Product: Nvidia GeForce GTX 580
Company: nVIDIA
Authour: James Prior
Editor: Charles Oliver
Date: November 9th, 2010
Introducing GF110

At NVIDIA's GPU Technology Conference in September 2010, NVIDIA rather surprisingly disclosed a long term CUDA GPU roadmap that not only informed us of the codenames of the new architectures through 2015, but their expected time frames as well.

Tesla; Fermi; Kepler; Maxwell
Tesla; Fermi; Kepler; Maxwell

Given this information, it should come as no real big surprise that the GeForce GTX 580 is based on the Fermi architecture - an existing product design currently in market. We've seen this type of revamp before, with the 'Tesla'' architecture 8800GTS -> 9800GTX -> GTS 250 revamps. However, this new branding of Fermi is a little difference - there's a much larger performance increase being touted, with better performance/watt claimed also.

GF110 - Fermi
GF110 - Fermi

NVIDIA came under a lot of fire for their initial Fermi offering, despite offering the fastest single GPU consumer graphics card on the market. The GeForce GTX 480 was criticized for being hot and noisy  - and incomplete, with one Streaming Multiprocessor and its corresponding 32 CUDA cores, 16 ROPs and Polymorph engine disabled. Despite the 700MHz core clock and 924MHz memory clock being the only announced speeds, there were accusations of being slow and underperforming - despite demonstrated game and consumer application performance that exceeded the competition.

With the GeForce GTX 580 no such accusation can be levied - not only are there more CUDA cores, there are no fused-off parts (at least, not that are part of the public design parameters for performance) and bumped clocks - 10% higher in standard form for the engine, and ~8.5% increased on the memory clock. This is not simply the result of 8 months of binning, this new chip is the result of a design re-work using NVIDIA's and TSMC's experience from making GF100.

Micro-Architecture Improvements

Additionally, per core micro-architecture improvements are in place - if the GPU from the GTX 580 was reduced to 480 cores, at 700MHz (like a vanilla GF100 found in GTX 480), it would still be between 5 and 12% faster than GF100. This is due in part to NVIDIA implementing full speed FP16 texture filtering, that was previously reduced on GF100; and improved Z-cull, which aids in heavily tessellated scenes. Combining the higher clocks and increased SM count, performance increases of up to 30% over GTX 480 are claimed. The extra Streaming Multi-Processor is important, as it not only adds 48 CUDA cores, but also 16 ROPs and a tessellation engine (Polymorph). Tessellation is one of NVIDIA's strongest areas, with their geometry throughput far exceeding AMD's Cypress. Like GF100, Double Precision is supported but not at 'full speed'; it's not relevant to current consumer applications, but at least it's there for hobbyist developers to play with.

It should come as no surprise that NVIDIA looked at their Polymorph engine and found room for improvement - tessellated scene performance is increased, when combined with the Z-cull tweaks and additional Polymorph engine (and bumped clocks). NVIDIA claim up to 11x the performance of AMD's Cypress GPU, meaning unless AMD has done something truly remarkable with Cayman, the tessellation beat down will continue. Like AMD, NVIDIA advocates adaptive tessellation, but their designs are less sensitive to high tessellation factors that produce small polys - HAWX 2's 6px triangle size will be a good demonstration of this.

As you can see, there are twelve memory chips located on the front of the board, two per memory controller. There also appear to be eight power phases, six for the GPU core and two for the memory bus and chips themselves. Physically, the GeForce GTX 580 GPU is very striking, thanks to it's large size and imposing heat spreader.

content not found

Copyright 2022 ©

You may not use content, graphics, or code elements from this page without express written consent from

All logos are trademarks of their original owners. Used with permission.