This is basically Nvidia adapting the tech MS is introducing with XSX - so AMD will be fine.
-
Ever wanted an RSS feed of all your favorite gaming news sites? Go check out our new Gaming Headlines feed! Read more about it here.
-
We have made minor adjustments to how the search bar works on ResetEra. You can read about the changes here.
NVIDIA announces RTX I/O - PCIe 4.0, on-GPU storage decompression, collaberating with Microsoft on DirectStorage
- Thread starter Earvin Infinity
- Start date
True, there's also that to factor in for comparisons
That's the point of the discussion/comparison, to figure out where or which implementation excels cause we know there are multiple bottlenecks (SSD speed, Priority levels, coherency, decompression e.t.c).
We have a lot of info (asides real world games benchmark/usage) on sony's custom implementation (The patent and cerny breakdown) to be able to make some form of informed guesses, you are right that we don't have that much info on nvidia.
PC was always going to get the top tier stuff at the end of the day, this was no debate.
Even cerny mentioned SSD drives topping the PS5 towards the end of the year and we have samsung drives doing that now.
The implementation of said solution is the interesting part, kinda like how AMD implements RT 'cores' compared to NVIDIA.
PS5 technology *is* revolutionary. Not purely for the technology, but simply because they considered the data transfer end to end. That has been ignored on PC for years already. PCs have fast SSDs. And most of the time you can barely tell the difference from a regular SATA SSD in games because of the various filesystem bottlenecks that get in the way.
MS intends to improve that iwth Direct Storage - hopefully it'll work well and any improvement is welcome. Nvidia has further plans to bring some of the decompression onboard the GPU so it'll bypass some of the back and forth with CPU/system ram which will help even more.
Its a great solution. But is sorely needed and its a little surprising it has taken so long. It doesn't undermine Sony's comments which were last year, and we're only now hearing about directstorage and Nvidia's plans 18 months later. Both are good. Both will be needed to get traction with multiplatform developers (you need addressable market to drive a baseline performance target)
It is going to be very hard for PC to match PS5 this decade as it does not have unified memory and will be difficult to universally implement the same level of customisations so all developers can rely on such a high level of performance. Given Sony also bagged the bleeding edge of compression technology too just in time for all developers on its platform, it is unlikely PC will be able to match PS5 regarding asset streaming capabilities from storage. It could work around that somewhat by relying on more system memory though. In terms of load times, it might be able to get close once we start to see faster CPUs, GPUs and storage solutions such as Sony's ReRam (PCIe5).
Looks like this is still a while away so load times will look embarrassing compared to PS5 for a while on PC but this will start to get it to towards a more respectable speed once Microsoft finally update Windows next year with DirectStorage and games start to take advantage of it.
Even assuming a dedicated decompression block in GPU this might require additional VRAM usage.
(NVMe: NAND -> opt. SDRAM) -> (GPU: VRAM)
(GPU: VRAM) -decompress-> (GPU: VRAM)
PS5
("NVMe": NAND) -> (IOCPLX: SRAM)
(IOCPLX: SRAM) -decompress-> (GPU/CPU: VRAM)
The optional SDRAM inside is used to buffer data, hold allocation tables. Cheaper drives can use system memory instead. PS5 is a hybrid. From NVMe point of view it uses system memory, but it is really just the IOCPLX memory. In the next step it gives the IOCPLX the opportunity to on the fly decompress data going to system VRAM.
(NVMe: NAND -> opt. SDRAM) -> (GPU: VRAM)
(GPU: VRAM) -decompress-> (GPU: VRAM)
PS5
("NVMe": NAND) -> (IOCPLX: SRAM)
(IOCPLX: SRAM) -decompress-> (GPU/CPU: VRAM)
The optional SDRAM inside is used to buffer data, hold allocation tables. Cheaper drives can use system memory instead. PS5 is a hybrid. From NVMe point of view it uses system memory, but it is really just the IOCPLX memory. In the next step it gives the IOCPLX the opportunity to on the fly decompress data going to system VRAM.
Kraken - or other RAD compression technologies - aren't exclusive to Sony. Sony has covered the license for RAD's tech for all PS5 devs, but devs can license it to use it in their PC games. It's a good question as to what algorithms RTX IO and the like will support, but I'd be surprised if they don't offer acceleration for Kraken decompression and others beyond that.
Faster SSDs paired to a capable GPU will meet and beat PS5's bandwidth, at least as an average over a window of time. Where PS5 might have a longer running advantage might be in consistency of performance in the software stack, in an ability to have a guaranteed worst case performance - which are not insubstantial considerations in informing certain use cases of out-of-RAM data. We need more detail on how DirectStorage and RTX IO will work. From a system design point of view, putting the stack into dedicated hardware is also 'more efficient' than spending generic GPU resources. But the point of a higher end PC is you can just spend an excess of resources in a way you wouldn't dream of on a console.
