AMD Kaveri APU features the Onion + bus like the PlayStation 4
- Thread starter onQ
- Start date
Betanumerical
Veteran
Onion and Onion+ share a 10GB/s read/write bus.
What remains to be seen is what difference the two approaches make (10GB/s coherent read/write) and (30GB/s coherent read).
Sorry x the OT and the eventual mistake.
But from VGleaks table is seems clear the 68gb/s BW for CPU.
http://www.vgleaks.com/durango-memory-system-overview/
I read wrong?
Also it seems 30gb/s Read/write ...
For the PS4:
- the CPU has it's own <20GB/s bus.
- onion/onion+ are a separate bus with 10GB/s read and 10GB/s write (you could consider that a 20GB/s coherent read/write "bus", but that might not be entirely true). [onion+ is basically a "flag" to change caching]
For the XB1:
- there is a 30GB/s coherent read/write bus.
In terms of how much bandwidth the CPU is expected to use, AFAIK vgleaks 'borrowed' these docs from MS literature, so the figures are probably realistic (4GB read, 4GB write, 1.5GB write combined per module). [e.g. ~20GB/s for the pair, including coherent usage]
http://www.vgleaks.com/durango-memory-system-example/
It IS read/write, on the other thread someone has been trying to spin that the having no full cache coherency thus it means no coherent memory, or as someone like to keep on repeating, coherent read only.
Let's try this one more time, coherent memory means that a pointer can be passed from the CPU to the GPU directly. Whether you need flush the cache or not (cache coherency) is an implementation detail, in X1's case, the code decides whether GPU wants snoop into CPU cache or not, and that the CPU can't snoop other components' cache.
This is just an exemple of the memory usage (it is just an exemple) still is visible the connection with The ddr3 block. I suspect that a part from 30gb/s bus there is also a bus direct to DDR3 capable of 68gb/s bandwith.
The diagram is clear!!!
To me it seems "Onion X1" 68GB/s, and Coherent BUS 30GB/S
Last edited by a moderator:
http://www.vgleaks.com/durango-memory-system-overview/
I know it's probably OT but if we are heading down this path. The other diagram has a bunch of zeros near the bottom, so which one is which?
Old story...
One is the peak BW.
The other one is just one exemple on BW usage by the X1, it is not the medium or the working values but just an exemple out of 1000 of the possible BW usage.
Reading the vgleaks memory exemple is clear that there are 2 Bus inside X1 absolutelly indipendent one from each other:
DDR - Northbridge with 68gb/s bandwith
Coherent bus 30gb/ s bandwith.
Let me say that this is quite a huge difference from Sony set up!
Last edited by a moderator:
I don't recall seeing anything to suggest a hard split into read and write.
Incidentally, what is this odd shorthand you have of using 'x' for 'for'? Where does that come from? It seems very non-standard and an unnecessary confusion.
Opss, sorry for that... It is an abbreviation in use in my country...
I have never thought before it was not in use in other part of the world!
I think you are still spinning, there's no evidence suggesting that PS4 CPU can snoop GPU cache, in fact, it's quite the contrary.
Last edited by a moderator:
No... x stand for "for"... X stand for x .... Well never mind
Onion is the GPU bus that snoops the CPU caches.
The CPUs rarely draw data from it, although I haven't ruled out a very small time window where the memory pipeline might forward a GPU write to a CPU in the small number of cycles it has before it makes it to memory.
If people want to measure CPU bandwidth, it's the bandwidth of the connections leading out from the CPU section that would need to be measured. One case that doesn't appear to be possible from the descriptions of older APU bandwidth measurements is the case of adding coherent with write-combining bandwidth. There was a common bandwidth ceiling they both had to share.
In those systems, memory bandwidth was much lower, so it isn't clear if that may have been tweaked this time around.
The CPUs rarely draw data from it, although I haven't ruled out a very small time window where the memory pipeline might forward a GPU write to a CPU in the small number of cycles it has before it makes it to memory.
If people want to measure CPU bandwidth, it's the bandwidth of the connections leading out from the CPU section that would need to be measured. One case that doesn't appear to be possible from the descriptions of older APU bandwidth measurements is the case of adding coherent with write-combining bandwidth. There was a common bandwidth ceiling they both had to share.
In those systems, memory bandwidth was much lower, so it isn't clear if that may have been tweaked this time around.
It seems that a few people assume that's the case, that in Liverpool the CPU can snoop GPU cache via Onion, thought I have my doubts because the whole Onion+ bus design implies that GPU cache can't be snooped by the CPU and hence relies on the selective flushing to make the data visible to the CPU.
Few had used this to spin that the other design is not fully coherent, but rather, coherent read only.
http://www.gamasutra.com/view/feature/191007/inside_the_playstation_4_with_mark_.php?page=2%29
Per Cerny:
He's saying that to have full coherency at the memory level via uncached access, which is Onion+.
He's saying the same thing, that even if compute is cachable in L2, you still need to flush to memory, though there's some optimization there for the graphics so that it doesn't flush entirely.
Therefore, I think I can conclude that the PS4 CPU does NOT snoop GPU cache, I wouldn't go as far as spinning it as coherent read only though ;-)
Also that HUMA requires HW cache coherency between CPU and GPU, neither can be HUMA by definition due the lack of GPU cache -> CPU coherency.
Last edited by a moderator:
Just to clarify my earlier post, the forwarding case I'm talking about is the possibility that the memory request queue that the Onion bus put the GPU write on can forward data if a CPU read to that same cache line hits the queue, or if the memory controllers can satisfy reads from their write buffers.
The GPU caches wouldn't be involved, hence why Onion+ meets AMD's threshold for coherence--although I am not sure that's the same one for HSA because HSA's bar seems a bit lower.
Yea sorry, I re-read your post and got that part, the GPU cache's either already flushed or not involved at this point.
This' different than what some are claiming that CPU can snoop GPU cache on the PS4.
I've seen no indication that the CPU can snoop the GPU cache on PS4.
Similar threads
