• Hey, guest user. Hope you're enjoying NeoGAF! Have you considered registering for an account? Come join us and add your take to the daily discourse.

NVIDIA announces RTX I/O - PCIe 4.0, on-GPU storage decompression, collaberating with Microsoft on DirectStorage

Thank you Mark Cerny.

Gpu decompression is not new though. What is exactly new with the new cards that hasn't been done before? I would guess it's the better software support.

It only shows decompression using the gpu though (which is not new), but what about data management, check-in, etc.?
 
Last edited:
I knew PC hardware would pull ahead fast, but this......Nvidia just buried the consoles from a performance AND pricing standpoint......
I honestly wasn't expecting DOUBLE the performance in a single generation like that. But pricing was a mental mind fuck. I can't wait to witness the salt mines from our strongest warriors on this site, that always troll PC based threads.



ilyG0UU.jpg
 
Last edited:

jimbojim

Banned
Some math:
PC 7GB/sec -> 14GB/sec = 24 Intel cores
Sony 5.5GB/sec-> 22GB/sec = 9 CPU cores
Pretty accurate difference between the skill of Sony and NV software engineers...

What is the decompression number for Nvidia I/O? Is it over 14GB/s or something i mixed up?
 
Yeah most definitely, go with AMD, they are cheap and powerful, especially now that Zen 3 is closer, Zen2 is getting cheap enough even in high-end.

looking on Newegg, with a Ryzen 7 3700 8 core 3.6GHz CPU and everything that goes into it. it'll run me $1200 or so without the 3080. I can probably cut that price down another $300 or so if I use my old power supply and SSDs. might just wait til Black Friday to see if I can get some good deals first.
 

M1chl

Currently Gif and Meme Champion
looking on Newegg, with a Ryzen 7 3700 8 core 3.6GHz CPU and everything that goes into it. it'll run me $1200 or so without the 3080. I can probably cut that price down another $300 or so if I use my old power supply and SSDs. might just wait til Black Friday to see if I can get some good deals first.
Well something aint right, what is your components. It's same think as my update with 3900x and Asus X570-E in the beginning of this year.
 

onQ123

Member
Note they also mentioned that that there were some similarities in the Raytracing hardware setup to the Xbox.

Actually no he was saying that using the same method that Microsoft used to say that they had 25TF when using ray-tracing they get 69TF of ray-tracing .
 

LordOfChaos

Member
Interesting that we essentially heard him agree with the next gen console vision. CPUs would become a bottleneck for SSDs this fast, a fast gen 4 SSD could use up to 24 cores and would be increasingly important for next gen games, and a solution to these problems.

Their performance range for next gen is also in tune with Cerny's numbers for the PS5, just somewhat over its 5.5GB/s raw at 7GB/s, 9 typical up to 20 max compared to 14GB/s stated compressed. The vision is surprisingly close to what Sony laid out.

20200901172408.jpg
 
Last edited:

ZywyPL

Banned
giphy.gif



So there goes all the console's supposed edge and superiority. Except those cards will actually be able to process all that data in less that 16ms, MUCH less actually it seems. However that 200GB number looks scary, we were promised lower game sizes due to no duplicated data, but by the looks of it, a 1TB will be able to hold just a few next-gen AAA titles at once :/
 

Dr_Salt

Member
So do you positively need pcie4.0 for this to work? Some people are saying it will work with 3.0 also and now im confused.
 

CuNi

Member
Let the information drip feeding begin!

I'm still awaiting word for NVEnc Gen 7 performance and capabilities. That shit must be insane by now in terms of quality to bitrate ratio.
 

Rikkori

Member
Guys, ofc the comparison is misleading re # of CPU cores. It's like talking about RT performance when you run it without RT cores. It would obviously never be done in that way because it's stupid, so it's an irrelevant number given purely for marketing purposes (no different than when they brag about how much faster 2080 is than 1080 Ti... with ray tracing). That being said, they obviously have a smart solution in place and spoiler alert: was possible with much older GPUs as well, because they actually have had this capability for many years in various forms - it just was never necessary because 1) SSDs are already barely used past the level of the shittiest SSD; and 2) even the fastest SSDs weren't that fast that you'd care. And it's something many of us pointed out to all the Sony fanboys when they were crying about how PC wouldn't keep up. On-GPU decompression is NOT NEW!

That being said, relax and enjoy being PCMR ;) 4K@120 HERE WE GOOO! 🎆🎆🎆

WPTRRMM.gif
 
Last edited:

Dr Bass

Member
Collaberate ....

