This a decent GPU for rendering with Indigo?

[Stinkie]

I'm looking into getting a new pc, and I know squat about hardware.

http://www.evga.com/Products/Product.as ... P4-2662-KR

[lycium]

I'd suggest getting a GeForce 580 Ti, with as much memory as you can The newer GeForce series, 6xx, aren't as good as the previous generation for GPU compute.

[Stinkie]

Hm, good to know. Thanks.

[suvakas]

There is GeForce 580 Ti
Or did you mean 560 Ti?

[lycium]

Sorry, I did indeed mean the 560 Ti. It has a better price/performance point than the 580, and probably doesn't produce nearly as much heat.

That card has been around for aaaages, even longer when I think that it's not very dissimilar from my GeForce GTX 470 which is now 2.5 years old

[suvakas]

I'd recommend 448 core version of 560 Ti. Faster than regular 560 Ti.

[Stinkie]

The newer GeForce series, 6xx, aren't as good as the previous generation for GPU compute.

In what sense, actually? Speed?

[pixie]

Personally I'm not much fan of single PT, so I'd rather go for a solution that could give me more cpu power, unless bidir PT cames for GPU, but I would wait till this happen since gfx board will eventually go lower in price.

[Juju]

The newer GeForce series, 6xx, aren't as good as the previous generation for GPU compute.

Any links to benchmarks / testing to verify?

[lycium]

This has been reported on a number of hardware sites testing CUDA and OpenCL performance, including this one from the excellent TechReport review:



AnandTech have similar benchmarks, and I hear it a lot from people who use Arion, Octane and so on.

It makes sense to me because there are a lot more processing "cores" for the same amount of onboard memory (cache, registers, ...) so memory-intensive kernels, such as those used in rendering, will simply be more resource-constrained and less able to hide memory latency.

[Juju]

aah, thanks.

I thought the 6xx GPU's had many more CUDA cores but running at lower speed, thus lower power usage and more computing power?

Comparison from nVidia info on specifications for models below:
- GTX 680: 1536 CUDA cores at 1006 MHz base clock, 128.8 bn/sec texture fill rate, 256-bit memory interface.
- GTX 580: 512 CUDA cores at 2004 MHz base clock, 49.4 bn.sec texture fill rate, 384-bit memory interface.
- GTX 480: 480 CUDA cores at 1848 MHz base clock, 42 bn/sec texture fill rate, 384-bit memory interface.

on paper: GTX 680>GTX580>GTX480.

Looks like I'll have to troll through some sites to follow up on these benchmarks and testing?

[lycium]

My synopsis is that the 6xx series is optimised for games, and generally simpler kernels. There's no question that it's way more power efficient for such tasks, and has an optimised memory controller etc, however we're interested in a very different application (complex compute kernels).

PS. Where in SA are you from? I used to live in Pretoria, then moved to Cape Town

[Juju]

PS. Where in SA are you from? I used to live in Pretoria, then moved to Cape Town

I'm from Cape Town northern suburbs, I live in Durbanville and work in Stellenbosch.

BTW, you do know that the comparison you refer to benches the OpenCL performance? That is AMD implementation, nVidia has CUDA implementation. Which is more efficient?

[Zom-B]

BTW, you do know that the comparison you refer to benches the OpenCL performance? That is AMD implementation, nVidia has CUDA implementation. Which is more efficient?

OpenCL isn't a "AMD implementation" but a free standard that is suported by AMD, Nvidia GPUs and some newer Intel GPUs too. All moder CPU support it too! Support for Mobile ARM devices is comming... Xeon Phi (Knight Corner or Intel MIC) will also work with OpenCL.

CUDA is Nvidia only and has therefor no real future...

[Juju]

O I C

looks like I'm seriously out of date with the latest tech...