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

Status
Not open for further replies.

Hmm... They did confirm it was 8 core Zen2, so... Could they have some special CPU customizations? Or is the dev just excited about a more powerful CPU? :p

I mean, what would there really be available on the CPU in terms of customizations that would be significant? The best thing would be to add more cache, but there's practical reasons why you wouldn't do that on a console. If anything, I expect it to be a disparity in disk performance. The rumours had PS5 ahead on disk performance since very early on. It looks like Xbox has done a lot of work in terms of decompression hardware, bypassing filesystem latency issues, and coming up with a way to reduce bandwidth consumption by using a virtual memory page table and being access parts of full textures instead of loading the whole thing into memory to sample a small portion. That said, 2.4 GB/s is not near a high-end SSD. There's a lot of room for Sony to have something more like a gen3 or gen4 SSD in the 3.5-5.0 GB/s range. I honestly think whatever the difference in TF is, whether Sony is higher or lower, disk performance will probably be more noticeable to the end user experience. Everything on the GPU side can scale easily with a minimal visual impact.
 
I honestly think whatever the difference in TF is, whether Sony is higher or lower, disk performance will probably be more noticeable to the end user experience. Everything on the GPU side can scale easily with a minimal visual impact.

Yeah. If Sony has some wizardly that loads games in 2-3 seconds and Microsoft is 8-10 seconds, side by side that's huge. But who the hell cares if it's 10 seconds, it's leaving behind the minute-plus load times that are common on today's console (and PC) games that will be the step forward. But yeah, what could be so exciting about the CPU. :???:
 
Yeah. If Sony has some wizardly that loads games in 2-3 seconds and Microsoft is 8-10 seconds, side by side that's huge. But who the hell cares if it's 10 seconds, it's leaving behind the minute-plus load times that are common on today's console (and PC) games that will be the step forward. But yeah, what could be so exciting about the CPU. :???:

Well, I didn't really believe the idea that games would be sampling textures directly form SSD on a per-frame basis, but that's exactly what series X is designed to do. Sony is likely doing the same thing, and if their disk is faster, that could mean a lot. Whatever their limit is they can sample per frame, or per game world chunk, they could have a large advantage. Mind you, I don't think those things fit sequential read patterns, so the numbers we'd be looking at would be smaller than the sequential read numbers, but it could be very significant.
 
That said, 2.4 GB/s is not near a high-end SSD. There's a lot of room for Sony to have something more like a gen3 or gen4 SSD in the 3.5-5.0 GB/s range.
Current SSDs can not do those speeds for long periods without overheating. If Sony lets the drive run at variable speeds, they could quote higher numbers but to get the kind of speed people here are expecting, they will need to spend more on the SSD than the SOC.
 
Current SSDs can not do those speeds for long periods without overheating. If Sony lets the drive run at variable speeds, they could quote higher numbers but to get the kind of speed people here are expecting, they will need to spend more on the SSD than the SOC.

Yes, 2.4 GB/s is I'm assuming a peak number and not a typical or sustained number. Sony having a 3.5 or 5 GB/s SSD would also be peak numbers. I'm sure someone more enlightened than me could inform us of the difference in random read performance of something like a 5 GB/s SSD vs a 2.4 GB/s SSD. I really don't know what the difference would be. My brain doesn't work in IOPS. We'd really have to see how many IOPS each disk is capable of, I think.
 
Yes, 2.4 GB/s is I'm assuming a peak number and not a typical or sustained number. Sony having a 3.5 or 5 GB/s SSD would also be peak numbers. I'm sure someone more enlightened than me could inform us of the difference in random read performance of something like a 5 GB/s SSD vs a 2.4 GB/s SSD. I really don't know what the difference would be. My brain doesn't work in IOPS. We'd really have to see how many IOPS each disk is capable of, I think.
From the Digital Foundry article on the XBSX : "the 2.4GB/s of guaranteed throughput is impressive" What does this wording mean?
 
