3GB Geforce GTX580 VS 2GB Geforce GTX680?

[m ratta]

Hi Guys,

I want to get a new Graphics card specifically for CUDA acceleration in Premiere Pro. The Two cards I'm looking at are:

3GB Geforce GTX580
2GB Geforce GTX680

The 680 is obviously faster, more power efficient and has new technology implemented which to me seems more specifically designed for games. However, it has 1GB less RAM then the 580. My question is, which is more important for CUDA acceleration, faster card or more RAM? Which will prove more beneficial in terms of performance (price not an issue)?

I will be running it on a 4.6ghz 17 PC with 32B RAM. Any help would be greatly appreciated.

Thanks guys.

[Subhadip Sen]

GK104 is Nvidia's finest gaming chip since G80 back in 2006. The key word here is 'gaming'. GK104 has been stripped down, the compute cache structure is gone, FP64 has been scaled back massively, a narrower memory bus, fewer PolyMorph engines and fewer ROPs among other compromises. The net result is Nvidia has finally not just closed in the huge gap to AMD on gaming efficiency (in every way - from performance per die area to performance per Watt) but in fact edged out AMD for the first time in over 5 generations. At what cost? Compute performance is very erratic. AMD's HD 7970 is winning by enormous margins (~400%) in OpenCL benchmarks, and even a GTX 560 Ti 448 (let alone a 580) is faster. FP64 / double precision is even worse, not helped by the fact that Nvidia has further crippled it to differentiate Quadro and GeForce. In short, GTX 680 is awesome for DirectX gaming. As for compute, it depends on the application. My guess is much of Premiere Pro's CUDA acceleration is single precision, in which case GTX 680 may be faster as FP32 is up 90%, but that is just an approximate guess that needs to be verified. For now you might want to go for the GTX 580 3 GB - prices will be dropping sharply so you could score a good deal. Also, 2GB may be a bottleneck for 4K+ resolutions, and 4GB variants are only coming in April.

PS: GTX 580 3GB are already selling for $450 and may drop even further as the $350 AMD 7870 is faster.

[Tom Lowe]

Get the 4GB GTX 680.

When I switched from the 1.5GB GTX 580 to the 3GB GTX 580, my render times became MUCH faster.

[Jeff Kilgroe]

I recommend the EVGA GTX680 FTW Edition. :) That's their OC'd 4GB model... Of course, I don't have one yet, but have it on order...

I don't think anyone is actually shipping the 4GB models yet. And all the 2GB models are out of stock everywhere you look since they just hit shelves and every gamer out there wants one.

Looking at specs comparisons, I'm in agreement with what Subhadip says above vs. AMD... Unfortunately, as great as the new AMD 79XX cards may be, they're a non-starter for many of us. Too many of my apps I need to work in on a daily basis rely on CUDA. Wish I could change that, but that's the current reality of it.

[Mike P.]

The 680 has over 1500 CUDA cores... how much does the 580 have? Also, the memory bandwidth (speed) is higher too...almost 200gb/s... That's worth waiting for, especially if you've got the coin. If you don't have the cash, you should wait anyway for the inevitable 580 price drops.

[DJ Meyer]

Why do folks adamantly recommend the 680 over the 580 for CUDA applications, without having tested them? Early benchmarks I've seen on general tech sites show it not being much faster than the 560 or 580 in general compute performance, due perhaps to the changing of the way they are counting the cores and prioritization of gaming performance over CUDA performance.

Are there some more CUDA specific benchmarks out that I have missed?

Hoping to get one to test for myself.

[Subhadip Sen]

The 680 has over 1500 CUDA cores... how much does the 580 have? Also, the memory bandwidth (speed) is higher too...almost 200gb/s... That's worth waiting for, especially if you've got the coin. If you don't have the cash, you should wait anyway for the inevitable 580 price drops.

GTX 580 has 512, but it's not as straightforward. Kepler is a vastly different marchitecture to Fermi. Previously Nvidia's design philosophy was few high efficiency big shaders at high clocks (2 x core clocks). So, GTX 580's 512 shaders were 'hotclocked' at 1550 MHz. With Kepler Nvidia has adopted AMD's model - many small, dense not-as-efficient shaders at low clocks (AMD 7970 has 2048 shaders at 925 MHz). GTX 680 has 1536 shaders clocked at 1006 MHz. Kepler shaders are smaller and less complex, most notable FP64 has dropped to 1/24th rate versus 1/8th for Fermi. So, the numbers are not really comparable. However, what is up is FP32 and thus single precision FLOPS. Way up, at 90+%. FP16 texture fillrate is also up massively, though this has little relevance to CUDA. Memory bandwidth is identical, 192 GB/s, GTX 580 has a wider 384-bit memory bus. In more complex compute (CUDA/OpenCL/DirectCompute) applications GTX 580 is still faster. Perhaps Premiere Pro is not as complex, relying mostly on single precision brute force. Let's wait for benchmarks to see how well GTX 680 performs in Premiere Pro.

[Tom Lowe]

I'm with Jeff on that 4GB FTW card :)

[Petri Teittinen]

GTX 580 has 512, but it's not as straightforward. Kepler is a vastly different marchitecture to Fermi. Previously Nvidia's design philosophy was few high efficiency big shaders at high clocks (2 x core clocks). So, GTX 580's 512 shaders were 'hotclocked' at 1550 MHz. With Kepler Nvidia has adopted AMD's model - many small, dense not-as-efficient shaders at low clocks (AMD 7970 has 2048 shaders at 925 MHz). GTX 680 has 1536 shaders clocked at 1006 MHz. Kepler shaders are smaller and less complex, most notable FP64 has dropped to 1/24th rate versus 1/8th for Fermi. So, the numbers are not really comparable. However, what is up is FP32 and thus single precision FLOPS. Way up, at 90+%. FP16 texture fillrate is also up massively, though this has little relevance to CUDA. Memory bandwidth is identical, 192 GB/s, GTX 580 has a wider 384-bit memory bus. In more complex compute (CUDA/OpenCL/DirectCompute) applications GTX 580 is still faster. Perhaps Premiere Pro is not as complex, relying mostly on single precision brute force. Let's wait for benchmarks to see how well GTX 680 performs in Premiere Pro.

Very wise words, Subhadip.

[Jeff Kilgroe]

Why do folks adamantly recommend the 680 over the 580 for CUDA applications, without having tested them? Early benchmarks I've seen on general tech sites show it not being much faster than the 560 or 580 in general compute performance, due perhaps to the changing of the way they are counting the cores and prioritization of gaming performance over CUDA performance.

It will be interesting for sure. But one thing to note, even though they have taken a different approach with the Kepler architecture, the new GTX680 sits on a PCIe 3.0 interface, so there's larger data path running in and out of the card. The outgoing data path on the GTX680 card itself is a good bit larger than on the GTX580. Currently, the CUDA scores show it to be about the same in performance overall when compared to the GTX580 and a bit more erratic. That says a lot to me right there, as nVidia typically increases the performance out of their video cards, both in graphics and CUDA performance with subsequent driver releases.

If you need something for CUDA today that works. The GTX580 is the one. The 2GB may or may not be a bottleneck on the GTX680, depending on what you're doing. Personally, I'm waiting for the 4GB models with the higher clock speeds. They will be at least one more driver revision down the road when those ship too.