23 January 2020

Analyse This: Next Gen Consoles (Part 5)

Last time I wrote about the possibility of the next gen consoles utilising the latest NVMe PCIe Gen 4 SSD controllers from Phison and SMI. There's an update to this rumour taken from the work profile of an ex-employee of Phison where the person mentions that they worked on the project team that were designing the controller that would be used in the Xbox Series X (labelled as Xbox Scarlett by the engineer). The controller in question is, according to the profile, the PS5019-E19. Unfortunately, this is a bit disappointing.
Looking back at the performance numbers listed by Anandtech, the E19T (which I will assume will have similar, if not identical, metrics to the E19 mentioned in the engineer's profile) is not that much better than the existing PCIe Gen 3 controllers in their lineup. Yes, it consumes less power than the higher end models but there's actually not much "next gen" about this controller other than being on the PCIe Gen 4 compatibility list.

The IOPs are slightly higher than existing Gen 3 drives, as are the sequential reads and writes. However, it's important to note that, for game performance - sequential transfers are not the dominant factor - IOPS are a better indicator of performance. Similarly, performance is not usually a static metric - the longer the drive operates, the performance of the transfers usually begin to decrease - especially with cheaper or lower-end controllers. 

The IOPs of the E19T controller top out at around 440k for reads and 500k for writes. With a 4 KB block size, that gives us a theoretical throughput of around 1.8 GB/s read and 2.0 GB/s for writes. A 14 Gbit/s GDDR6 module is able to operate up to 1.75 GB/s in transfers so I guess Microsoft is safe-ish in their goal of using a portion of the hard drive as virtual RAM but it's a very close number... By comparison, the currently existing E16 reaches around 600k read IOPS and 500k write IOPS on AMD's Gen 4 X570 which gives a little overhead in terms of dropping-off of theoretical performance.

To put this all into context - the Gen 4 SSD won't really make a big difference in install times (disc reads are very slow in comparison and internet speed is usually pretty slow too) and, if output recording is going to be a feature then that will also be writing to the disc at the same time as any virtual RAM capability - along with general loading of assets for the game in question. All those features combined require constant access to the SSD and the bandwidth available and those numbers above are a bit too close for comfort to the available bandwidth of a GDDR6 module (assuming they're using 14 Gbit/s modules and not 12 Gbit/s modules...). This is also assuming a 1 TB drive size - if the Microsoft were to use a 512 GB drive, the IOPS decrease due to reduced parallelism from having fewer NAND chips on the board.

Sure, the recording feature itself is unlikely to have much effect on the avilable bandwidth (1440p @60fps, 127 kbps, 48 kHz could take up to 6 MB/s in bandwidth) but that is constant work being done across the PCIe interface and when combined with the use as virtual RAM and asset streaming could, in certain circumstances, saturate the controller.

Given that the whole array of GDDR6 modules will be operating over, at minimum, a 256 bit interface, we're looking at north of 448 GB/s total bandwidth available on the RAM side of the equation so the SSD would be a relative bottleneck if required for a lot of loading/unloading of assets which would normally reside in memory (e.g. textures or level geometry).

NOT an actual event!!! (I hope!)

Moving onto the implementation in nect gen consoles, I hadn't fully realised the ramifications of switching to NMVe SSDs from the traditional 5400 rpm HDDs. First off, I realised that constant writing to the drives would result in dead drives and shortened lifespans - depending on the recording quality - which could be terrible if these are not user replaceable. It was easy to see the performance delta between the laptop HDDs and the PCIe Gen 3/Gen 4 NVMe drives and the effect that would have on load times. I could see the reduction in power usage from around 5 - 6 W to 1.5 - 3 W but what I hadn't accounted for would be the effect on the thermals of the system.

Like most other people I'd been focussing on the thermal performance of the APUs in the PS5 and Xbox SX but it looks like the Renoir chips are highly performant, despite being at lower wattages. If an R7 4800H can come close to an i9-9980HK in CPU score at 45 W in terms of CPU performance, that bodes well for the next gen consoles. Saying that, I still think the 4800U 15 W part is more indicative of the performance level we'll see in the consoles (and it's not that far behind the 4800H!). That doesn't tell us anything about the wattage required to run the GPUs of course...

However, I am now realising that a major issue for these consoles will be the heat generated by the SSD. Der8aur showed that allowing heat to build up on these devices effectively kills the transfer speeds to 10-20% of theoretical performance and can even result in their crashing and refusing to work at all until a system reboot. 

While in a normal system and with average use, the SSD won't even reach very hot temperatures but constantly shuffling data onto the drive through the recording feature, virtual RAM and asset streaming could really cause a lot of heat generation and thus it seems that, for these compact systems, some sort of heatsink will be required to be installed in order for it to not kill the system. It could also further help explain the design of the SX and hint that the SSD will not be user replaceable as it would be more efficient (in terms of cost) to place it under the combined heat piping of the APU and other board components.

I really hope that Microsoft and SONY have a good solution for this issue that still allows for user replacement of the drives.

There's a certain expectation that SONY will have an initial reveal of the PS5 in February - possibly as early as the 5th so we may not have long to wait for comparisons of the general features of the two consoles.

No comments: