DavidGraham
Veteran
It was a first time launch after a long shutdown.
Playstation 5 [PS5] [Release November 12 2020]
- Thread starter BRiT
- Start date
Ok. Sorry. But I would still do a perf counters check.
Yep that’s almost certain. It does offer a lot of options for creating worlds that persist more though - you could perhaps cut down a tree in an open world and it would still be gone when you get back to that location, or you could have a tree grow back or whatever. But some of that would have been possible earlier and then testing / QA probably made that a nightmare
That's a tiny, tiny write stream.
if you look at the design of MS's external card its made in a way that the metal casing contacts an internal heatsink that gets air flow in the console
https://www.bing.com/images/search?...1,2,6&ccid=nam/4KPQ&simid=146906518330&sim=11
damn, that's a good eye you have there
What if I could deform every mesh and put graffiti on every texture ... [emoji16]. The write stream I agree would probably still be small. The fact that every object needs to be stored only once though with an id associated to it and low latency could allow this to be far more efficient and possible.
Bring on a Dreams that can make massive worlds and only have regional thermos [emoji16]
its easier when you can take it apart and see whats what
edit - Also i don't know what Sony has in terms of nand or controller since i haven't seen the system. I will say however that the speeds Microsoft has chosen for their nand does not typically get super hot like some of the faster nvme pc ssd's. I have an nvme in my wife's computer and one in mine and the one in mine is faster but also gets a good deal hotter. I've since added a heatsink to mine and hers and on mine its hot to the touch while hers is just warm. I may actually add a fan to mine.
I believe that is one reason why MS went with the speed they did , it allows a hot swappable memory card type feature in a small foot print.
I am excited to see how sony deals with BYOSSD and how they plan to keep those cool
Last edited:
Persistence and bashing all manner of things are some of the things I would sure like to see with the new systems. Might even be a differentator in terms of gamplay.
Going back to the most recent Digital Foundry article and video on the PS5's architecture, I had some questions concerning the Tempest engine.
The disclosure reiterated that there is a customized CU that has a DMA-based memory subsystem.
The additional details:
The article states that the engine could consume up to 20 GB/s of bandwidth. However, the video at 19:00 states that Cerny indicated the engine could consume 20% of system bandwidth, which would be a more sizeable 90 GB/s.
I'm not sure which one to go by, though the larger value would make sense to merit a mention.
The CU works at GPU clock speed, which may mean it's in the same area as the GPU.
The CU has ~100 GF of performance, roughly that of Jaguar in the PS4, and it does so by providing 64 FLOPs/cycle.
The CU supports the running of 2 wavefronts, one for the system and one for the game. (edit: there is a patent with some similar ideas to this)
Some items that need reconciling, that could be errors or signs of additional alterations.
The 64-FLOP figure is consistent with Jaguar's 4 MUL and 4 ADD per clock over 8 cores, but not consistent with the usual way FMA is counted as 128 per cycle for GCN/RDNA.
Even at 64, the GPU's clocks would put performance over 140 GFLOPS. It would be around 1.6 GHz that the ~100 GF figure would make sense. If this supposedly works at GPU clock, is this incorrect or an indication of what a baseline clock that the GPU uses as a base?
The 20 GB/s bandwidth figure might be reasonable, though potentially not as illuminating as the larger number. The 20% figure, or 90 GB/s, doesn't mesh with a clock at 2.23GHz. It can make sense for a 64-byte connection running at 1.6 GHz, or 32 bytes in each direction at that speed, consistent with the width of other infinity fabric stops like the CCX.
The disclosure reiterated that there is a customized CU that has a DMA-based memory subsystem.
The additional details:
The article states that the engine could consume up to 20 GB/s of bandwidth. However, the video at 19:00 states that Cerny indicated the engine could consume 20% of system bandwidth, which would be a more sizeable 90 GB/s.
I'm not sure which one to go by, though the larger value would make sense to merit a mention.
The CU works at GPU clock speed, which may mean it's in the same area as the GPU.
The CU has ~100 GF of performance, roughly that of Jaguar in the PS4, and it does so by providing 64 FLOPs/cycle.
The CU supports the running of 2 wavefronts, one for the system and one for the game. (edit: there is a patent with some similar ideas to this)
Some items that need reconciling, that could be errors or signs of additional alterations.
