Hi,
Does the GTX295 uses half of the PCI bus for one of its GPUs and the other half for the second one? so if I use a PCI-E16
each half only sees 8 and therefore the bandwidth is halved?
Is it the case even if one half of the 295 doesnt process any work, the other half still gets only 1/2 of the PCI bandwidth?
thanks
eyal
When I bandwidthTest one half of my GTX 295, I get 5818 MB/sec with pinned memory. That’s the full speed of the PCI-Express 2.0 bus.
BR04 (the PCIe splitter/switch on there) is actually a switch, not just a splitter. You’re not getting 8x out of it to each card, you’re getting something that hovers between 16x and 8x to each device depending on whether they’re hitting the bus simultaneously.
Hi,
I ran the bandwidth tests on the machine (AMD, linux, with 4 GTX295) and those are the results:
1 card:
-bash-3.2$ ./bandwidthTest --memory=pinned
Running on......
device 0:GeForce GTX 295
Quick Mode
Host to Device Bandwidth for Pinned memory
.
Transfer Size (Bytes) Bandwidth(MB/s)
33554432 5484.5
-bash-3.2$ ./bandwidthTest --memory=pageable
Running on......
device 0:GeForce GTX 295
Quick Mode
Host to Device Bandwidth for Pageable memory
.
Transfer Size (Bytes) Bandwidth(MB/s)
33554432 2365.6
2 cards (with 4 cards the results were ~ the same):
-bash-3.2$ ./bandwidthTest --memory=pinned
Running on......
device 0:GeForce GTX 295
Quick Mode
Host to Device Bandwidth for Pinned memory
.
Transfer Size (Bytes) Bandwidth(MB/s)
33554432 1468.2
-bash-3.2$ ./bandwidthTest
Running on......
device 0:GeForce GTX 295
Quick Mode
Host to Device Bandwidth for Pageable memory
.
Transfer Size (Bytes) Bandwidth(MB/s)
33554432 1296.7
the hardware guy in our company says the hub ( tmurray, is that the BR04 you mentioned?) is now the bottleneck.
Weird that the performance drops with the second card and not lineary…
Is this something specific to the board we use? Are there other boards that give 6GB/s for 4 cards?
Again sorry for my ignorance in hardware :)
thanks
eyal
Which slot did you put the second card into, and what motherboard do you have? In my ASRock Supercomputer motherboard, the connectivity of each slot drops from x16 to x8 if I put two cards right next to each other. You seem to be getting the equivalent of a drop from PCI-Express 2.0 x16 to PCI-Express 1.0 x8 or PCI-Express 2.0 x4, which is unexpected even under that hypothesis. Does your motherboard manual give any guidance on the performance of using multiple slots at once?
I put them in slots 0 and 1. Hopefully today we’ll try slots 0 and 2. And indeed it seems it dropped to x8.
The slots issue is not important though no? it doesnt help me if I cant put all cards in the slots.
I’ll ask the IT manager to have a look at the manuals again…
thanks
eyal
Sure, but this test is mostly to determine wither the bandwidth loss is due to the motherboard itself, rather than the PCI Express switch embedded in the GTX 295. If putting cards in adjacent slots drops the bandwidth by a factor of 2, that is to be expected given the limitations on the number of PCI Express lanes on current motherboards. If adjacent slots drop by a factor of 4, that sounds like a motherboard bug (or “limitation”) that is worth contacting the manufacturer about.
The question is all about who to blame. :) So far it isn’t clear what’s going on, but I suspect the problem is not the GTX 295.
All GTX 295 are sold out in all the major online stores. I heard Nvidia is coming out with a new version of the card. Some say it will be this month. Does anyone have any idea?
Not sure I would get too excited yet. I’ve seen the GTX 295 supply fluctuate a few times on Newegg in the past, and thought the same thing. :)
If the plan is to replace the GTX 295 with another dual-GPU solution, I’m not sure where they would go with it. NVIDIA never seems to debut entirely new chip designs in the the dual-GPU configuration, so the only remaining option would be to downclock the GPU in the GTX 285 and put two of those in a card. That would widen the memory bus to 2x 512 bits, and up the memory to 2x 1GB. Not a huge improvement, but is possible.
The other upgrade option would be the transitioning of the GT200 chip to a smaller process, which would be nice. Getting the GTX 295 power usage below 225W would let you power it off of two 6-pin PCI-E connectors, dramatically simplifying the options for quad card systems.
There is a third option: sales of the GTX 275 are high and cannibalizing the supply of chips that would normally go to the GTX 295. If you look at the specs of the 275 and the 295, you see that it is almost certainly the same chips used in both, just run at slightly different clock rates. If the GTX 275 is doing great, no point in spending chips to restock the GTX 295 that sells much more slowly.
However, this is all speculation. The most likely situation is still an inventory hiccup on a low volume product. :)
OK, maybe something is going on. I noticed this morning that Newegg has delisted all of the GTX 295 cards except the red card from EVGA, and that card was out of stock. If more of the same card were coming, the products would just be marked out of stock rather than removed entirely.
If these “single PCB GTX 295” rumors are true, that would be nice if it also lowers the price of the 295. (I could go for a price progression that is GTX 275 = ~$250, GTX 285 = ~$350, GTX 295 = ~$450…)
For me, the hardware knowledgeless ;), what does this mean “single PCB”?? what difference will it make ? hardware/software wise?
PCB stands for “printed circuit board.” Currently the GTX 295 is two circuit boards (each with one GPU and several memory chips) with a metal heatsink sandwiched in between. The rumor is that NVIDIA is switching the design to put everything on a single circuit board. Moving to a single circuit board will have no observable difference for performance, but it should make the GTX 295 a little cheaper to manufacture.