It will take a bit of time because this is still a bit early on the PC side, DirectStorage will only enter preview next year, RTX IO is still in development etc. But by this time next year I expect we'll be seeing end-to-end compressed bandwidth benchmarks between ssd and gpu - whatever about latency or worst-case requests - that beat PS5 numbers. This isn't to take anything away from PS5 (or XSX for that matter) - it's remarkable these days for consoles to launch with designs that, even momentarily, are doing things that will only be possible on PC later. And that a path is opening for software to dramatically scale up their use of storage IO on PC will also only benefit the consoles - it should accelerate the development of a software/tools ecosystem that will lean into the bet that's been made on IO in the consoles, among multiplats and not just first party exclusives.
I know that, even current-gen consoles benefit from that but only Sony are guaranteeing access to it with serious dedicated hardware support which allows for insane throughput and also standardising the level of SSD speed required for games to function. RTX IO looks to be a generation behind what the PS5 is doing and that's when you use a 7GB/S drive with it. Developers on PC likely be targeting SSDs much slower than that like SATA ones once they move on from HDD support as the baseline on PC.
The best part about his comment is that the new Nvidia cards were already in development before he made this comment. Sure, maybe he was under NDA, but than it's still missleading to suggest what he did.
Yep, get ready to get instantly in games while waiting to transfer data from external to internal SSD, whenever you want to play a game you haven't played in a long time and backed up on the external drive. The elephant in the room.
No, unless you think PS5 doesn't need RAM for the CPU, audio, ... in games and doesn't need RAM reserves for the OS. Let's say PS5 will have 13,5 RAM, because that's the figure the Xbox Series X has for games, then you need to reduce the number with the amount of RAM used by the CPU and so on. So saying the Nvidia GPU has only 10GB VRAM is misleading, when the console won't have more for the GPU, unless the developers save heavily on CPU, audio, ... . And that's not going to happen. 16 GB Unified memory sounds very nice, but the truth is all important components want a piece from the cake.
Nope, VE is a software and hardware solution.
Of course that's planned for years.
hahaha, made my day. A generation ahead? Really?
Nah, let's all come down shall we? I/O gets increasingly important, especially for consoles, because we don't have a huge RAM increase compared to current gen. On PC you could always increase RAM to offset some of the bandwidth requirements for the SSD. Compute performance is still more important.
I think your post is great and I just want to answer to this statement. I don't think it's a surprise, because SSD prices were high and consoles, where lots of gamers play, still had HDD as a baseline. You wouldn't use a NVME as minimum requirements for games anyway and thus developing all the technology and software with more time, when next gen hits, was the right choice imo.
Last edited:
Fantastic fit for the pure shithouse Gaming side is these past weeks.
IIRC, he also claimed both next-gen consoles have an SSD faster than anything in the PC space, but Sabrent released a PCI-E 4-based NVMe drive with a 5GBps read speed at the beginning of the year or so (the XSX's SSD tops out at 4.8GBps). I imagine he stopped being intimately involved with Unreal Engine R&D years ago, so I think he just doesn't keep a particularly close eye on the PC hardware landscape.
Last edited:
So for an idiot like me, what do i exactly need (apart from the 3080 im gonna get) to take advantage on this? Ive read from pèople here (though maybe in undertanded wrong) intel doesnt work with this right now?
I was looking between a i9-10900K and a Ryzen 9 3900X, but i suppose if what i understand is true, its better to get the ryzen in this case right (also cheaper with a similar performance). Ive always gone intel so it makes me a little scared changing (also I cant wait for the new cpus becuase i need my PC to work).
Also what motherboard would be recommended (i dont understand the difference between motherboards)? And I suppose i will need to get a SSD with a pci4.0 connection instead of 3.0 right? Any recommendations of what SSD 4.0 i should buy price/performance wise?
I was looking between a i9-10900K and a Ryzen 9 3900X, but i suppose if what i understand is true, its better to get the ryzen in this case right (also cheaper with a similar performance). Ive always gone intel so it makes me a little scared changing (also I cant wait for the new cpus becuase i need my PC to work).
Also what motherboard would be recommended (i dont understand the difference between motherboards)? And I suppose i will need to get a SSD with a pci4.0 connection instead of 3.0 right? Any recommendations of what SSD 4.0 i should buy price/performance wise?
My point was that 10 GB VRAM might not be enough to act as a buffer to compensate for potentially worse latency, which could reduce the amount of new data you can load per frame. If the latency is good, 10 GB are plenty. But that might require a hardware solution like PS5 has.
I don't think this will be a problem in the short term, because for something like portals (see Ratchet & Clank) latency doesn't seem critical. But it might be a problem for mid-gen / late-gen games when devs do more ambitious stuff.
