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

Considering how much the CPU gets hammered in TLOU and Spiderman I think it would be very beneficial on PC.

Also having it built in to the CPU would mean every piece of software could take advantage of it and not just games.

You keep forgetting that most people buying a Ryzen CPU aren't using it for gaming. With silicon real estate being such a precious commodity it doesn't make sense to include it in a general purpose CPU.

What you should be looking for is if any outside party asks for a semi custom CPU/APU from AMD and includes a similar decompression block. For example, someone like Asus releasing the Z1 based gaming handhelds. Or perhaps Steam. If even they choose not to include something like that, then it's because they don't believe the benefit outweights the cost of implementation in silicon, IE the silicon area could be better used for other things to accelerate gaming ... like more GPU cores.

Regards,
SB
 
You keep forgetting that most people buying a Ryzen CPU aren't using it for gaming.
Enough are, though. And clearly both AMD and Intel put a fair bit of effort into marketing towards the gaming market because it's not some small niche whatsoever. There's plenty of things in a GPP that aren't 'universally useful' but are still there because it's useful to enough people.

It is not some unreasonable ask if this really is going to be a big issue going forward. And with the I/O die and whatnot getting relegated to cheaper, older processes, it actually makes for a really good opportunity to include extra functionality on there where it's affordable(if there isn't already space for it...).
 
You keep forgetting that most people buying a Ryzen CPU aren't using it for gaming. With silicon real estate being such a precious commodity it doesn't make sense to include it in a general purpose CPU.

What you should be looking for is if any outside party asks for a semi custom CPU/APU from AMD and includes a similar decompression block. For example, someone like Asus releasing the Z1 based gaming handhelds. Or perhaps Steam. If even they choose not to include something like that, then it's because they don't believe the benefit outweights the cost of implementation in silicon, IE the silicon area could be better used for other things to accelerate gaming ... like more GPU cores.

Regards,
SB

I used AMD as they made PS5's SoC and not because it should only be them that implement such a thing.

Obviously it would also need to be included in Intel CPU's.

If on low end CPU's it will be the difference between playable and unplayable frame rates (CPU decompressions vs custom I/O block) then it'll be worth the transistor cost.
 
Considering how much the CPU gets hammered in TLOU and Spiderman I think it would be very beneficial on PC.

Also having it built in to the CPU would mean every piece of software could take advantage of it and not just games.

It's more the practical considerations within the PC eco system that make this solution not make sense. From a raw performance perspective there are definitely advantages to be had, but there's a few practical reasons why it wouldn't work:

  • As a hardware decoder, it would be fixed to a limited set of formats, Kraken and zlib in this case. That means to be useful, devs would have to package their games in those formats. But what incentive would they have to do that (and pay the associated licensing costs for Kraken) if only a tiny proportion of the newest PC's could perform the decompression in hardware? Doing this on the GPU ala the PC DirectStorage approach leverages existing hardware that is present in almost all gaming PC's already to do the same task and can later be transitioned to a dedicated hardware chip on the GPU if needed without anyone having to forgo the hardware assisted decompression while the transition happens (in compute to hardware form)
  • Having the decompression on the CPU side of the graphics PCIe bus means you are sending uncompressed data over that bus to the GPU - wasting bandwidth. The GPU method allows that data to be sent compressed which essentially doubles your effective PCIe bandwidth.
  • As mentioned above, it's a waste of silicon on any CPU that wouldn't be used for gaming, and so inflates the price of Ryzen CPU's vs the competition in segments where it's not advantageous.
 
Neural networks are already used to upsample BC7 textures on God of War Ragnarok for the PS5 version. They use Neural Networks to upsample textures (used in the performance mode) for the 4K mode.

Every frame, the texture streaming system detects textures that require upsampling , and sends them over to the upsampling system.

XxAjw4l.png


30% of the textures (diffused) are upsampled without NN. Normal maps are upsampled with NN. In total ~10GB of textures are produced in 10mn, that's around 42 4k textures per minute. They use FP16 to run the machine learning inference with plenty of optimization on PS5 GPU compared to their initial generic code. Their system needs ~9ms by frame.

The system seems to be dynamic based on ressource (time) availability. The more time they have, the more textures are upsampled from 2K to 4K.
The way we determine how much BC7 blocks to upsample is by tracking running averages of excess frame time and the duration of one evaluation of any specific network.

9CGcQc3.png
 
Last edited:
There were older AMD cards that had something like this for professional workloads. It's interesting to think about what things would look like for use in games using a slower, large storage drive on the GPU for sure.

Probably useful for uatx boards and sfx builds where number of m2 slot is limited.
I suggested this in the past since the price of nvme's have plummeted . However my suggestion was to use it to install a copy of a game or at least the texture assets so that you can bypass the need to move texture datas from an nvme to the cpu to the gpu .
 
Unlike the card in the past, isn't 't this one basically just an PCI-E adaptor for NVME card?
So it's basically a pass through adaptor.
Therefor would still need to go via the cpu, unless Direct Storage could do it I guess. But then it's still just basically an NVME drive attached through PCI-E. Good if you don't have adaptor on motherboard or need another one.
 
Seems we have another game with Direct Storage 1.2

Portal: Prelude RTX


Cool! It's not available to download here yet but I'll absolutely be getting that when it is. Interesting that they mention the reduced game footprint as well. Almost - put not quite - a 2:1 compression ratio for textures.
 
Seems we have another game with Direct Storage 1.2

Portal: Prelude RTX


Well it loads incredibly quickly although I don't really have a baseline to compare it with. I do have Portal RTX installed but that's on my HDD and naturally loads much slower. The game totally punishes my GPU though. pegged at 100% all the time and often not breaking 60fps with DLSS Auto and FG on. That's at my monitors output res of 3840x1600. Plays well enough though regardless.
 
Back
Top