Question GDDR6 latency on XBOX Series X and PS5

psolord

Golden Member
Sep 16, 2009
1,920
1,194
136
Hello everybody.

I didn't really know where to ask this question, because it could go to cpu, memory and gpu subforums.

In any case, since GDDR6 is mostly part of a gpu, I thought that the gpu subforum would be more appropriate, although the second part of the question, has to do with the cpu.

So anyone knows what kind of latency we are looking at, regarding GDDR6 on the new consoles?

If it is slower than normal RAM in terms of latency, what kind of performance penalty would that have, regarding the cpu stuff? I mean what would be the equivalent of a normal Zen 2 processor, paired with normal RAM on the PC, in terms of latency and performance?

thanks
 

psolord

Golden Member
Sep 16, 2009
1,920
1,194
136
I know they have a unified memory architecture, my friend.

However they have a cpu that does cpu stuff. I asked what happens when the cpu does these cpu stuffs, regarding the GGDDR6 latency. Both PCs and Consoles use Zen 2 architectures. So we can't just say they are not comparable.
 

Guru

Senior member
May 5, 2017
830
361
106
I mean GDDR6 is miles faster than DDR4 ram, coupled with the fact it's a custom build board that would mean the lattency would be lower than desktop counterparts, not higher.

We can't really know, but if you ask me, I'd say 12.5GB of usable GDDR6 ram by the cpu+gpu for games is much faster than any desktop available cpu+ram.
 

psolord

Golden Member
Sep 16, 2009
1,920
1,194
136
I mean GDDR6 is miles faster than DDR4 ram, coupled with the fact it's a custom build board that would mean the lattency would be lower than desktop counterparts, not higher.

We can't really know, but if you ask me, I'd say 12.5GB of usable GDDR6 ram by the cpu+gpu for games is much faster than any desktop available cpu+ram.

I am not sure this is how things are, but then again that's why I started this thread.

What I am asking, is not a matter of bandwidth. It's a matter of latency. And if I'm not greatly mistaken, latency has to do with the physical abilities of the chips to respond to new queries, in conjunction with their true frequency.


Your typical ddr4 3200 has a frequency of 1600Mhz and a cas latency of 15cycles. On the other hand, a RTX 2070 Super has 1750Mhz physical frequency, for 448GB/sec on its 256bit bus.

PS5 will be exactly the same. 448GB on a 256bit bus, therefore 1750Mhz physical frequency. XBOX is less straightforward, with its 10+6GBs on different speeds, so lets focus on the PS5 for now.

We are then essentially looking at 1600Mhz cas15 vs 1750Mhz cas whatever on the PS5 for cpu operations. The way I see it, anything above cas17 would be a handicap for the PS5. Could be better or worse for the XBOX, depending on how these 6+10GBs are mapped.

We also know that even Zen2 is very sensitive when it comes to gaming and memory latency. So, is there any info for all that?
 
Last edited:

ZGR

Platinum Member
Oct 26, 2012
2,052
656
136
This would be good info to find. I would be interested in seeing the latency and timings.

It should have higher latency than DDR4 3200 I would imagine.
 

SPBHM

Diamond Member
Sep 12, 2012
5,056
409
126
the priority is the GPU,
like on the PS4 the CPU has fairly high latency AFAIK, and the bandwidth is limited to 20GB/s while the GPU gets well over 100GB/s
 

itsmydamnation

Platinum Member
Feb 6, 2011
2,776
3,161
136
the priority is the GPU,
like on the PS4 the CPU has fairly high latency AFAIK, and the bandwidth is limited to 20GB/s while the GPU gets well over 100GB/s
i wouldn't make the same assumptions alot of that was driven by the cache design of jaguar. A Zen CCX can push and pull far more data, Jaguar L2 ran at 1/2 speed, 16bytes a cycle. Zen2 is 32Bytes full speed. But good cache systems reduce memory access requirements and Zen CCX is also alot better then Jaguars core complex.
 