I'll rephrase :) I think it's disappointing that more hasn't been done already to properly leverage the potential of SSDs for those gamers that can use them. Your performance generally scales as you increase CPU and GPU but SSD is a relatively flat bump with faster initial load but buying faster doesn't really scale in the same way you would hope
while this is a good step forward it's a long way from being standard. If both AMD and Nvidia provided drivers with direct storage decompression - even just using GPU compute -that would be a bigger step towards allowing devs to use it as a min/recommended spec.
if it's limited to RTX then it's still toorestricted
I wonder if we'll see platforms that can support such a system (Intel X with millions of PCIe capacity, for example) shipping motherboards with a built in decompression ASIC etc aimed at games/gamers.
I suppose this is sort of where Optane comes in, but perhaps this could me managed through some driver/MB software.
I suppose this is sort of where Optane comes in, but perhaps this could me managed through some driver/MB software.
Big time.
Actually game specific benchmarks though is gonna be something very new for GPU/desktop reviewers - there is almost nothing that focuses on this. We will have to create our own I imagine with next gen only games.
Yeah after reading the xbox blog posts I realized that MS use HW solution to decompress game assets and SW solution (?) for I/O . One interesting detail is that they also deploy a specific hardware block for this function instead of relying on the GPU like NVIDIA's solution. Does this imply the cost of decompressing assets on RNDA GPU to be excessive to the point it would have a significant performance penalty when playing games?
Also NVIDIA has not made a deep dive into their RTX IO tech, i.e we have no idea which component(s) is responsible for decompression (RT? tensor? shader?) If I would have to guess then perhaps it would be the tensor cores since LZ decoding is mainly interger math but I'm not even sure since RTX IO may support other compression algorithms as well.
For a console, fixed-function hardware just makes a lot of sense. You don't really need the flexibility and the speed and latency will be unmatched while having no effect on your other resources. Even if the performance impact were minimal, it'd still be a preferable approach in a closed ecosystem like this unless the latency was up to par AND the performance overhead was so minimal so as to be negligible.
Even the XSX reads into VRAM. The API required to do that on PC will release after consoles.
Ah, very true. But I'm amazed at what's happening right now and how fast it's happening. If the planets are aligning you could have a PC that beats nextgen consoles on all points before we even know what the consoles will cost.
It seems Direct Storage will maybe need new SSD hardware or some existing SSD able to directly adress VRAM in GPU and maybe new motherboard
It seems Direct Storage will work with new SSD and new motherboard
It seems Direct Storage will work with new SSD and new motherboard
I might be wrong, but I think it's more about the GPU being DX12U compliant.
No some SSD will not be able to use Direct Storage. I was surprised too because direct access to VRAM from storage is not how current motherboad work. Maybe some existing NVMe are compatible or people will need to buy a new SSD.
DirectStorage is coming to PC - DirectX Developer Blog
Earlier this year, Microsoft showed the world how the Xbox Series X, with its portfolio of technology innovations, will introduce a new era of no-compromise gameplay. Alongside the actual console announcements, we unveiled the Xbox Velocity Architecture, a key part of how the Xbox Series X will...
devblogs.microsoft.com
In any case, details are still missing, but yes, we might have SSDs compliant with DS.
Might be a whole console generation or more before we get something close to as good on PC and utilised. Multiple PC hardware generations at least, RTX IO is a great start for sure. PS5 solution is the dream, maybe only exclusive PS5 games will fully make use of it. However like I said before, more RAM on PC should help especially later on in the generation so probably won't be necessary for PC to have a solution that is as good as the one on the PS5. Something that is close in terms of loading speeds should be sufficient for the most part.
I think that is why Sony hasn't really shown off the lightning speed of the PS5 much yet as it will blow people's minds by the sheer speed of it compared to anything else as well as the UI integration and software features which PC will find hard to match during the whole generation. It's going to feel so futuristic.
Anyway I just hope Microsoft can speed DirectStorage release on PC up.
That's why Tim is so slick and why he's better suited for Fox news now. He's another salesman.
I can understand that. Finally we are here and it's exciting to see Xbox, PS and PC pushing I/O to the next level. I wonder what Nintendo will do next. I think their push will be ML to use DLSS to bridge the gap somewhat.
Yep, but please don't tell people this :P
oh man if i need to buy a new SSD and/or Motherboard then it's just not worth getting a 3000 card. i just paid £290 for my motherboard and £385 for my SSD. i'd also need to spend £150-200 for a new PSU. and then £650 for the GPU itself lol.
Honestly if you want the whole package in terms of featureset it's always worth it to wait about a year before the new gen consoles release before just building/buying a whole new computer.
Besides, there's never enough good games at/near launch. It's usually worth it to wait even with consoles.
You say this like the most common GPU out there isn't the 980ti lol
Devs will still need to worry about regular hard drives on PC.
