OTOY Forums

roeland wrote:Hello,

t_3 wrote:... so what about a nightly octane build?
just kidding

OK, you'll get your nightly build

We made a build with CUDA 4.2, for the Fermi and Kepler architectures. Licensed customers can find it in the RC testing forum.

--
Roeland

Leiurus wrote:I don't see anything "cool" here.

Leiurus wrote:@ t_3

So did I, all apologies if the brackets around the word has been misleading, a one-night rebuild is not cool, it's fantastic.

The very "uncool" thing is the direction Nvidia seem to take with their latest architecture products. Speculation it is indeed, I'm simply expressing my concerns and hopefully not predicting future.

They are influenced by a rather interesting thread related to this release I've read a few days ago, I can't spot it now as I'm not using my computer, I'll try to put an hand on it ASAP, would sincerely appreciate to have this forum members thoughts and inputs on it.

Fair Disclaimer: Yes, i know how hardware works low level, yes i'm not a "proper" coder, but i know my way with C, Lisp, VBA, PHP, and some Assembly code.

Guys, the Chip released by nvidia is the GK104. Nvidia starting with the Fermi family of chips (GFxxx) is doing chips specifically for the 3 platforms they sell: GeForce, Quadro and Tesla. The Chips ended with "0" are the flagship, and are most likely to be used with their high end parts (Quadro, Tesla and the High End Geforces)and the rest ending in a number 1-9 are designed to be used with the GeForce only, because are designed for gaming on mind. This makes the manufacture more economic and less prone to bad chips being produced. (Until now, you designed only one chip, and those who passed all tests were sold as High End , those who failed some tests were downclocked and have some internal parts disabled, and the most "burned out" were sold as entry level cards (cheapest).. This is what AMD and nvidia did for years).

But with Kepler, Nvidia decided that is more viable to create a Chip that is 100% oriented to gaming, another that is 100% oriented to Full OpenGL compilance, and another oriented for GPGPU computing. This way designs are simplified, thus gets cheaper to manufacture, the chips are less prone to failures, and drivers gets easier to code. Is a full win situation for Nvidia.

So, What have to do that with the fact that the GK104 (a.k.a Geforce 680) is slower on OpenCL and CUDA than the older counterparts? Simple: They upgraded the GAMING aspect of the Chip (literally in gaming they send AMD Tahiti to take vacations to the artic ) and to make sure the chip is viable, they simplified the GPGPU part of it, and powered the graphics part of it. This is highly technical so i'm not going to detail here (artists site, remember?? ). This allows 4 objetives: Better and cheaper Chip production, Focused and segmented market control, easier driver support for the chips, and capitalize more money per card. (Remember: THIS IS BUSINESS).

For the consumer, that may not be so great news, since this means you don't have anymore a product that excels to do 3 things at the same time, but 3 products that excels to do one thing and for the other 2 things, results are just average or low:

Geforce: Gaming card, requires good performance, quite some workaround for games, and good PhysX; doesn't require good GPGPU neither full OpenGL compilance, just DirectX compilance. Clocks are higher here.

Quadro: Professional Card, requires good performance, High precision drawing, full OpenGL and DirectX compilance, doesn't require good GPGPU, neither PhysX for gaming. Clocks are slower than gaming, but in applications gives more FPS and precision.

Tesla: It's all about GPGPU. The rest is secondary. Clocks tend to be slower here.

This is what is coming in the future for the GPU market. You want something, you pay for it. This is done because nvidia needs to protect their investments (they invested in raytracing, physics, gpu computing, and general purpose processing) and they expect to more people buys their Quadro and Tesla lines of products for GPGPU if you want to do these.

Of course AMD could change the game, but if you are a gamer, you sell the faster gaming card, not the faster GPGPU card. So is probably AMD will follow the same trend, since the gaming cards is what sells and produces more profits. Professional cards doesn't sell so much, but are not cheap.

TL;DR: It's all about business.