Nvidia GK110 spec et discussion.

NVIDIA GeForce Kepler 110 (GK110) Specs Detailed

NVIDIA surprised everyone by introducing the biggest GPU in history, codenamed "GK110", first as a mission-critical Tesla HPC GPU accelerator part, instead of a GeForce consumer graphics part. The GK110 is driving NVIDIA's Tesla K20 HPC GPU accelerator, which was detailed by NVIDIA in its Tesla Kepler press release. We are, however, more interested in what goes into building one of these little monstrosities.

To begin with, NVIDIA crammed a mind-boggling 7.1 billion transistors into the GK110. It is essentially a 2x upscale of the GK104, while retaining its essential component hierarchy. The number-crunching machinery still consists of streaming multiprocessors (SMX), which pack 192 CUDA cores, each. It's quite plausible that the GK110 silicon packs 16 such SMX units.

NVIDIA disclosed that the Tesla K20 offers a memory bandwidth as high as 320 GB/s. This is a particularly interesting number, which provides hints to the chip's memory interface. Given that today's fastest high-volume GDDR5 memory chips are qualified for 6.00 GHz, NVIDIA would only be achieving 288 GB/s, if the memory bus width were 384-bit. Taking into consideration the fact that GPU vendors conservatively set the clock speeds of GPU and memory on GPU compute accelerators, it sounds plausible that GK110 has a 512-bit wide GDDR5 memory interface, and that NVIDIA is achieving 320 GB/s with a memory clock speed of 1250 MHz (5.00 GHz effective), on the Tesla K20.

Launching the GK110 first as a Tesla part serves quite a few purposes. To mention a few, GK110 is launched to the market in a low-volume yet high-margin market; NVIDIA doesn't have to worry about volumes or pressure to launch the chip in its GeForce avatar, since it already plugged high-end graphics market with a splurgy $999 (read: $1200-ish) GeForce GTX 690. It's possible that GK110 makes its march to consumer platforms once 28 nm manufacturing has achieved maturity (in terms of volumes, it's a 13-year old).

En résumé.

GK110

3072 cuda cores

512bit memory bus

?? gb GDDR5

320gb/s de bandwith mémoire.

GK104 GTX680

1536 cuda cores

256bit memory bus

2gb GDDR5

192gb/s bandwith mémoire.

On est une bonne coche au dessus d'une GTX680 !

Edited May 15, 2012 by Boulard83

oui mais $$$

C'est de GPU de calcul, pas fait pour gamer....

C'est de GPU de calcul, pas fait pour gamer....

les K10 et K20 oui, mais le GPU "mainstream" sura sur la même base de Kepler.

15 SMX au lieux de 16 ? donc 2880 cuda cores ? à suivre ... !