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

[PSman1700]

If the XSX SSD solution isn't fast enough for that (previous gen game), then PS5's wont be either for true next gen games. Most likely because Gears 5 isnt tailored for those speeds. Much like most pc games on NVME right now.

 

[iroboto]

If the XSX SSD solution isn't fast enough for that (previous gen game), then PS5's wont be either for true next gen games. Most likely because Gears 5 isnt tailored for those speeds. Much like most pc games on NVME right now.

right forgot the bandwdith numbers being reported are likely a result of multi-threads pulling from the SSD as opposed to a single thread.

 

[anexanhume]

You're not constrained to 10GB, it's just faster to use.

And the whole concern is whether the textures you need are in fast RAM. The whole point of MS’s methodology is to keep textures out of RAM because they likely won’t be needed. To that end, it’s going to help you try and stay constrained to that 10GB to keep your BW metrics up.

 

[Xbat]

PCs don't care, they can always brute force their way out of any storage problem with more RAM, beyond 2020 most above average PCs will have 24GB of RAM or 32GB, problem solved.

I don't like how the term PC is used to somehow describe super high end PC which only a small portion in people actually own.

 

[iroboto]

And the whole concern is whether the textures you need are in fast RAM. The whole point of MS’s methodology is to keep textures out of RAM because they likely won’t be needed. To that end, it’s going to help you try and stay constrained to that 10GB to keep your BW metrics up.

JITT - Just In Time Textures

 

[iroboto]

I don't like how the term PC is used to somehow describe super high end PC which only a small portion in people actually own.

well, it's okay, it doesn't mean it's efficient use of resources or your money.

 

[anexanhume]

JITT - Just In Time Textures

Getting JITTY with it.

 

[PSman1700]

I don't like how the term PC is used to somehow describe super high end PC which only a small portion in people actually own.

That's the nature of PC gaming, consoles are (at launch) the sweet spot in performance/price ratio, which you most likely cant match with a equal pc build, in special the XSX (12+TF GPUs are rather expensive right now). On the other hand, a rather small portion of the total console users will own the latest console hardware in the first year. Hence the reason we are going to see games still being produced for older systems and pc for a while.

 

[PSman1700]

XSX still loading faster then that ultra pc rig?
nvm, dont think he's using a top end nvme, hence the slow loading. XSX should be faster then 99% of SSD out there anyway.

 

[DavidGraham]

It'll be proportionally the same for RT as rasterising. If PS5 is 80% of XBSX when rasterising, it'll be 80% of XBSX on the GPU side when raytracing. And if rendering 80% of the pixels, it'll need cast 80% of the rays.

Xbox has higher memory bandwdith, this should help making the gap wider in RT.

 

[Dictator]

I mean, even in this demo here:

When you see the XSX load, you still see pop-in happening. I'm not sure if this is a result of poor code or what not, I'm open to suggestions, but even with the SSD and the CPU and the GPU; it's still present even if for a split second.

I imagine that is because the game is still utilising texture steaming pool sizes from the old consoles. That would be hard coded I imagine.
BC games would also not be coded to use the New ssd api. That would require a patch.

 

[mrcorbo]

Weren't we hearing similar arguments about bandwidth to flops with Xbox One versus PS4? It seems like more marketing than an actual point.

There may be a point about balance, maybe streaming assets is more important than more flops at a certain point but if that's the case Sony should be presenting some info to support that design philosophy.

No. We weren't hearing similar arguments. What I was saying had nothing to do with bandwidth to flops.

 

[DuckThor Evil]

And the whole concern is whether the textures you need are in fast RAM. The whole point of MS’s methodology is to keep textures out of RAM because they likely won’t be needed. To that end, it’s going to help you try and stay constrained to that 10GB to keep your BW metrics up.

There is 16GBs in total. 2.5GB is for the OS, so the remaining 13.5GB is for games. The OS portion will come out of the 336GBs memory.

Given those numbers I think it's quite logical to assume that the portion of game data that doesn't require the highest bandwidth can and will fill up that 3.5GB (336GBs is probably an overkill for a lot of things...), and in the worst case the data that would see benefit from the highest bandwidth will only spill over very marginally to the slower section of the pool. Personally I think 10GB out of 13.5GB is a very generous portion of fast graphics memory out of the total.

 

[mrcorbo]

The memory speed hardware specs actually seem to be really close, if you average the total bandwidth of the 16GB on Xbox you get 476GB/s vs the 448GB/s on the PS5. So it’s going to be very interesting to learn what the advantages of doing this end up being.

Why would you average them? I don't think that's going to model the bandwidth distribution very accurately.

 

[Shifty Geezer]

Xbox has higher memory bandwdith, this should help making the gap wider in RT.

Again, it's proportionally the same. PS5 has 80% the GPU resources and 80% the RAM BW, giving the same BW per RT unit. XBSX has 46.7 GB per TF, and PS5 has 44 GB per TF. If rendering to a display 80% the resolution, the RT aspect should be nigh identical.

 

[almighty15]

Again, it's proportionally the same. PS5 has 80% the GPU resources and 80% the RAM BW, giving the same BW per RT unit. XBSX has 46.7 GB per TF, and PS5 has 44 GB per TF. If rendering to a display 80% the resolution, the RT aspect should be nigh identical.

Or you could look at it as PS5 has 12.44GB's per CU and Series X has 10.76GB's per CU.

 

[N00b]

Again, it's proportionally the same. PS5 has 80% the GPU resources and 80% the RAM BW, giving the same BW per RT unit. XBSX has 46.7 GB per TF, and PS5 has 44 GB per TF. If rendering to a display 80% the resolution, the RT aspect should be nigh identical.

That's nonsense. PS5 has to divide available memory bandwidth between GPU and CPU, so effective GPU memory bandwidth will be much lower than the theoretical maximum and less deterministic. XBSX GPU will have the fast memory all to itself (unless devs are morons) and CPU will almost always use slower memory. That way it will be much, much easier to actually utilize GPU resources to the max and do heavy data lifting on the CPU at the same time (think BVH for example). I expect the XBSX to run around PS5 in circles when it comes to RTRT performance.

 

[Shifty Geezer]

Or you could look at it as PS5 has 12.44GB's per CU and Series X has 10.76GB's per CU.

As PS5's are clocked higher, per flop should be a better comparison of RT capabilities which should be tied to CU count and clock speed same as the ALUs.

 

[Barrabas]

For game developers, I have an impression that a maximum TF number is in use when they are designing game engines and games?
If so, is there any average percentage of the maximum TF number they design the game for? I guess there must be a bit overhead for demanding scenes in the game or else the frame rate drops to low. Let's say when I play GOW, my PRO is sniffing on the 4.2 TF when I use my heavy Runic blade attack and burning meteors are raining all over the place with enemies flying everywhere. On average game exploring I use say 3 TF? Am I way off here?

 

[anexanhume]

There is 16GBs in total. 2.5GB is for the OS, so the remaining 13.5GB is for games. The OS portion will come out of the 336GBs memory.

Given those numbers I think it's quite logical to assume that the portion of game data that doesn't require the highest bandwidth can and will fill up that 3.5GB (336GBs is probably an overkill for a lot of things...), and in the worst case the data that would see benefit from the highest bandwidth will only spill over very marginally to the slower section of the pool. Personally I think 10GB out of 13.5GB is a very generous portion of fast graphics memory out of the total.

I completely agree, especially with it being monitored by software. Sony also made the comment you only need the next 1 second of textures in memory versus 30 seconds with the old memory topology.