Current Generation Hardware Speculation with a Technical Spin [post GDC 2020] [XBSX, PS5]
- Thread starter Proelite
- Start date
- Status
That's ridiculously trolling. 5.5 GB/s is the bare minimum as the raw speed of the drive for data that doesn't compress at all. A 1.5:1 compression ratio would see a typical 8-9 GB/s , and as data compresses higher, the transfer rate increases. Median transfer rate may be 9 GB/s, with the mean quite possibly going higher than that.
Yes i agree, the SSD/decompressor can reach 17GB/s or more. It's an impressive feat. But what i ment was that PS3's ram (GDDR3?) has a higher chance of achieving such speeds on average. It if didn't, the PS3's ram setup must have been kinda bad even for it's time.
I don't think the typical speeds of the SSD will be as high as 17gb/s on average or sustained. Whereas for PS3's ram (or any comparable ram setup) will.
If the SSD's in the consoles would be (in special PS5) having a typical average bandwith of 20gb/s mark (with compression etc), that would be very impressive indeed.
That's still lower then PS3's ram setup though?
Or maybe Sony did not test games with oodle texture for the road to PS5 video but only with Kraken. This is another possibility.
The SSD would be competing with some pc's higher end DDR3 ram setups (in transfer atleast) if those speeds are going to be achieved. I can see why additional DDR4 ram was omitted in these consoles.
No. The SSD is not as fast as PS4's RAM. It's not as fast as PS3's RAM. If you massage the numbers as much as possible, accept a typo of 'PS4' for 'PS3', and pick the worst case example of the average obtained transfer rates for PS3...
...with 'as fast as PS3' being 15.5 GB/s, the latency of SSD is still an order of magnitude off. It's a nonsense statement that cannot sanely be rationalised. If you really want to argue it's not nonsense and the guy just said the wrong thing, try suggesting he meant Cell Read Speed from Local Memory.
Seriously, no-one should be trying to justify this one. It's plain wrong.
What Cerny actually said is that the IO engine would be able to cope with up to 22GB/s of compressed data, if the data was ever compressed that well.
D
Deleted member 7537
Guest
And 9GB/s is already quite a lot, even if Oodle manages to give them higher bandwidth maybe it's just not needed. At 9GB/s you can stream 150MB per FRAME (60fps). As other have mentioned, the downside is that latency of the SSD is gigantic compared to what RAM can do. But that only means you need to warn the SSD to load assets into memory a few miliseconds before they are needed by the GPU. Is damn impressing.
No. The SSD is not as fast as PS4's RAM. It's not as fast as PS3's RAM. If you massage the numbers as much as possible, accept a typo of 'PS4' for 'PS3', and pick the worst case example of the average obtained transfer rates for PS3...
...with 'as fast as PS3' being 15.5 GB/s, the latency of SSD is still an order of magnitude off. It's a nonsense statement that cannot sanely be rationalised. If you really want to argue it's not nonsense and the guy just said the wrong thing, try suggesting he meant Cell Read Speed from Local Memory.
Seriously, no-one should be trying to justify this one. It's plain wrong.
At least "speed " is a bad word maybe if he has said the bandwidth memory is comparable it would be acceptable. I agree.
Last edited:
Recent Discussion on Sustained I/O spun off to it's own topic: https://forum.beyond3d.com/threads/...oughput-loading-of-data-in-games-spawn.61838/
Isn't 5.5 GB/s the peak maximum speed of the drive, because it's the sequential speed, not random reads? So compression ratio will most likely be something like 1.5:1 typically, but raw read speed will be something less than 5.5 GB/s in practice? You're looking at less than 8.25 GB/s usually. It is a 12 channel controller, but sustained random read speed will be less than 5.5 GB/s unless I'm misunderstanding the specs.
D
Deleted member 11852
Guest
I've read nothing to suggest that PS5's SSD read speed is limited only to sequential data. In his Road to PS5 presentation, Mark Cerny referenced the Spider-man technical port-mortem and how data had to be organised in sequentially both for streaming games and as it related patches to games take longer to install because PS4 tries to keep certain data in sequential chunks and this issues were not raised as being issues for PS5, not does it make sense for this to be the case. Sequential reads are about avoiding seeks, which is moving a physical HDD head and waiting for a disc to spin, neither of which exist with solid state storage.
We always describe storage specs in terms of peak ability to provide data. If there's management overhead, that can reduce the average data rate you obtain in use, but that's a moot data-point applicable to all storage tech. In comparing PS4's SSD to XBSXs to a PCIe 4 M2 on PC, the ratio is something like:
5.5 : 2.4 : 4.95
Uncompresed. If you want to scale those figures down by some theoretical 'attained transfer speed', you can do so, but the ration will remain, and quite possible increase for those platforms with leaner, more specialist file IO.
hmm... I think even standard RAM won't do non-sequential as fast either.
Which is why we moved to Structure of Arrays vs Array of Structures if we want higher performance from our memory.
Being able to fill a whole cache line with a pre-fetch is what lets us extract performance. If your data is randomly placed everywhere, you won't be able to maximize the cache lines, and it won't know where to pre-fetch.
I expect SSDs to operate in similar principles.
8ch drives currently do 800k to 1M iops with 4K random read, depending on the controller.
4ch drives are 400k to 500k iops (half the parallelism regardless of nand speed)
PS5 is 12ch so it should be 1.2M to 1.5M iops at 4K
Basically, going with 16K block size would make random read as fast as sequential on any of those drives. As long as the queue is kept filled with requests, and the requests spread evenly across all channels.
4ch drives are 400k to 500k iops (half the parallelism regardless of nand speed)
PS5 is 12ch so it should be 1.2M to 1.5M iops at 4K
Basically, going with 16K block size would make random read as fast as sequential on any of those drives. As long as the queue is kept filled with requests, and the requests spread evenly across all channels.
D
Deleted member 11852
Guest
Memory controllers always read more than the specific data that is being accessed but when it comes to storage, 64kb of data here, or there, or 1mb here or there, or hundreds of megabytes here or there, makes very very little difference.
I don't understand the math here if you guys want to help me out.
I just pulled a random SSD here from a site;
If I look at random read I get 300K IOPs. Do I just multiply 300K by 4K? If so, I'm only getting back about 1200 MB/s
Is this the correct method?
D
Deleted member 11852
Guest
It depends what these figures represent? Are they the end-to-end performance figures of OS, filesystem, I/O controller and SSD? If so, if you change the filesystem or the OS or the controller, do they change?
- Status
Similar threads
- Locked
