AE CS6 11.0.1 CUDA BENCHMARK PROJECT – test your graphics cards!

[Ian Mapleson]

That’s quite good for a single 580! 8)

Note that I just keep hunting for new/used Thermaltake Toughpower units listed
on ebay, 1kW or more. I’ve bought half a dozen 1475W units, saved more than $2K
doing this. Most were new or barely used, listed as normal auctions, won them
for less than half normal new price in every case. The 1475 is excellent, easily
handles four 580s and an oc’d 6-core. For two 580s I’d get a 1kW. This is of course
more than one really needs, but the idea is to have plenty of headroom to ensure
that the feed for the CPU is stable even when all GPUs are under load, plus of
course AE can stress the entire system, so only factoring in load power for the
GPUs is unwise. If it was just a single 580 though then I’d use an 850W PSU.

Ian.

——–
SGI Guru

[John Garland]

Hi,
Been using this benchmark for years, finally decided to post scores.
supermicro x8dah dual xeon 5675
96 gig ddr3 1333
two 3gig GTX580 classified

finished 2 min, 55 sec

We have a quad opteron H8qmi with 4 8439 se and 96 gig of ram as well. It has two GTX590 classified GPUs. It has done 2 min 40 sec before.

[Ian Mapleson]

That’s an astonishly good time with just two 580s. It would be very useful to know how
you’re getting such a quicke time than I do with two 3GB 580s, some other aspect of
the system… the dual CPUs perhaps? Which would be odd as this isn’t supposed to be
a CPU test. Storage setup? Hmm, most intriguing.

Ian.

——–
SGI Guru

[Teddy Gage]

Maybe the 96 gb RAM? Could be there was a RAM or HDD bottleneck with this project originally. or he’s lying. 🙂 Would have to know rest of pc specs to figure it out…

[John Garland]

I think it is the number of PCIE lanes the dual CPU board can support. The x58 chipset with one cpu is 24 lanes, the dual xeon x8 supermicro is 36 or 38 or 48.

[John Garland]

I think that is possibly why you don’t see much improvement with three or four cards. Each card is going down to six effective lanes with a single CPU on four cards. It was already choked a little bit on two cards. Supermicro has those big GPU workstations that have four x16 slots with no multiplexing of the lanes. The motherboard I have has two x16 slots, and a x8 which runs the raid card with no sharing of lanes.

The op system and scratch drive are on separate SSD’s. The raw video data is stored on a 8 drive Raid 10 for redundancy and speed.

[Teddy Gage]

Fascinating! Makes perfect sense.

[Ian Mapleson]

Depends on the mbd; I use the P9X79 WS, which does not have any lane disadvantage,
or at least not with two anyway.

Besides, there’s no evidence at all just now that this test remotely requires any
significant degree if continuous PCIe traffic whatsoever. We’re making unwarranted
assumptions here and that’s unwise if the info we post is to be useful in the
decision making processes of potential readers. Proper analysis is required.

Ian.

——–
SGI Guru

[John Garland]

I don’t know how much bandwidth this benchmark needs, but as for the cards themselves I am sure if they only needed eight lanes they would only have made the card with eight lanes. Sort of like the comedian Ron White joke “I don’t know how many bouncers it was gonna take to throw me out of the bar, but I knew how many they had”. We know how many lanes the card has.

It of course depends on the processor and board. i7-3930k support 40 pcie lanes. throw a third card on it and you are at 48 lanes. some lanes are already in use for other motherboard resources. on the other hand i7-2600K only has 16 lanes. I think your own testing showed how much worse the 2600 was with two and three cards.

my old 5675 processors don’t have built in pcie lanes, they have to be handled by dual 5520 tylersburg northbridge chipsets off the chip, for a total of 72 lanes to handle the motherboard resources.

You asked what the difference in performance might be and that is my best educated opinion. When I jumped my ram from 48 to 96 GB it didn’t help at all. Processor clock doesn’t seem to matter past a point. Lane bandwidth seems to be the main difference between the two setups.

When I put these two GTX580 in a board with a single northbridge, I7-990X, and 24 gig ram at higher speed and clocks, they only score 3 minutes 59 seconds. About the same as when I put a single GTX590 in the dual xeon machine which is of course choked to 16 lanes.

There is a sweet spot for every price range.

Dual cpu isn’t necessary with a 17-3930 and two cards, but if you want to get the most out of three cards, it will be needed.

[Ian Mapleson]

John Garland writes:
> … but as for the cards themselves I am sure if they only needed eight lanes they would
> only have made the card with eight lanes. …

That’s not how such cards are made at all. 😀 They’re all just natively x16 of whatever PCIe
revision.

> It of course depends on the processor and board. i7-3930k support 40 pcie lanes. throw
> a third card on it and you are at 48 lanes. …

That doesn’t matter if the board has PCIe switches, and many do.

> … I think your own testing showed how much worse the 2600 was with two and three cards.

I only have a 2700K atm.

> my ram from 48 to 96 GB it didn’t help at all. …

So far I found the biggest difference for GPU rendering was made by the RAM clock. In one
case, dropping the RAM from 2133 to 1866 increased the render time by 10%.

However, I’ve not tested PCIe speeds. Will try this at some point.

> Dual cpu isn’t necessary with a 17-3930 and two cards, but if you want to get the most out
> of three cards, it will be needed.

For AE, that I can well imagine, but this conclusion is AE specific. Other test like Arion
don’t show such dependencies at all, ie. the scale nicely irrespective of the platform. I’ve
alrady observed with other tests that AE does not properly exploit multiple GPUs in numerous
cases, it’s just not written that well, especially since it doesn’t offer any kind of round-
robin option.

Ian.

——–
SGI Guru