GTX 680 Performance (vray RT) & DirectCompute

[Delineate]

Has anyone tested the GTX 680's? Can anybody confirm the speculated poor performance for Vray RT?

Does Vray RT use DirectCompute? and can gtx 680's poor DirectCompute performance potentially be a driver issue?

(desperate to know, 3 cards are arriving tomorrow, and could potentially exchange them last minute for 3 gtx580 or 2 gtx590's)

thanks thanks thanks

[Delineate]

http://vr-zone.com/articles/nvidia-k...te-/15332.html

[Delineate]

Single Precision FP on the 680 doesn't look so bad. Does Vray RT even use double precision floating point?

[vlado]

Does Vray RT use DirectCompute?

No.

Best regards,
Vlado

[vlado]

Does Vray RT even use double precision floating point?

At the moment, no.

Best regards,
Vlado

[Jiri.Matys]

GTX680 is strictly gamers card without a lot lot LOT of things for GPGPU computing. This is why the card is cheaper and colder with smaller chip. There will be a full GTX 6xx card with full support of GPGPU computing later, I think after one or two months.

[Delineate]

Just cancelled 3 GTX 680's which were meant to arrive tomorrow. Not enough info out there, & lots of pessimism.
Still no one benchmark it in vray RT with the Indoor benchmark scene?

[Paul Oblomov]

In GK104 cuda cores are simplified a lot. Waiting for GK110

[vlado]

So here are some preliminary results with the CUDA code compiled with the 4.2 rc2 CUDA toolkit (not the final version yet) for the benchmark scene; OpenCL code is compiled by the respective driver. This is our latest GPU code, which is somewhat more complex than before.

GeForce 680 GTX/Driver 301.10
OpenCL kernel: 3m 39s
CUDA kernel: 3m 3s

GeForce 580 GTX/Driver 301.10
OpenCL kernel: 3m 15s
CUDA kernel: 2m 57s

GeForce 580 GTX/Driver 296.10
OpenCL kernel: 3m 15s
CUDA kernel: 2m 57s

Best regards,
Vlado

[Dariusz Makowski (Dadal)]

Basicaly its disaster hehe.... curse u nvidia !!!!

[vlado]

Basicaly its disaster hehe.... curse u nvidia !!!!

On the bright side, at least with the 4.2 toolkit the CUDA code is actually faster than OpenCL

In any case it seems to me that nVidia wanted the 680 to have roughly the same compute performance as the 580...

Best regards,
Vlado

[Dariusz Makowski (Dadal)]

I have bad feeling they want to cut off GTX GPU rendering sollutions from happening. I wonder if it will kick them back in the ass since a lot of people simply wont buy quatro/tesla and go to ATI - once they finnaly start working...

And yea the CUDA works very nice !, on a side note is the Shellac mode in Vrayblend working with 2 materials or I'm blind? =D it looks like it work mrrrrr...

[Delineate]

so with this cuda method there's no waiting time at the start when compiling?

[vlado]

so with this cuda method there's no waiting time at the start when compiling?

No waiting; the GPU code is precompiled so it starts right away.

Best regards,
Vlado

[Delineate]

that's great, Would be cool to compare the amount of time it takes to render the scene from the moment you press render... to get an overall time taken (including compile time for openCL) not so much to compare 580 to 680 but to compare overall openCL to CUDA.

as for the 680, the 1500 cuda cores seems to be nothing but marketing.