The 64-FLOP figure is consistent with Jaguar's 4 MUL and 4 ADD per clock over 8 cores, but not consistent with the usual way FMA is counted as 128 per cycle for GCN/RDNA.
Even at 64, the GPU's clocks would put performance over 140 GFLOPS. It would be around 1.6 GHz that the ~100 GF figure would make sense. If this supposedly works at GPU clock, is this incorrect or an indication of what a baseline clock that the GPU uses as a base?
The 20 GB/s bandwidth figure might be reasonable, though potentially not as illuminating as the larger number. The 20% figure, or 90 GB/s, doesn't mesh with a clock at 2.23GHz. It can make sense for a 64-byte connection running at 1.6 GHz, or 32 bytes in each direction at that speed, consistent with the width of other infinity fabric stops like the CCX.
anexanhume
Veteran
The fact that 20 occurs in both numbers make me think one was a misinterpretation.
There's a quote in the article that explains it:
If you pay attention during the video, you can also make out a protrusion on the central aluminium chassis that matches the placement of the internal drive.
Maybe he is referring to system memory bandwidth.
Or for better words the audio processor may be attached to a cpu coherent bus.
Last edited:
Helps when it was explicitly mentioned in the DF "teardown" video of the XSX.
No way the Audio DMA is taking anything like 90G/s,
The total DMA's for audio will likely be very small, BUT DMA gets it's efficiency from large transfers.
So unless they have some streaming mechanism setup - which might be likely, i dont know
even 10Gb/s transfer could impact the rest of the system like 30Gb/s of video transfers.
The Tempest engine's CU is fed by a DMA engine, which usually falls under IO coherent or something similar.
The video mentioned that they needed to be careful about allowing the CU to pull too much bandwidth, and at least naively a GCN CU is more than capable of pulling that much.
Some of the figures give for the CU sound like they could come from a stripped down GCN as readily as they could come from RDNA, given how sparse the details are.
I just recall the xbo audio processor that is connected to the cpu coherent bus through an iommu and axi bridge with an audio related dma on it.
Last edited:
Theoretical limit and realistic impact in real-world games? 20 GB/s is a LOT of audio data. Uncompressed audio at 48,000 Hz 16 bit is about 96,000 bits/s. 20 GB/s would be 28,000 concurrent sound streams per second. Tempest can only mix 512 sounds which would be a peak audio data rate of 48 MB/s assuming uncompressed. At that data rate you can go fricking 24 bit 96 kHz and hardly touch the bus. Data usage must be for 3D data or something. I think 20 GB/s is loads.
Using 90 GB/s for audio would leave PS5 with <= 358 GB/s for Graphics and everything else. If they manage to get maximum decompression to always happen that's another 22 GB/s use, which takes you to <= 336 GB/s for Graphics and everything else. That is barely faster or exactly the speed of the "slow" memory section of SeriesX at 336 GB/s that some seem concerned about and significantly less than the "fast" memory section at 560 GB/s.
So I don't think it's going to be 90 GB/s for audio.
So I don't think it's going to be 90 GB/s for audio.
What if you made a VR game that is audio only like a simulation of what it is like to be blind? Or a 3D or even 2D music game?
But seriously it will clearly be up to a dev how much you want to use, but you’re not likely to run into an upper limit quickly that much seems clear.
But seriously it will clearly be up to a dev how much you want to use, but you’re not likely to run into an upper limit quickly that much seems clear.
What if you made a VR game that is audio only like a simulation of what it is like to be blind? Or a 3D or even 2D music game?
But seriously it will clearly be up to a dev how much you want to use, but you’re not likely to run into an upper limit quickly that much seems clear.
I had encounter in half life alyx that reminded me how poorly audio is made in many games. It's a vr game and I saw some flying drones coming to kill alyx. Lost track of them, heard the buzzing sound, turned toward sound and I was so surprised to see the drone exactly where the sound was coming from. Not every game does audio as well as alyx does.
I'm so happy sony is putting effort into audio. It's about time someone takes audio to next level. Of course I'm hyped, might be sony doesn't manage(initially) to produce cerny's vision but I hope they endure and really work on this for long term. Something like gran turismo with super good audio in psvr2, count me in!
Similar threads
- Locked
- Deleted member 7537
- Games Industry
- Sticky