NVIDIA GeForce GTX 460 Review - GF104 and the budget Fermi
More than just a shrink
The Need for GF104
UPDATE: The new GTX 460 is already for sale at Newegg starting at $199 - check it out after reading our review!!
It is a normal trend for us to see both NVIDIA and ATI take their current generation GPU designs and shrink them down for lower cost and lower performing markets. For example, we first saw the G84 GPU from NVIDIA (known as the 8600 series of GPUs) just about 5 months after the release of the G80-based GeForce 8800 cards. The dies are smaller, cheaper to make, and more power efficient, as a general rule of thumb. The same general trend is seen with the AMD/ATI cards as well but on slightly differing schedules.
This time around, the GF104 design, a remake of the GF100-based Fermi GTX 400-series cards, is much more critical for NVIDIA. While the GF100 is a powerful GPU it has some critical deficiencies when put head to head with the Radeon 5000 cards from ATI; it is larger, hotter and just generally hard to make. Because of that we know that NVIDIA's margins on GF100 are not where they want them to be and the highly competitive market they current sit in makes the financials all that more difficult.
GF104 and the new GeForce GTX 460 are an attempt to address these issues by offering a smaller version of the Fermi architecture from the outset rather than take larger GPUs and disabling portions to sell at a lower cost. GF104 should be smaller, cheaper, and be more easily produced at much higher yields. At least, that's the theory.
Let's see what the new $199 (768MB) and $229 (1GB) Fermi cards from NVIDIA are made of.
A surprise shift in the new GPU
The new GeForce GTX 460 and the GF104 not only have fewer CUDA processing cores than any other Fermi card out there (336 to be exact) but much more has changed under the hood than that. As it turns out, NVIDIA has taken this as a chance to reorganize the architecture at a pretty basic level.
Also worth noting is that just like the GF100, the GF104 is being released with one SM completely disabled - the remaining seven SMs add up to the total 336 CUDA cores.
Most interesting is the fact that the PolyMorph Engines are now balanced quite differently with one per 48 SMs rather than one per 32 SMs. Remember that NVIDIA has been highly touting its advantage in tessellation performance and games that use the technology though this change moves the tessellation performance per CUDA core value down some. While the GTX 480 saw 30 cores per PolyMorph Engine the new GTX 460 will see 41.5 cores per tessellation engine.
For reference, here is a GF100 GPU with the same quarter size comparison
With such a dramatic shuffling of the GF100 architecture it seems obvious that NVIDIA found a couple of things worth changing as they built the GF104 chip for use in the GTX 460. By reducing the number of PolyMorph Engines per CUDA core in the GPU NVIDIA has lowered the amount of relative tessellation performance in comparison to the GTX 480 and GTX 470. I doubt they have lowered it too much but NVIDIA obviously thought the tessellation engines were idling a bit too much so by increasing the number of cores for shader processing per PolyMorph the balance should be shifted in the other direction.
This type of adjustment happens pretty often as GPU companies move from process node to process node or between redesigns like this. NVIDIA is simply adjusting its estimates for GPU performance and utilization across games, GPGPU functionality, etc, in hopes that they will find better power efficiency in the long run.