Is that like where you work with someone while yelling at them that they suck?

This is great for the Series X. The more adoption we see for this on the PC side means we'll see more 3rd party games utilize it.

This has absolutely nothing to do with the Series X, and never will.
 
Last edited:

MarkMe2525

Member
Collaberate ....

Is that like where you work with someone while yelling at them that they suck?



This has absolutely nothing to do with the Series X, and never will.
So the fact that nvidia has integrated support of an API that will be used to develop games for PC and series x simultaneously has nothing to do with the series x? Are you sure about that?
 
Last edited:
Well something aint right, what is your components. It's same think as my update with 3900x and Asus X570-E in the beginning of this year.

here's what I picked out from Newegg

AMD Ryzen7 3700X 8 core 3.6GHz
$290
ASRock B550M PRO4 MicroATX motherboard
$120
Corsair DDR4 3600 32GB (2 X 16 GB) ram
$115
DIYPC tower case
$90
EVGA 1000W power supply
$200
Intel 1TB QLC SSD
$115
Noctua CPU 140mm fan
$75
Win10 pro 64
$150
Thermal Paste
$8

these comes to around $1163 or so and like I said, just kinda picked them off the store. and I can probably use my old power supply and my SSDs to cut the cost down to $850 or so. adding the $700 of the 3080, that comes to $1550 or so. any better deals out there? I'm not in a hurry hence the mention of Black Friday.
 

M1chl

Currently Gif and Meme Champion
here's what I picked out from Newegg

AMD Ryzen7 3700X 8 core 3.6GHz
$290
ASRock B550M PRO4 MicroATX motherboard
$120
Corsair DDR4 3600 32GB (2 X 16 GB) ram
$115
DIYPC tower case
$90
EVGA 1000W power supply
$200
Intel 1TB QLC SSD
$115
Noctua CPU 140mm fan
$75
Win10 pro 64
$150
Thermal Paste
$8

these comes to around $1163 or so and like I said, just kinda picked them off the store. and I can probably use my old power supply and my SSDs to cut the cost down to $850 or so. adding the $700 of the 3080, that comes to $1550 or so. any better deals out there? I'm not in a hurry hence the mention of Black Friday.
Alright that WIn 10 buy it here: https://software-codes.com/collections/top-sale/products/windows-10-pro Bam 146USD off.

Everything else is understandable, I though you are buying just mobo+CPU+RAM...
 

RaySoft

Member
Last edited:

ANIMAL1975

Member
Looking forward to the PS5 regulars now accept the industry is following Microsoft on IO not Sony. That the Microsoft engineers were not liars after all that direct storage was not just PR. So games will be designed around direct storage since the industry is adopting it.
Lol of course they have to follow Ms, they use DirectX.
And in what fantasy world Sonys proprietary hardware or software libraries for Playstation, are shared with the pc industry during their development, or even used by them afterwards lol? It's proprietary and customized. They are only for Playstation.
Sony could of worked with amd, intel and nvidia to get their IO the standard. You guys said all the velocity was all pr don't back down now. You guy have 100 pages of trashing it in the official thread. Stand up eat your SSD crow.

Again you don't know what you're talking about, Sony worked with who they needed to work with: the developers, that's the most important input in the matter.
And what came up was a completely hardware based approach, the i/o complex on the SoC. Listen to the PS5 presentation again _ everything is hardware supported and automatic, no special dev time or optimizations needed.
So your so-called standard solution that needs direct storage/velocity architecture (software support) will not be a great advantage.
 

PaintTinJr

Member
I can't tell if I'm interpreting the info correctly, but ...am I right in thinking that RTX IO is a PCIe NIC(network interface card)with a slot for a Gen4 NVMe SSD, that plugs into a new motherboard with PCIe 4.0 bus, and then you need a Nvidia RTX 30 series GPU that also plugs into the PCIe 4.0 bus?
If that is the setup, I assume that the reason PCIe 4 is needed is due to the bi-directional comms between the NIC and GPU will be capable of saturating half the bus at peak transfers, so going up to PCIe 4 leaves enough lanes for a system to still use all the legacy PCIe 3 stuff too without interference.

Even though the numbers look competitive to the XsX XVA and PS5 IO complex, I was surprised that a quick control+F on the nvidia page didn't find the word "latency". From the info provided I would still expect the latency of XVA to be less, maybe 1/2 or a 1/3- because adding a NIC: a switch, and adding external switching increases latency as a general networking rule - and I still expect the IO complex to be 1/5 or less of the latency of the XVA; especially now if this solution is the external design version of the integrated XVA solution.

