NVIDIA's latest Fermi core has 3 billion transistors.
NVIDIA sparked a mini revolution in the graphics world when it released the 8800GTX some time ago, and while it was superceded with last week's 5870 launch it seems they're doing their darndedst to get back into the spotlight.
Details have been released about their latest generation core, codenamed Fermi, that the next generation of gaming cards will be built around. While the model described is the Tesla compute-oriented product, gaming tech will sport similar features.
It boasts a phenomenal 3 billion transistors on a 40nm manufacturing process, much more than twice the current transistor count of NVIDIA's own GT200 core at 1.4 billion (and even higher than AMD's RV870 core at 2.15 billion).
These new transistors bring about a huge amount of stream processors, newly renamed to CUDA Cores and bumped up to a total of 512 over the current 240 in the present cards.
Also finally making an appearance is GDDR5 memory on a very wide 384-bit memory bus, with a theoretical maximum of 6GB usable to the card, depending on clockspeed it will be a higher bandwidth than AMD's offering.
In the die shot above the major changes can be seen quite clearly in the very logical layout of the core; with sixteen rectangles each designating a SM cluster with 32 CUDA Cores within, a huge L2 cache sitting in the middle strip and DRAM interfaces all around the edges.
With all these transistors and new features being added the core itself is going to be physically large, even considering the process drop to 40nm - but exactly how large and how hot remains to be seen.
However the gaming performance should definitely stack up if these specs remain unchanged for the highest end gaming card, and could potentially be higher considering the architecture tweaks and possible clockspeed increases.
It looks like the GPU scene just heated up all over again.
Issue: 133 | February, 2012