Current Generation Hardware Speculation with a Technical Spin [post GDC 2020] [XBSX, PS5]

Status
Not open for further replies.
I/O should never be done by the CPU. You may need a CPU (an application) or other hardware device to initiate a transfer but ideally you don't want the CPU involved again until the CPU needs to utilising some of the data. Back when 8-bit CPUs were cutting edge the CPU mostly handled I/O, literally being interrupted every byte or series to bytes to move data from a buffer-space into RAM. I/O should not work like that any more. DirectX API calls are not strictly related to I/O though so maybe you're referring to something else!

PS5 should be very efficient at this. There is 6 level's of priority. CPU can setup what to fetch hopefully well in advance assuming engine is well written. The dedicated io controller and dma controllers will handle the actual loading/decompression/coherency/cache line flushes/... On top of it feels like ps5 has a monster of io controller and has excellent hw support outside cpu to do massive amount of io. CPU load should not be that much compared to something that had more naive hardware.

Having those requirements for game to load very fast, instant resume etc. will force minimum bar on developers.
 
I/O should never be done by the CPU. You may need a CPU (an application) or other hardware device to initiate a transfer but ideally you don't want the CPU involved again until the CPU needs to utilising some of the data. Back when 8-bit CPUs were cutting edge the CPU mostly handled I/O, literally being interrupted every byte or series to bytes to move data from a buffer-space into RAM. I/O should not work like that any more. DirectX API calls are not strictly related to I/O though so maybe you're referring to something else!
well then, this would be false, because all of that is being done on your CPU right now if you're running windows. Without direct storage, you're eating some huge penalties on I/O.
Somewhere in there the call to disk usage is unrolling into something large with massive overhead.
 
It is a little interesting that some of the dynamics of the PS4/XB1 are a little reversed in a way.
The PS5 is now more "balanced" by getting rid of inefficiencies and using helper modules (Tempest vs SHAPE) and I recall cache coherency being big thing for the XB1 as well.
The Tempest being something that could possibly be unused like the XB1 Kinect maps a tiny bit but an extra CU and development costs vs a 100 dollar peripheral is hardly equivalent.
You could suggest that the increase in the PS5 GPU speed matches the XB1 CPU advantage. Of course the XBSX still has a CPU frequency advantage.
I don't consider the PS5 SSD to be like the ESRAM although it is tempting to see it that way with a bespoke highly parallel I/O system that could kind of look like the ESRAM solution very superficially. The SSD thing was there from the beginning I would think.
The variable clock is a slightly better match for a ESRAM candidate but I bet that was something that Sony was interested in early on as well.
On the other side the XBSX has lots of CUs and lots of bandwidth and even a sort of Onion/Garlic Bus thing going on with the 2 memory speeds.
Still this time around there is a lot more similar than different and either console will make the games look great until folks start hankering for ...ugh... 8K.
 
mmm, IO logic runs in cpu...not hardware block. But decompression too?.
Something tells me that XSX is all stock RDNA2 but they added fancy names to software solutions all around: sfs, velocity...
As @More_law_is_dead says they even stole the thunder of virtual memory from Sony being the first to announce it, when with the io and flash controller they have wont be able to move data to gpu fast enough.
The problem here is the lack of a good communication marketing from Sony. They have to clearly sell their advantages. It doesnt cost much to say a Naughty Dog ICE team developer to write a blog article today, not tomorrow, like the one from Matt Philips but concrete to PS5 architecture, explaining clearly the architecture advantages...

They just needed to show a demo...
 
Other thing AMD release a useful patent for big reduction of the effect of CPU stealing bandwidth to CPU, The CPU call to RAM will be the priority and collides less with GPU memory call. Great choice because CPU are more sensible to latency and it will let more bandwidth to the GPU. I don't remember the patent I need to search.
 
I think both needed to do that to convince people nvme ssd vs 5400r disks.

Yes but for this one I know some stuff like I told since weeks the PS5 SSD will have compressed speed above 7 gB/s and said since May 2019 on era the patent is the SSD PS5(minus coherency engine). Sony have a really great internal demo but they didn't show it.
 
If the file io is being done on the OS core what does it matter especially if they've got it down to a 10th usuage? As long as latency is ok.

I wouldn't dismiss MS audio capabilities just yet. It's not like they have been negligent all these years, where audio is concerned they've been very pro active imo.
 
Yes but for this one I know some stuff like I told since weeks the PS5 SSD will have compressed speed above 7 gB/s and said since May 2019 on era the patent is the SSD PS5(minus coherency engine). Sony have a really great internal demo but they didn't show it.

Not doubting that, but both havent shown anything. Im sure both have tested the hardware and see what they can do, internal studios etc. Probably not the time for it, yet.
 
If the file io is being done on the OS core what does it matter especially if they've got it down to a 10th usuage? As long as latency is ok.

I wouldn't dismiss MS audio capabilities just yet. It's not like they have been negligent all these years, where audio is concerned they've been very pro active imo.
Because surely from the 2,4 GB GB/s raw that arrives from the SSD without the custom logic to move that data the effective speed when it arrives to the RAM pool and the GPU cache will be way lower than those 2,4. Thats why making streaming from SSD in XSX is just a marketing chexkbox.
 
Because surely from the 2,4 GB GB/s raw that arrives from the SSD without the custom logic to move that data the effective speed when it arrives to the RAM pool and the GPU cache will be way lower than those 2,4. Thats why making streaming from SSD in XSX is just a marketing chexkbox.
It's a bit too early to be making such definitive statements.
We know very little about MS solution and how it will perform.
All we can say is that Sony seems to have implemented a strong all round solution.
well, textures in your back need 5 GB/s (Cerny dixit) "effectives" to be loaded without popping. You should believe that already.
I'm sure MS would say you need 3.6Ghz CPU otherwise you can't mentain 60fps in next gen games.
Or same for bandwidth, TF, etc.
 
Current solutions are based around 20 MB/s. We're getting over 2400 MB/s on SeriesX. If compression rates are typical, then it's getting over 4800 MB/s.

Game textures will be fine.
Current solutions stream the next seconds of gameplay, not the current vistas.
 
And that will always be a valid solution. Next-Gen doesn't suddenly make it an invalid solution.

Games will have more double the memory to use as well, going from 5 GB for base consoles to 13.5 GB.
Yes but in one solution you could have 10 GB of textures in RAM for current view and in the another 7 suppossing 3 GB of textures stramed as the another 3 will be seen in some seconds.
 
Status
Not open for further replies.
Back
Top