From the Digital Foundry article on the XBSX : "the 2.4GB/s of guaranteed throughput is impressive" What does this wording mean?
It means something impossible AFAICS. You can't just guarantee a minimum 2.4GB/s on solid state storage. Imagine the sequential CPU cycles and access times needed to transfer e.g. two and a half million 1 byte files per second.
 
It means something impossible AFAICS. You can't just guarantee a minimum 2.4GB/s on solid state storage. Imagine the sequential CPU cycles and access times needed to transfer e.g. two and a half million 1 byte files per second.
Thanks, I found the wording curious.
 
It means something impossible AFAICS. You can't just guarantee a minimum 2.4GB/s on solid state storage. Imagine the sequential CPU cycles and access times needed to transfer e.g. two and a half million 1 byte files per second.
hmm, I think in this example, nvme is pretty good at this, at least better than traditional controllers.
Yea I think nvme supported up to 64K threads for file transfer. If your block size is 4K, you're going to hold a lot of those 1 byte files in a single block. It may not be all that terrible. I've never seen a test/benchmark like that before though. Curious to see how a PC would perform.
 
hmm, I think in this example, nvme is pretty good at this, at least better than traditional controllers.
Yea I think nvme supported up to 64K threads for file transfer. If your block size is 4K, you're going to hold a lot of those 1 byte files in a single block. It may not be all that terrible. I've never seen a test/benchmark like that before though. Curious to see how a PC would perform.
It's still pretty hard to guarantee that kind of throughput in a worst-case scenario, which is what the wording implies.
Besides, it would be much more interesting (marketing wise) to shout out peak numbers to the skies, as a SSD capable of that kind of minimum throughput would need to have stellar sequential throughput numbers.


I would expect they mean sustained throughput, and probably for sequential reads and writes. Probably just a poor choice of words on Digital Foundry's part.
That's what I think as well.
 
From the Digital Foundry article on the XBSX : "the 2.4GB/s of guaranteed throughput is impressive" What does this wording mean?

Many pc nvme ssd's (can) overheat. This leads to throttling and poor performance. Likely this means that the xbox solution is cooled well enough to not have to throttle speed below 2.4GB/s. Another problem on some pc drives is that the performance depends on amount of free space on disk. Likely this means that there is enough hidden spare blocks to keep 2.4GB/s speed up even when disk is close to full.
 
Many pc nvme ssd's (can) overheat. This leads to throttling and poor performance. Likely this means that the xbox solution is cooled well enough to not have to throttle speed below 2.4GB/s. Another problem on some pc drives is that the performance depends on amount of free space on disk. Likely this means that there is enough hidden spare blocks to keep 2.4GB/s speed up even when disk is close to full.

Good points. I still think it's peak sequential read, but I suppose they can say that they guarantee each unit will hit that mark in a sustained fashion when copying large files etc.
 
Another problem on some pc drives is that the performance depends on amount of free space on disk. Likely this means that there is enough hidden spare blocks to keep 2.4GB/s speed up even when disk is close to full.

This is common practice on consoles. Unfortunately it's also common for them to just use drives with the advertised size and cut on the available space, instead of using larger drives and provide the advertised amount as usable space.

It's just an industry-wide practice, so I'm not exactly counting on either console maker doing it any other way.
 
Besides, it would be much more interesting (marketing wise) to shout out peak numbers to the skies, as a SSD capable of that kind of minimum throughput would need to have stellar sequential throughput numbers.
which I would expect would be the normal thing to do with marketing; I can assume that this was made for GDC and thus for developers in which these metrics will matter more to them ?
 
Cloud based work when not being used to run games.

I suppose there can be Azure instances that could fit in the bounds of a Series X system, or perhaps one with a more fully populated GDDR6 allotment and maybe a link to a more substantial storage system.

What are the chances that RDNA 2 has cache growth?

