This is just a little subset of the discussion from the Nixxes interview and hopes there'll be some insight, extended to a question about the current rate of progress for DS and what it's aiming to accomplish and how long it might take to get there.

Raw speeds probably are the least important factor indeed. Even ordinary sata III ssd's comply quite well in these other areas (seek times, latency etc), probably good enough to meet min requirements and scale up from there to m2 kind of hardware.

Obviously there's more to it than just throughput... PS5 is a lot more efficient in how that data comes off the drive and is processed and put directly into memory which is shared across CPU and GPU... there's absolutely no getting around that.

PS5 and their reliance on this ultra efficient I/O block and SSD bandwidth is born from a lack of system RAM. On PC you don't lack RAM. Games can also scale to take less RAM for PCs with lower amounts of RAM. It's like a bus ride. Yes the trip takes a bit longer, there's a couple extra stops along the way, but your bus holds more than 2x the amount of people.. so instead of making multiple trips you might only have to make one.

And theres where direct storage/RTX IO/AMD's variant come into play right, on modern hardware that should be close enough to what consoles do efficiency wise. I just doubt that games are going to be held back due to the PC as a gaming platform for Sony going forward.
 
I'm curious where you've seen that?

I know that with the current CPU based decompression, Direct Storage hands off the decompression task to the application which allows the application (game) to use whatever decompression algorithm is wishes on the CPU before handing back to the API.

I've seen no details at all of how that will work for GPU based decompression though. There would need to be a common GPU compression format for games to target (i.e. a single game can't use an AMD and NV compression scheme simultaneously). That's unlikely to be supplied by either AMD or Nvidia (or Intel) and so either GPU's would have to support some mix of 3rd party formats (much as CPU's do now - but I'm not sure any 3rd party GPU formats exist) with the application choosing which it wishes to use like the current CPU decompression model, or Microsoft would provide the format as part of the API. I don't see why they would need separate API's from the GPU vendors to handle this part of the work. That would largely defeat the point of having a common storage API in the first place.



Actually this is exactly what SmartAccess Storage is designed to address in the PC (bypassing system RAM for a direct path between SSD and GPU memory), and I assume RTX-IO too, although the marketing is more vague around that.
Your post brings up a good point. It might be why RTXIO and SAS's decompression is late to the game. MS probably doesn't want to touch the decompression side as choosing a decompression scheme or rolling their own would probably warrant an antitrust investigation. As DS has the capacity to relatively monopolize decompression for PC gaming.

Alternatively, Nvidia and AMD may be working with tool makers like RAD Games (I guess its Epic now) to convert and produce a GPU version of their decompression schemes. Or give tool makers the time to work out and decide which freely available non patent gpu schemes they want to support with their tools.
 
Last edited:
Is Valhalla designed to max out the consoles IO, or was it designed to support PC IO and then just uses the consoles for faster loading?

Doubtless the latter, but my point was simply to demonstrate a game that requires an SSD at higher settings but can get along with an HDD (and very weak CPU/GPU) at the lowest settings as an example of why consoles don't need to be held back by slower PC's thanks to scaling.

It's equivalent to PC GPU;s being way, way more capable than consoles in shader abilities but games never being designed around that peak abliity set because devs target a lower spec, whether consoles or just a lower, more mainstream PC spec.

I'm not sure it's quite the same because PC's scale whereas consoles don't. It should be quite feasible to design a game around the PS5's IO capabilities and then scale that back to lower end PC's. Whether a dev feels it's worth the effort rather than just targeting a lower performance point is another matter though.

And again, I'm not specoifically arguing about games being 'held back' - I'm wondering what tech level is required in PC to match consoles in functionality and when that's due.

I think the answer to that (from the functionality point of view) is when the next iteration of Direct Storage lands with GPU based decompression - whenever that may be!

I'd rather learn about specifics. eg. How important is GPU decompression?

I'd say important for loading, but not that important for streaming where the data transfer rates are pretty modest. Doubtless it can help reduce load times though since the CPU is usually going to be stretched on other activities at those points while the decompression workloads will be simultaneously high.

What GPUs are capable of supporting that already and/or will there need to be a period of upgrading before a signfiicant portion of PCs can use that feature?

Any DX12 GPU supports Direct Storage.

Until then I'm sure games will be designed to support all platforms but there'll also come a time, maybe 2035 say, when the minimum spec is 'PS5 level tech' and games won't scale down lower than that. So given that point of feature parity, what are the milestones PC needs to adopt to reach that point?

Of the 3 main requirements (outside of the full implementation of Direct Storage itself) we're already at ~94% on two of them. The last one - an NVMe drive, or specifically an NVMe drive of 5.5GB/s or higher if we're talking about a 100% match for PS5 will probably take some time to propagate through the majority of the PC gaming market. But my point here though is that isn't needed before games can target PS5 level performance on PC because those games can scale down. PS5 level IO performance should be doable on PC today in a Direct Storage enabled game (with a big CPU to make up for the lack of decompression hardware), but the fraction of PC's able to achieve that level will be currently pretty small.
 
