HANDS ON GUIDE

4 - Nvidia has a different perspective

Nvidia is calling its GPGPU efforts 'thread computing' rather than 'stream' and its CUDA API uses a standard C programming interface ..

Launch this Guide

Nvidia’s answer to ATi’s CTM GPGPU programming is called the Compute Unified Device Architecture (CUDA). However, CUDA is focused on the company’s 8000-series GeForce cards and latest Quadro FX 4600 and 5600 boards, all of which use a Unified Shader Architecture. Previous generations don’t support it. ATi’s CTM supports hardware from R580 (the Radeon X1900 series) onwards so its last two generations can be harnessed fully for GPGPU usage.

Nvidia is trying to differentiate its GPGPU implementation from the ‘other manufacturer’s’ by calling it thread computing rather than stream computing. The difference is that the graphics hardware processes instructions from cache rather than video memory, in groups of at least 32 threads at a time. Nvidia claims this is faster than ATi’s GPGPU, which must process through texture memory, although there aren’t any directly comparable benchmarks available to prove this. Another major difference with CUDA is that it uses a standard C programming interface. This makes it easier to program for than CTM.

Not to be outdone by ATi’s Stream Processor products, Nvidia has also now launched its own range of GPGPU-specific products called Tesla, which like Stream bear an uncanny resemblance to graphics cards. In fact, the underlying hardware is based on the same G80 core as the GeForce 8800, although the configurations are optimised for stream processing. Tesla comes as an individual card, an external box containing two cards (similar to the Quadro Plex external graphics system), or a complete 1U rackmount server.

Latest Hands on guides

These are the most popular features articles over the past three months.

Mobile Broadband

Compare prices

Fastest, cheapest 3G mobile broadband dongles from 3, Vodafone, T-Mobile and Orange
from just £10/month