Does 32MB L3 let XSX emulate OG XB1, or do they offer a X1X like compatibility for that?
Zen CCXs can generate 32 bytes of traffic in each direction, so two of them can take an aggregate of 64 bytes in each direction per clock. That would be gated by the speed of the fabric clock, which is 1.6 GHz in a system with PC32000 DDR4, with the official ceiling being at ~1.8 GHz. Splitting the ESRAM space across two L3 caches could barely nudge things over the original Xbox One's read bandwidth, but not the Xbox One S.
This is assuming an ideal level of connectivity between the L3 and GPU, which is not a data point we have.
The recent Renoire disclosures indicate that it has a 2x wider IF fabric, which might allow the CCXs to transfer more over the fabric at more reasonable clocks, making the raw bandwidth possible.

The hierarchy of Zen and the GPU architecture at least right now aren't amenable to this sort of sharing, since the GPU cannot write to the L3 and treating this traffic like regular cached traffic would probably saturate the home agents being used by the CPUs to order their transfers (benchmarks tend to put pure inter-CCX traffic at a significant penalty).
Beyond that, the CPUs would desperately need at least some L3, and using it as ESRAM would lock them out entirely.

It means something impossible AFAICS. You can't just guarantee a minimum 2.4GB/s on solid state storage. Imagine the sequential CPU cycles and access times needed to transfer e.g. two and a half million 1 byte files per second.
The architecture would set some guidelines for what would be offered. At a minimum, cache line fills are 64 bytes and DRAM transfers require at least 32 bytes.
The virtual memory system is going to have a 4KB minimum granularity, some forms of GPU paging go to 64KB, and the file system may only provide a certain level of reasonable file granularity.
The promised level of performance may be within those bounds.

One possible interpretation to the guarantee is that for an access pattern that gets 2.4 GB/s or more on a fresh and empty SSD will be guaranteed to be readable at that speed and within reasonable latency bounds even if the drive is full, undergoing garbage collection, or the data is old. Many SSDs have issues with performance degradation in those scenarios, and it's an area where I'm curious how the console makers intend to handle things if they have lower-level access to the drives.
Drive makers have easily messed this up in the past, and there's a risk for either Microsoft or Sony messing things up the more they customize the lower levels of the hierarchy.
 
From the Digital Foundry article on the XBSX : "the 2.4GB/s of guaranteed throughput is impressive" What does this wording mean?

The guiding principles that MS went by when starting the design of Project Scarlett back in 2016 were.
  • At a minimum doubling everything in the XBO-X, specifically the GPU. Hence the 12 Tflop target. This plays into the next point.
  • Guaranteed minimum performance.
    • Hence why MS are keen to stress that all clocks are locked.
That 2.4 GB/s is the minimum guaranteed performance of the SSD.

It means something impossible AFAICS. You can't just guarantee a minimum 2.4GB/s on solid state storage. Imagine the sequential CPU cycles and access times needed to transfer e.g. two and a half million 1 byte files per second.

Normally that is true. However, as stated in the DF article, MS have gone to great pains to ensure locked (guaranteed) performance across the breadth of the XBSX design. This includes the SSD.

They've completely redone the storage interface API (DirectStorage). They've included custom logic in the SOC to deal specifically with the high speed requirements (IO and decompression) that they were going for. According to MS, without that custom logic achieving a guaranteed 2.4 GB/s raw transfer rate/4.8 GB/s compressed transfer rate/6+ GB/s including decompression of data would have required 5 of the 8 Zen 2 cores in XBSX.

Unlike current PC NVME drives, all of the logic appears to be either hardcoded in the SOC itself or in the software API layer. This allows them to bypass 3rd party control of transfer rates.

As well, the fact that the external custom NVME drive is using a solid metallic case implies that cooling of the NAND chips is a top priority to ensure that high thermals don't reduce the capability of the chips themselves to maintain a minimum level of performance. Again, something that current PC NVME drives suffer from which MS are addressing.

I wouldn't blame anyone being skeptical until the hardware is actually in customer's hands. However, this isn't information coming from PR which can inflate or misrepresent information. This is information coming directly from Xbox's lead architect.

Regards,
SB
 
Status
Not open for further replies.
Back
Top