WEBVTT 00:00:00.240 --> 00:00:08.559 So, as a follow-up to my recent video where I took two GTX 580s, threw them on 00:00:05.920 --> 00:00:15.200 the Gigabyte G1 Gorilla motherboard, and I tested in dual 16x mode in SLI and 00:00:12.080 --> 00:00:17.920 dual 8x bandwidth mode in SLI. I am 00:00:15.200 --> 00:00:23.119 going to be taking an even higher an even more demanding card, the Radeon 00:00:21.000 --> 00:00:30.080 HD6990, and I'm going to be running it in a single card configuration in a 16x 00:00:25.840 --> 00:00:32.640 slot, an 8X slot, and finally a 4X slot 00:00:30.080 --> 00:00:37.360 to find out what difference there is. In terms of performance, I'll be using the 00:00:34.000 --> 00:00:39.760 3D Mark 11 extreme preset. Now, please 00:00:37.360 --> 00:00:45.920 note that the extreme preset will not show as much of a difference as one of 00:00:42.399 --> 00:00:48.879 the lower quality presets, uh, such as 00:00:45.920 --> 00:00:53.600 the entry or performance presets. But what it does give us is a more realistic 00:00:51.440 --> 00:01:01.280 real world scenario because most people are not going to be using a 6990 to run 00:00:57.120 --> 00:01:02.559 Quake 3 at 1024 x 768. So, you're going 00:01:01.280 --> 00:01:07.920 to be turning up the details, you're going to be turning up the res. So, that is the scenario that we will be looking 00:01:05.840 --> 00:01:13.119 at. Something to bear in mind before I get too far into my testing here is that 00:01:10.400 --> 00:01:18.159 the top slot at 16 and the second slot at 8X are both running off the PCI 00:01:15.680 --> 00:01:22.640 Express controller integrated into the Sandy Bridge CPU on this board which is 00:01:20.560 --> 00:01:27.600 clocked at 4.7 GHz by the way to eliminate any CPU bottlenecking or as 00:01:24.960 --> 00:01:33.640 much as possible. The bottom slot actually runs off of the chipset at 4x 00:01:31.119 --> 00:01:38.000 speed. So, it has an additional disadvantage that is not shared by the 00:01:36.320 --> 00:01:41.600 top two slots. So, just bear that in mind. Take the results with a bit of a 00:01:39.360 --> 00:01:47.360 grain of salt, but I don't have another platform handy where I can show you 16x, 00:01:43.759 --> 00:01:51.880 8x, and 4x where they will all be equal. 00:01:47.360 --> 00:01:55.040 So, my 16x result is a score of 00:01:51.880 --> 00:01:57.680 3,330. So, I'll be back with 8x and 4x 00:01:55.040 --> 00:02:02.399 results shortly. Now, this may look like a stupid configuration. I've got my 6T90 00:02:00.159 --> 00:02:06.320 and a GTX 580 here. But the reason I've installed the GTX 580, which won't 00:02:04.320 --> 00:02:12.520 affect my benchmark scores in any way, is to force these PCIe lanes to go from 00:02:08.959 --> 00:02:16.480 16x to 8x8x. So, you can see here within 00:02:12.520 --> 00:02:19.760 GPUZ, the 6990 reports as running at 00:02:16.480 --> 00:02:21.520 16x. And while that may be true because 00:02:19.760 --> 00:02:26.959 it's running internal Crossfire where each GPU is getting 16x bandwidth from 00:02:24.480 --> 00:02:32.160 the uh from the lane spplitter that's on the PCB 00:02:29.239 --> 00:02:37.280 itself. The actual lane splitter itself is only 00:02:33.560 --> 00:02:39.040 getting if we look here at the GTX 580. 00:02:37.280 --> 00:02:44.720 So you can see that the actual bandwidth to the slot is only 8x because it is 00:02:42.080 --> 00:02:52.040 correctly split at this time. So my benchmark results are 00:02:48.519 --> 00:02:56.080 3,312 3D marks. So we have seen 00:02:52.040 --> 00:02:59.200 a difference in performance between 16x 00:02:56.080 --> 00:02:59.200 and 8x 00:02:59.640 --> 00:03:08.840 312. Okay, very small, very small. And I 00:03:03.280 --> 00:03:11.200 will be back with PCIe 4X results very 00:03:08.840 --> 00:03:17.200 shortly. So at the end of the day, this video becomes more about how if you 00:03:13.680 --> 00:03:20.239 internally bottleneck a card um 00:03:17.200 --> 00:03:22.640 according to the computational resources 00:03:20.239 --> 00:03:28.159 located on the card that the amount of PCI Express bandwidth you supply it with 00:03:25.200 --> 00:03:32.799 is actually going to make very little difference in terms of performance. Now, 00:03:30.239 --> 00:03:37.200 something to bear in mind too about this particular testing setup today. You can 00:03:34.640 --> 00:03:40.799 see I had to nudge my motherboard over a little bit to get the uh 6T90 to fit 00:03:39.200 --> 00:03:47.440 into that bottom slot on my tech station. But so, I've got it in the PCIe 00:03:43.959 --> 00:03:50.080 4x slot. Now, think about it this way. 00:03:47.440 --> 00:04:00.400 Basically, I've got two GPUs. I have on a 6990, I pretty much have two 6970 00:03:54.080 --> 00:04:00.400 GPUs. Okay? So, so bear with me here. 00:04:00.760 --> 00:04:08.040 So, by the comparison I did the other day where I was comparing 00:04:05.799 --> 00:04:13.439 16x6 and 8x8x SLI, this is basically running it 00:04:11.599 --> 00:04:19.400 on this motherboard, this card in 16X mode is pretty much like running two 00:04:15.840 --> 00:04:24.520 cards, two high-end single GPUs at 00:04:19.400 --> 00:04:28.560 8x8x, 4x4x in the 8x slot, and then at 00:04:24.520 --> 00:04:29.840 2x 2x in that 4x slot at at the very 00:04:28.560 --> 00:04:36.000 bottom that that's essentially the amount of bandwidth we have available 00:04:31.680 --> 00:04:37.520 per GPU and we still see very very 00:04:36.000 --> 00:04:42.240 little fluctuation in terms of the overall score. So, what I may have to do 00:04:39.040 --> 00:04:44.960 in the coming uh next little while is do 00:04:42.240 --> 00:04:51.840 an updated video about this where I find a slightly less uh GPU speed 00:04:49.280 --> 00:04:55.840 bottlenecked benchmark and uh demonstrate maybe more of a more of a 00:04:54.000 --> 00:05:00.000 typical or worst case scenario for how much the PCI Express bandwidth affects 00:04:58.000 --> 00:05:02.639 single card performance. So, thank you for checking out this little episode of 00:05:01.360 --> 00:05:07.880 Linus Tech Tips. Don't forget to subscribe for more unboxings, reviews, and other computer videos.