The GTX 460 is the latest addition to Nvidia's DirectX 11 compatible range of graphics cards. Aimed squarely at the mid-range of the market, it should be a great buy for those looking for a decent amount of bang for their buck. Let's find out is it lives up to expectations.
Nvidia's first set of DirectX 11 graphics cards didn't exactly set the world alight. They were all based on a new architecture called Fermi and used a new chip based on that design, the GF100, that on paper looked very impressive. Unfortunately for Nvidia it didn't quite live up to its billing. Heat, power consumption, and primarily manufacturing problems plagued the GTX 480 and GTX 470 and neither delivered the performance to justify these compromises. What's more, Nvidia charged extortionate prices for them (£450+) when they first came out – it really wasn't a good episode for the company.
A few months later it released the GTX 465, which used a heavily cut down (i.e. bits disabled) version of GF100 that was meant as a stop gap to tackle the mid-range market. However, it also failed to deliver on all fronts.
Now Nvidia has finally released a card based on a brand new chip, designed from the ground up to be smaller and less powerful but that should deliver in terms of all the other aspects of a graphics card; heat, power consumption, noise, and pricing. That card is the GTX 460 and the chip it's based on is dubbed the GF104.
We won't cover the detailed inner workings of the Fermi architecture in this review (for that we refer you to our review of the GTX 470) but we will give you a brief overview of the differences between GF100 and GF104.
GF100 contains 512 Stream Processors (SP), or Cuda cores as Nvidia likes to call them, which are the main processing units of the chip. These are split up into sets of 32, in what's called a Streaming Multiprocessor (SM). This also adds four texture units, some cache, thread schedulers, and the polymorph engine, which handles geometry processes like tesselation.
Four of these SMs are then clustered together to form a General Processing Cluster (GPC), which basically just adds a raster engine (the bit that converts 3D models into 2D pixels). Finally, four of these GPCs are added together along with further thread scheduling components, the host interface, memory controllers, ROPs (48 of them split into six sections), and 512KB of L2 cache. The result is one enormous and hot chip that contains a whopping 3.1bn transistors. However, due to production problems, the GTX 480, has one of the SMs disabled so you actually get 480 SPs, 15 SMs, and 60 texture units.
For GF104, all the basic building blocks are very similar, but their proportions have been tweaked. So, you get 384 SPs, split up between just eight SMs, giving you 48 SPs per SM. To deal with these extra SPs, each SM also has double the number of texture units. Each SP has also had a second dispatch port added for more efficient thread handling. Finishing things off, GF104 has 32 ROPs, split up into four sections. The result is that despite at first glance looking like it is only half the chip that GF100 was, GF104 is considerably more than this – something that's reflected in its transistor count of 1.95bn.