Ultimately, my main concern for a solution like this, is the number of points of potential trouble shooting when it doesn't perform as expected - and reminds me a little bit of SLI/Crossfire teething troubles and the original 3dfx Voodoo passthru card when you got a blank screen launching a game. A complex multi-component solution like this in the PC space shouts design workaround to me, and it makes me wonder if Nvidia's lack of access to a x86 license to make x86 APUs is now a major problem and a potential for their demise - or acquisition by someone like Intel - in the next 5years.

Even if all those RTX IO numbers work out great, and the latency is on par with the PS5 IO complex - which would be an amazing feat IMO - the RTX IO still looks like a half-way-house solution to me, because the data/memory solution is even more fragmented than before - and questions and potential installation nightmares of if or how the RTX IO SSD can be used as the boot drive for Windows, and if not, what is the gaming performance cost of limitation? - while on the console side, the benefits of unified memory for gaming and the impact on game design is probably about to really take off now that the new consoles have brawny CPUs and heavy weight GPUs to share complex simulations across both types of compute.

But none the less, very interesting new PC tech solution and new exciting times ahead (IMHO).
 

Dr Bass

Member
So the fact that nvidia has integrated support of an API that will be used to develop games for PC and series x simultaneously has nothing to do with the series x? Are you sure about that?

Yes because the Series X won't have the hardware design to utilize the API. And it's an Nvidia thing not AMD ...
 

psorcerer

Banned
I can't tell if I'm interpreting the info correctly, but ...am I right in thinking that RTX IO is a PCIe NIC(network interface card)with a slot for a Gen4 NVMe SSD, that plugs into a new motherboard with PCIe 4.0 bus, and then you need a Nvidia RTX 30 series GPU that also plugs into the PCIe 4.0 bus?
If that is the setup, I assume that the reason PCIe 4 is needed is due to the bi-directional comms between the NIC and GPU will be capable of saturating half the bus at peak transfers, so going up to PCIe 4 leaves enough lanes for a system to still use all the legacy PCIe 3 stuff too without interference.

Even though the numbers look competitive to the XsX XVA and PS5 IO complex, I was surprised that a quick control+F on the nvidia page didn't find the word "latency". From the info provided I would still expect the latency of XVA to be less, maybe 1/2 or a 1/3- because adding a NIC: a switch, and adding external switching increases latency as a general networking rule - and I still expect the IO complex to be 1/5 or less of the latency of the XVA; especially now if this solution is the external design version of the integrated XVA solution.

Ultimately, my main concern for a solution like this, is the number of points of potential trouble shooting when it doesn't perform as expected - and reminds me a little bit of SLI/Crossfire teething troubles and the original 3dfx Voodoo passthru card when you got a blank screen launching a game. A complex multi-component solution like this in the PC space shouts design workaround to me, and it makes me wonder if Nvidia's lack of access to a x86 license to make x86 APUs is now a major problem and a potential for their demise - or acquisition by someone like Intel - in the next 5years.

Even if all those RTX IO numbers work out great, and the latency is on par with the PS5 IO complex - which would be an amazing feat IMO - the RTX IO still looks like a half-way-house solution to me, because the data/memory solution is even more fragmented than before - and questions and potential installation nightmares of if or how the RTX IO SSD can be used as the boot drive for Windows, and if not, what is the gaming performance cost of limitation? - while on the console side, the benefits of unified memory for gaming and the impact on game design is probably about to really take off now that the new consoles have brawny CPUs and heavy weight GPUs to share complex simulations across both types of compute.

But none the less, very interesting new PC tech solution and new exciting times ahead (IMHO).

I'm pretty sure that DirectStorage will use a different proprietary fs just for game installations. Thus separate IO card is not a bad solution. It kinda solves all the problems without introducing too much latency.
Still probably miles away from PS5 but may be pretty close to XBSX (and they may use NIC interface too btw).
 

MarkMe2525

Member
Yes because the Series X won't have the hardware design to utilize the API. And it's an Nvidia thing not AMD ...

I want to give you a chance but are you saying that AMD and the XsX will not support direct storage? Do yourself a favor before replying, type in "directstorage" into your Google machine.
 