Your post brings up a good point. It might be why RTXIO and SAS's decompression is late to the game. MS probably doesn't want to touch the decompression side as choosing a decompression scheme or rolling their own would probably warrant an antitrust investigation. As DS has the capacity to relatively monopolize decompression for PC gaming.

Alternatively, Nvidia and AMD may be working with tool makers like RAD Games (I guess its Epic now) to convert and produce a GPU version of their decompression schemes. Or giving tool makers the time to work out and decide which freely available non patent gpu schemes they want to support with their tools.

Yes agreed this is a definite possibility, especially given AMD's wording around this of: compression algorithms that Direct Storage "promotes, supports, endorses and asks ISV's to design to". As opposed to simply "supplies/provides/includes".
 
Or the PC games can use Oodle/Kraken and run the decompression on the GPU just like some games even on the PS4/XOne have done. This is already provided by RAD Tools.

The only requirement might be a license, but that might be covered if using UE. I didn't see what Epic did after acquiring them.
 
Yes agreed this is a definite possibility, especially given AMD's wording around this of: compression algorithms that Direct Storage "promotes, supports, endorses and asks ISV's to design to". As opposed to simply "supplies/provides/includes".
Where's this from?

nvm, I went back a page and found your post. lol
 
Last edited:
Obviously there's more to it than just throughput... PS5 is a lot more efficient in how that data comes off the drive and is processed and put directly into memory which is shared across CPU and GPU... there's absolutely no getting around that.

AMD Smart Access Storage (and possibly RTX-IO) seem to be largely addressing this. Under SAS for example data should be transferred direct from SDD to either CPU or GPU memory (depending on which processor it's needed by) where it's decompressed in real time ready for immediate use by the relevant processor. The only downside is that CPU consumed data still consumes CPU cycles on decompression, but this is by far the smaller proportion of data and so shouldn't be a big overhead. Some may be tempted so think that the CPU or GPU decompression stage represents an extra step over the console where this is done in hardware. But the hardware decompressor will still add latency to the transfer. Sure it can decompress the stream in real time, but so can a GPU - and probably a CPU when relieved of the burden of the GPU data.

Where's this from?

 
AMD Smart Access Storage (and possibly RTX-IO) seem to be largely addressing this. Under SAS for example data should be transferred direct from SDD to either CPU or GPU memory (depending on which processor it's needed by) where it's decompressed in real time ready for immediate use by the relevant processor. The only downside is that CPU consumed data still consumes CPU cycles on decompression, but this is by far the smaller proportion of data and so shouldn't be a big overhead. Some may be tempted so think that the CPU or GPU decompression stage represents an extra step over the console where this is done in hardware. But the hardware decompressor will still add latency to the transfer. Sure it can decompress the stream in real time, but so can a GPU - and probably a CPU when relieved of the burden of the GPU data.



ok I watched the video.

I think it's pretty clear what's happening. DirectStorage will have it's own supported compression format, and will ask developers to design to that spec.

It's a typical MS API in that sense.

RTX I/O and SmartAccessStorage is basically AMD and Nvidia taking advantage of their unique hardware capabilities, to streamline that pipeline even further by adhering to MS' spec, but doing what they can through their hardware architectures to make it more efficient.
 
ok I watched the video.

I think it's pretty clear what's happening. DirectStorage will have it's own supported compression format, and will ask developers to design to that spec.

It's a typical MS API in that sense.

RTX I/O and SmartAccessStorage is basically AMD and Nvidia taking advantage of their unique hardware capabilities, to streamline that pipeline even further by adhering to MS' spec, but doing what they can through their hardware architectures to make it more efficient.

Don't forget though that both emphasise the direct path between storage and GPU, I.e. cutting out the system memory copy which Direct Storage alone does require. Neither is completely explicit in this, but SAS in particular is hard to interpret any other way.
 
Don't forget though that both emphasise the direct path between storage and GPU, I.e. cutting out the system memory copy which Direct Storage alone does require. Neither is completely explicit in this, but SAS in particular is hard to interpret any other way.
CURRENTLY requires... because the GPU vendors aren't ready yet. The "future DirectStorage" slide shows that it won't always require that system memory copy step.

1660771249588-png.6768
 
Or the PC games can use Oodle/Kraken and run the decompression on the GPU just like some games even on the PS4/XOne have done. This is already provided by RAD Tools.

The only requirement might be a license, but that might be covered if using UE. I didn't see what Epic did after acquiring them.

Yep, Doom Eternal uses Oodle but not sure if it uses GPU decompression, although with how fast the game loads (Compared to other similar games) it might??
 
CURRENTLY requires... because the GPU vendors aren't ready yet. The "future DirectStorage" slide shows that it won't always require that system memory copy step.

1660771249588-png.6768
That chart doesn't really indicate that. This is basically just talking about how they plan on doing GPU decompression. I've seen nothing to suggest that DS will ever be removing the copy to system RAM step, though.
 
Back
Top