Posts: 6.069
Threads: 68
Joined: May 2010
Reputation:
167
Location: Grenoble, France
07-02-2015, 12:21 PM
(This post was last modified: 07-02-2015, 12:22 PM by gregory.)
Yes. As Ref said it depends on the RAM speed too.
I'm working on some PCIE IP, so I know the spec quite well.
There are lots of overhead. You have overhead of physical layer (128b/130b). You have a packet overhead (I need to double check it but it is around 4-5 DW so 16-20B). In order to reduce the latency real data payload of packet is around 128B. There is also some extra packet to communicate the room available for transfer (credit mechanism, ack/nack). So globally I think RAM is faster than PCiE. (Ram has some overhead too)
Ofc I'm too slow with my tel...
Posts: 3.526
Threads: 6
Joined: Dec 2013
(07-02-2015, 12:08 PM)refraction Wrote: Hmm not sure, the theoretical bandwidth of a PCIe 3 graphics card is 32GB/sec, which is obviously faster than system ram. That is theoretical however, much like dual channel memory is not a 50% increase in speed.
as a general rule, PCIe 3.0 isn't any faster than PCIe 2.0 which is a theoretical maximum of 16GB/sec, but again that could be a system ram limitation.
isn't that memory bandwidth nullified if you write to the vram directly? pcie hotlinked with the memory controllers on die. or not?
Posts: 6.069
Threads: 68
Joined: May 2010
Reputation:
167
Location: Grenoble, France
Actually a faster gpu might helo too. When you read back a target, the gpu must have finished to render it first. It is often worst, because GPU is often late (vs the cpu)
Posts: 6.069
Threads: 68
Joined: May 2010
Reputation:
167
Location: Grenoble, France
Well game are optimized to use the correct bandwidth, if the link is saturated perf go down very quickly. Besides GPU mem increase a lots. So you can store texture for severals frames. And finally API slow evevery thing down.
Posts: 369
Threads: 35
Joined: Jul 2014
Reputation:
4
BTW shouldn't DX12 / Vulkan (looking at you gregory) grant even a lower access without the need for those pesky GPU-CPU Transfers