Last edited:
giphy.gif



So there goes all the console's supposed edge and superiority. Except those cards will actually be able to process all that data in less that 16ms, MUCH less actually it seems. However that 200GB number looks scary, we were promised lower game sizes due to no duplicated data, but by the looks of it, a 1TB will be able to hold just a few next-gen AAA titles at once :/

Sony I/O solution is still better, my friend. This is however a step in the right direction, and one would suspect a hardware based solution (accelerators on the die, just like what Sony and AMD have come up with) within the next two years or so. Well, I hope so, at least.
 

ratburger

Member
I can't tell if I'm interpreting the info correctly, but ...am I right in thinking that RTX IO is a PCIe NIC(network interface card)with a slot for a Gen4 NVMe SSD, that plugs into a new motherboard with PCIe 4.0 bus, and then you need a Nvidia RTX 30 series GPU that also plugs into the PCIe 4.0 bus?
If that is the setup, I assume that the reason PCIe 4 is needed is due to the bi-directional comms between the NIC and GPU will be capable of saturating half the bus at peak transfers, so going up to PCIe 4 leaves enough lanes for a system to still use all the legacy PCIe 3 stuff too without interference.

Even though the numbers look competitive to the XsX XVA and PS5 IO complex, I was surprised that a quick control+F on the nvidia page didn't find the word "latency". From the info provided I would still expect the latency of XVA to be less, maybe 1/2 or a 1/3- because adding a NIC: a switch, and adding external switching increases latency as a general networking rule - and I still expect the IO complex to be 1/5 or less of the latency of the XVA; especially now if this solution is the external design version of the integrated XVA solution.

Ultimately, my main concern for a solution like this, is the number of points of potential trouble shooting when it doesn't perform as expected - and reminds me a little bit of SLI/Crossfire teething troubles and the original 3dfx Voodoo passthru card when you got a blank screen launching a game. A complex multi-component solution like this in the PC space shouts design workaround to me, and it makes me wonder if Nvidia's lack of access to a x86 license to make x86 APUs is now a major problem and a potential for their demise - or acquisition by someone like Intel - in the next 5years.

Even if all those RTX IO numbers work out great, and the latency is on par with the PS5 IO complex - which would be an amazing feat IMO - the RTX IO still looks like a half-way-house solution to me, because the data/memory solution is even more fragmented than before - and questions and potential installation nightmares of if or how the RTX IO SSD can be used as the boot drive for Windows, and if not, what is the gaming performance cost of limitation? - while on the console side, the benefits of unified memory for gaming and the impact on game design is probably about to really take off now that the new consoles have brawny CPUs and heavy weight GPUs to share complex simulations across both types of compute.

But none the less, very interesting new PC tech solution and new exciting times ahead (IMHO).

I think you're over-thinking this. Any PCIe device can communicate with any other on the PCIe bus. The GPU just needs to know how to properly request data from the SSD since there's a file system involved.

From this official nvidia article:
RTX IO is supported on all GeForce RTX Turing and NVIDIA Ampere-architecture GPUs.
So since it will work on 2000-series cards, it can't be a physical slot on the video card.
 
Last edited:
Some math:
PC 7GB/sec -> 14GB/sec = 24 Intel cores
Sony 5.5GB/sec-> 22GB/sec = 9 CPU cores
Pretty accurate difference between the skill of Sony and NV software engineers...

Jesus you are a FUD machine. You gonna spread anymore FUD about Ampere to Turing TF relative performance metrics? There's no way you say what you said in the other thread with a straight face. Puts a lot of your other technical discussions in serious question now.

im really confused i dont know if you're trolling or praising Sony xD

He's a Sony fanboy, so the latter.
 
Last edited:

psorcerer

Banned
Jesus you are a FUD machine. You gonna spread anymore FUD about Ampere to Turing TF relative performance metrics? There's no way you say what you said in the other thread with a straight face. Puts a lot of your other technical discussions in serious question now.

So far all my "FUD" was on point. On Ampere. I.e. literally all my predictions manifested.
 
Last edited:
So far all my "FUD" was on point. On Ampere. I.e. literally all my predictions manifested.

No, you were dead wrong about the IPC gains. You implied 30 TF Ampere was 15 TF Turing. When Ampere is the newer architecture and has (at least) 10% IPC gain over Turing.

Literally you'd have to use BS "math" or some Common Core level tripe to get "30 TF Ampere = 15 TF Turing", which is what you posted. So your analysis is fraudulent.
 

psorcerer

Banned
No, you were dead wrong about the IPC gains. You implied 30 TF Ampere was 15 TF Turing. When Ampere is the newer architecture and has (at least) 10% IPC gain over Turing.

Today NV presentation never mentioned any 10% increase AFAIR.
But they did mentioned 2x per core shader performance (whatever that can mean).
 
Last edited:
Top Bottom