What kind of mainboard will you be using for 4 double-slot (and double-PCB) cards? :) Please show pics :D
Wikipedia thinks it will have a shader frequency of 2.4 GHz? Ha, we’ll see about that.
Notice that the bus width has been halved, so perhaps they needed to radically increase the shader frequency to compensate.
GDDR3 → GDDR5 doubles the bandwidth per clock so you can compare 512 bit GDDR3 with 256 bit GDDR5 by clock speed.
The reportedly used Hynix memory has 7 Gbps, so that translates into 7*256/8 = 224 GB/s :):)
It is more likely they need to increase the shader freq to keep the memory busy ;)
Doubling both bandwidth and FLOPS in 3-4 months? I’ll believe it when I see it.
Which shouldn’t be long ;)
Well, these are actually just specs from the Hynix website. Couple it with a 256 bit memory bus (which ATI is doing with GDDR5 at this time) and you get the 200 GB/s figure, so that one should not be impossible to achieve. That will already mean double performance in a lot of cases without doubling the GFLOPS number…
It would actually be quite in line with this graph from NVISION: http://gallery.techarena.in/showfull.php?photo=13977
So we can still be hoping for this one ;) http://gallery.techarena.in/showfull.php?photo=13976
Well, that’s bandwidth internal to the card…but PCI-Express 3.0 is due out later this year, which would also double the device-host bandwidth!
GPU makers don’t run RAM at its max rated speed. They often go for tigher timings and lower freq. Even at 7Gbps, that’s only a 60% increase, not a doubling.
Still, 30-60% boost in bandwidth is much more reasonable than a 200% jump in FLOPs. Doubling shader frequency, esp without introducing a completely new arch, sounds very fishy. GPU frequency has never jumped like that. I’m sure in the end FLOPs will also increase about 50%.
And this is all vs the GTX280, not the GTX295. We should all be moving to multiGPU. It’s the present as much as the future.
It’s more housekeeping to do multiGPU, and I am bad at housekeeping ;)
But you are right.
According to specs, (see for example fudzilla, the on board memory of the GTX285 will be ~2.5 GHz DDR3. This should be compared to ~ 1.2 GHz DD3 memory on GTX280 (according to Nvidia system monitor).
Does this mean that memory bandwidith of GTX285 is two times higher than GTX280, or I am missing something here?
The 285’s memory is 2484 Mhz, the 280’s is 2214 MHz, so there is a very small on-device memory bandwidth boost on the 285.
Where it gets confusing is that this is DDR “double data rate” so many people will double the clock frequency to give you “effective” clock rate.
That’s already been done to these numbers. The ~1200Mhz speed you saw for your 280GTX is undoubled.
It causes endless confusion, but the marketing people prefer whatever interpretation gives them bigger values.
Even worse, GDDR5 is still called “DDR” but it actually is effectively quad rate. AMD uses this for its modern cards, and it’s extremely likely that NVidia’s next batch of cards will as well… it makes both the chip and board design much easier since you have hundreds fewer bins and traces.
I’m using a Foxconn Destroyer. Right now I’m able to get 3 cards up and running. The fourth card shuts down the (1200W) power supply as soon as I push the power button. So, I’ll have to wait for a supplementary power supply to power the fourth card (or support powering two of the cards). Hope that’s enough to have all of them running at full speed.
Please excuse the crappy image quality. It’s been taken with the mobile phone…
Actually, I think the doubled numbers are more meaningful. The “true” “clock frequency” is just the signal on the CLK input pin. It could be made 100x slower, and the chip internally would have its own oscillator that uses the external signal just for synchronization. This is in fact what GDDR5 is doing. It is more accurate to say that the CLK signal (for the purpose of signal quality) is 1/4 of the true clock rate. People should only talk about the actual 2.4 or 7 GHz data rate, instead of confusing matters with the 1.2 or 1.75 GHz CLK signal.
Well, it turns out the power supply is only able to run two cards. So, I’m choosing (and I see the FASTRA guys did the same) the Thermaltake Toughpower 1500W, which has 40A on each of the two PCIe 12V rails while the bequiet 1200W had only 20A.
Will keep you posted when the new PSU has arrived (probably tomorrow)…
I’m also planning for a mega-GPU box to pick up cute girls, so thanks much for sharing the issues beasts like this introduce.
I was about to laugh at a “1200W” PSU that only gives 20A. But it seems there are 6 rails at 20A each. Not too bad… sounds like two rails per card would be great. But I found this very good review of the beQuiet you have, and I think its fatal flaw is identified:
Thanks for pointing that out!
Can’t seem to get that fourth card to work :wacko: The Thermaltake is running smoothly with three cards, now, and handles four cards (as in: fans start up and keep running). But, there’s no beep and no booting. If I take the 1200W bequiet just to power the fourth card, the system beeps but does not boot (no display on any of the graphics adapters). Maybe it’s the board or the cards.
If someone from NVIDIA could confirm that a four-card eight-GPU GTX 295 setup is possible, I’d appreciate that.
EDIT: Some further testing showed that even when using a secund PSU to power the fourth card the system won’t boot. So, I’m guessing it’s the mainboard (maybe a power issue there) or something like that, which would be a bummer.
I wonder if this is the problem with the power draw on the PCI-Express slots themselves. Each physical x16 slot is supposed to be able to provide 75W of power to the card, with the rest of the power coming from the 6 and 8 pin connectors. With 4 cards, that’s 300W of power draw on the motherboard, not counting what the CPU, RAM, and chipset require. Does your main PSU have a third rail for the motherboard, and how much power can that supply? Also, I’ve seen some motherboards with an extra power connector (usualy a 4-pin molex connector like on IDE drives) in order to give the PCI-Express slots an extra boost.
That’s what I thought about as well. The PSU should be able to take it as it has one 20A rail each for mainboard and CPU. However, I will give the MSI K9A2 Platinum V2 board a try as this is what the FASTRA guys are using. It has an AMD chipset instead of the nVIDIA one on the Foxconn Destroyer, also it has the molex connector to “stabilize” power supply to the PCIe slots. For now, the system is running pretty stable with three GTX 295.
Update: The MSI K9A2 Platinum V2 is the solution. I noticed that it has the extra molex plug on the board to support PCIe power supply and it seems that is what is needed. I would not run any more than one hard disk, though, as the drives draw their power from the same (40A) rail that powers 2 of the cards via PCIe power plugs and the supplementary power for PCIe on the board. The system has run my Kernel on 8 GPUs now for a about 15 minutes without a hiccup.
So, my recommendation is to stick exactly with the specs on the FASTRA homepage regarding case, board and power supply. Get the latest and fastest CPU for that board - my Phenom X4 9950 (2.6 GHz) runs at 100% load all the time. External Image
Observations: That thing is loud. After about 2-3 minutes the graphics cards’ fans run at full speed. Also, no heating is required in the room (about 20 square meters) the system is running in. When I leave the window open over night (at about 0-5°C outside temperature) it is comfortably warm in the morning. I don’t think that’s cost effective heating, though :wacko:
You need more case fans, the internal case temperature is getting too hot.
Also if the heat in the room gets uncomfortable, I suggest a window fan such as this one: http://www.amazon.com/Honeywell-HW-628-Twi…n/dp/B0000CGQYA It works really well.