Is there a PCIe 4.0 Nvme that recommended in light of these news? Theoretically what would be the difference between say the new Sabrent 7gbps nvme and their 5gbps?
I know the PS5 favours faster drives but on pc are we likely to see difference between early gen and these new 7gbps drives?
I know the PS5 favours faster drives but on pc are we likely to see difference between early gen and these new 7gbps drives?
I would honestly wait.
If it's true that bypassing the CPU/DRAM is not possible on current motherboards designs, compatible NVMe SSDs and PCIe Gen 4 are the least of your problems.
Current SSDs via PCIe Gen 4/3 x4 are connected to the PCH, while PCIe x16 GPUs are connected directly to the CPU, so in theory bypassing the CPU/DRAM and feed data directly into VRAM is not possible.
Last edited:
If he doesn't keep a close eye on the PC hardware landscape why is he offering his opinion, considering it would be took by most people based on his position in the market as some sort of authority.
I don't have an opinion on the validity of his claims, but Sweeney talks so much shit I take everything he says with a grain of salt.
I feel like Sony put their focus on the wrong tech for PS5, they could have just followed the "standards" created by other players and saved some money. The speed difference won't be a deal breaker anyways.
Could someone help me out with this technology? They are mentioning PCIE 4.0 here and there but is it a prereq. for the benefits of RTX IO? Or it would have advantages with PCIE 3.0 also? If with both then what is the difference and why is PCIE 4.0 mentioned?
Also as far as I know there isn't any Intel CPU with PCIE 4.0 support yet but there are motherboards supporting it. In that case if my CPU does not support PCIE 4.0 but my motherboard and SSD supports it will I still get the PCIE 4.0 benefits of this technology or a PCIE 4.0 compatible CPU is also needed so in case of intel it'll be useful only within 1-3 years, when PCIE 4.0 support will finally be implemented?
Also as far as I know there isn't any Intel CPU with PCIE 4.0 support yet but there are motherboards supporting it. In that case if my CPU does not support PCIE 4.0 but my motherboard and SSD supports it will I still get the PCIE 4.0 benefits of this technology or a PCIE 4.0 compatible CPU is also needed so in case of intel it'll be useful only within 1-3 years, when PCIE 4.0 support will finally be implemented?
The mobo needs to have PCIE switch for it to work as it is consumer implementation of NVIDIA GPUdirect. Mobo is finally going to matter again.
I think this announcement actually strengthens the case for sticking with a cheap $100 1TB 3.5GB/s PCIe 3.0 drive. The hardware decompression will effectively double that to 7GB/s putting it well ahead of the effective 4.8GB/s of the Series X. The Series S/X is the common baseline that games utilising NVME SSDs will be built around not the PS5.
The Series X SSD is only really PCIe 4.0 in name only as it uses just 2 PCIe 4.0 lanes which is essentially the same as the bog standard 4 lane PCIe 3.0 drives and even then it doesn't come close to saturating the practical maximum bandwidth of that bus of 3.5GB/s only having a native read bandwidth of 2.4GB/s.
The PCIE 4.0 SSD hype has meant many have lost sight of the fact that Microsoft have specced a low end NVME drive built to a cost. The rumours around the controller suggest a DRAMless QLC drive and the fact that it is only a 2 lane 2.4GB/s device tend to support that. Its ace up the sleeve was Direct Storage and hardware decompression and it's great to see that translate over to the PC.
This is obviously all theoretical at the moment but I am personally delighted that Nvidia and Microsoft are pushing ahead with bringing this technology to PC.
In consumer PC's the GPU PCIE lanes are typically terminated at the CPU itself, while the SSD lanes are terminated at a secondary controller. To get data from the SSD to the GPU, the CPU needs to receive the data from the SSD, move it to main memory and then read it from main memory to the GPU memory. RTX IO purports to allow the GPU to read directly from the SSD. This is not possible with PCIE3 or PCIE4 that are terminated in the way that is described above as there is no "wiring" between the GPU and SSD.
To achieve this, a Mobo needs to include a PCIE switch that routes PCIE packets between the CPU, SSD and GPU, so you'll need a Mobo with a supported PCIE switch and support for PCIE4 to fully support RTX IO.
Isn't this supported already on all new motherboards with PCIe4?
I think in deciding what hardware you want, it might be less about what you baseline need, and more about how you forecast software scaling aspects of quality of performance - or not - with increasing levels of bandwidth of this kind. If software will scale some or other aspects of 'quality' or performance beyond 7GB/s, you might want something more than the minimum. It's hard to know how software will scale on this in the future, but I think an open-ended roadmap of increasing performance on the PC - as introduced yesterday - certainly won't hurt the chances of devs adopting technology that scales things along that axis.
What standards? There are none for this.