Gideon

Golden Member
Nov 27, 2007
1,646
3,711
136
We have PS4 on linux Geekbench 4 scores for jaguar. So GDDR5 for Jaguar is about the same as slower mobile ram Latency wise (~120ns). Not great, not terrible:


I can see higher-clocked Zen 2 with GDDR6 doing better, but nowhere near DDR4 speeds (in latency, bandiwidth is multilpe times better).
 
  • Like
Reactions: ZGR

Saylick

Diamond Member
Sep 10, 2012
3,171
6,405
136
We also know that even Zen2 is very sensitive when it comes to gaming and memory latency. So, is there any info for all that?
It's probably better than what you'd get on desktop Ryzen since the consoles use a monolithic die, so there's no IO die for the data to go through on the consoles. That should improve the gaming performance of the Zen 2 cores.

Edit: You are right in that the overall latency will depend on how loose the timings are on the GDDR6 memory, but the lack of an IO die should help compensate for that. Regardless, the Zen 2 cores will be much faster than the Jaguar cores they replace, in both IPC and clocks. Microsoft advertise 4x the CPU performance, which seems in line with expectations: more than 2x the IPC with 2x the clocks.
 
Last edited:

SPBHM

Diamond Member
Sep 12, 2012
5,056
409
126
i wouldn't make the same assumptions alot of that was driven by the cache design of jaguar. A Zen CCX can push and pull far more data, Jaguar L2 ran at 1/2 speed, 16bytes a cycle. Zen2 is 32Bytes full speed. But good cache systems reduce memory access requirements and Zen CCX is also alot better then Jaguars core complex.

8 years later and massively better CPU architecure perhaps it's not the best idea to assume too much you are right, and it's not ideal but a quick check on geekbench from the ps4 pro and a similarly clocked desktop Jaguar had memory latency numbers in the same range.

I would still assume that the priority is the GPU since these are very heavily reliant on bandwidth and CPU access can slowdown the performance quite a bit

PS4-GPU-Bandwidth-140-not-176.png


xbox one
XBox_One_SoC_diagram.jpg


I think it's clear the priority in this.

NDPS4_5-565x425.jpg

PS4

AMD did make a Zen1 based APU with 8GB GDDR5 with that Chinese "console" running windows, perhaps there are numbers from it around...
 
  • Like
Reactions: ZGR

IntelUser2000

Elite Member
Oct 14, 2003
8,686
3,785
136
Latencies are going to be higher on small dataset access due to higher frequency memory technologies generally being at higher latency, sharing the system memory between CPU/GPU and increased parallelism.

But, consoles have an immense advantage of being on the same platform for years it rarely matters. Lower latency is needed on PC because they are so general purpose and the code is optimized for that.
 

FaaR

Golden Member
Dec 28, 2007
1,056
412
136
So anyone knows what kind of latency we are looking at, regarding GDDR6 on the new consoles?
If anyone here knows for sure they're bound by ironbound NDAs and whatnot. So what you'll be left with is speculation! :)

Modern DRAM tech tends to have roughly the same latency no matter what bus it attaches through - DDR, GDDR or HBM for example all have similar latency because the latency largely stems from the DRAM die fundamentally being an analog device and the time it takes to charge its capacitor arrays and read back said charges and whatnot. So GDDR6 will probably follow the same pattern.

Btw... Memory latency is measured in - typically - ns, not clock cycles. 2 cycles @100MHz or 20 cycles @ 1000MHz = same latency. :)

What latency the CPU finally sees is a different matter though. Probably radically higher effectively than in a PC with dedicated memory controller for the CPU, due to sharing same memory pool with the GPU and every other device in the entire system. And CPU will still be connected via infinity fabric of course to what is essentially a separate northbridge (even if on the same chip instead of a separate), containing said memory controller, quite similar to current desktop zen 3.

It'll be okay, though. CPU memory latency is pretty terrible in PS4 and xbone (not counting on-chip SRAM in the latter), and they still perform quite well all things considered. :)