Question about multithreading in hardware mode
#31
1) you do realize he's talking about software mode, right ?
2) use hardware mode.
We're supposed to be working as a team, if we aren't helping and suggesting things to each other, we aren't working as a team.
- Refraction
Reply

Sponsored links

#32
(07-02-2015, 09:11 PM)xemnas99 Wrote: You can prove it by setting the extra rendering threads to 3 in case your CPU has 4 cores, then run a game and look at the CPU usage. If the system RAM or another component is the bottleneck, the CPU usage should be lower than 50% because the CPU has to wait. Then again, the extra rendering threads would be kind of pointless if the CPU wasn't the bottleneck.

i'm not saying that the CPU is not the bottleneck ofcourse it is, i'm saying that the GPU is faster cause it have a much superior bandwidth Which is demanded heavily in any rendering process
Reply
#33
(07-02-2015, 06:29 PM)xemnas99 Wrote: No, the reason is the CPU is used for rendering.

Anyway, I don't understand the talk about PCIe. The bandwidth of the system RAM is lower than that of PCIe, so PCIe can't be the bottleneck.

http://www.gamersnexus.net/guides/1349-r...nel/Page-3
http://www.corsair.com/en-us/blog/2014/s..._synthetic

PCIe G3 is 15GB/s (likely much less in real situation, my guess is around 12-13 GB/s) by direction. So I have the spec in front of my eyes. Each packets of data contains 16B of header + an optional 4B of digest. Then the physical layer will add another 4B for start packet and 4B for end packet, + another 4B CRC. Small transaction have a big overhead. I think the typical max_payload is 128B, so there is 20-25% overhead due to protocol.

https://devtalk.nvidia.com/default/topic...s-missing/

As a side note, if you can configure the max payload in your bios, I suggest to put at least 512B Wink
Reply
#34
(07-03-2015, 11:36 AM)Serial Hacker Wrote: i'm not saying that the CPU is not the bottleneck ofcourse it is, i'm saying that the GPU is faster cause it have a much superior bandwidth Which is demanded heavily in any rendering process

I'm not sure I follow you. So you agree that the CPU is the bottleneck, but at the same time you still insist that the memory bandwidth is what makes the hardware mode faster? It seems contradictory to me. Anyway, you can ask the users who use APUs or iGPUs (HD Graphics 5000 or higher) if the hardware mode is as slow as the software mode.

(07-03-2015, 12:34 PM)gregory Wrote: http://www.gamersnexus.net/guides/1349-r...nel/Page-3
http://www.corsair.com/en-us/blog/2014/s..._synthetic

PCIe G3 is 15GB/s (likely much less in real situation, my guess is around 12-13 GB/s) by direction. So I have the spec in front of my eyes. Each packets of data contains 16B of header + an optional 4B of digest. Then the physical layer will add another 4B for start packet and 4B for end packet, + another 4B CRC. Small transaction have a big overhead. I think the typical max_payload is 128B, so there is 20-25% overhead due to protocol.

https://devtalk.nvidia.com/default/topic...s-missing/

As a side note, if you can configure the max payload in your bios, I suggest to put at least 512B Wink

Thank you for the information. By the way, the RAM in the first article runs at 2400 MHz. I'm not sure, but I think most PCs have 1600 MHz RAM. I forgot the dual-channel architecture, but for actual games it only makes subtle differences:
[Image: shogun2-2.jpg]
[Image: cinebench-results.jpg]
Reply
#35
Yes, first link is DDR4 (that why I give a 2nd link). In our case the limitation is the PCIe. Actually even if the PCIe has the same speed as the RAM, it means that a RAM-VRAM transfer is 2 times slower than a single RAM transfer Wink
Reply
#36
(07-03-2015, 07:20 PM)xemnas99 Wrote: Thank you for the information. By the way, the RAM in the first article runs at 2400 MHz. I'm not sure, but I think most PCs have 1600 MHz RAM. I forgot the dual-channel architecture, but for actual games it only makes subtle differences:

It depends, if you're using an iGPU, the more channels you have on your ram can make a huge difference. Going from Single to Dual channel memory can increase your FPS by 20-100%, providing that the memory bandwidth was your limit, in most cases on modern iGPU's, it is. Faster ram does also help in this case

Heres some numbers: http://www.hardware.fr/articles/872-6/pe...eld-3.html
[Image: ref-sig-anim.gif]

Reply




Users browsing this thread: 1 Guest(s)