NVIDIA's big spring launch is the other green thing to look forward to, the new GeForce GTX TITAN X. It's surprising how NVIDIA could go on to build not just its performance-segment new-generation "Maxwell" chip (the GM204), but also its enthusiast-grade one on the existing 28 nanometer node, which is due to delays in the implementation of the newer node at TSMC. The GM204 in its GTX 980 avatar showed us that Maxwell was no joke, and that even with existing silicon fabrication nodes, NVIDIA could bring about more performance-per-watt gains than you'd expect from a shrink to a new node, maybe even 20 nm. If NVIDIA would have told us that the GM204 is built on 20 nm, we would have had every reason to believe them.
It's only natural then that NVIDIA will keep up the momentum with the tanker-loads of cash it earned with the GM204 to develop the big "Maxwell" chip using whatever is available (i.e. being stuck with 28 nm due to factors beyond its control). The fruition of that idea is the slick new GM200 silicon that is making its consumer debut with the company's new flagship product, the GeForce GTX TITAN X reviewed today.
With the GTX TITAN X, NVIDIA is continuing its product model of selling an extremely highly priced "have-it-all" product, much like Intel sells its have-it-all Core i7 Extreme CPUs at a similar price-point. The trend started with the 2013 debut of the GTX TITAN that also saw a refresh in the GeForce GTX TITAN Black. NVIDIA cleverly named the dual-GPU product based on that silicon the GTX TITAN Z (does that sound like 'Titans?'), and we're guessing that the TITAN X sounds like 'Titan-next' in some accents. Certainly So-Cal.
The GM200 could wind up being a lesson in VLSI design if the GeForce GTX TITAN X ends up meeting the performance-per-watt trends set by its smaller GM204-based siblings because it features a jaw-dropping 8 billion transistors in a square 601 mm² die, the biggest ever on the 28 nm node. NVIDIA gave it the same exact TDP rating of 250W as the previous-generation, which peaks our interest.
So what's all the fuss about? Roughly 50% more number-crunching muscle than the GTX 980, a 50% wider memory bus, and three times the memory. The GTX TITAN X is the first single-GPU consumer graphics card to breach the 10-gigabyte onboard memory mark by offering 12 GB of it. You get 3,072 CUDA cores based on the "Maxwell" architecture, 192 texture memory units (TMUs), 96 raster-operations units (ROPs), and a 384-bit wide memory interface. The GPU clock is a little over 1 GHz, with an additional 8% GPU Boost to frequency; while the memory ticks at 7.00 GHz, which gives the chip 336 GB/s of memory bandwidth. It may not look like an improvement over the previous generation, but then NVIDIA's new lossless texture-compression magic steps in, which NVIDIA claims to improve "effective" bandwidth by around 15%.
GTX 780 Ti
GTX Titan X
|Shader Units||2560||2688||1664||2816||2880||2048||2x 1536||3072||2x 2816|
|ROPs||64||48||56||64||48||64||2x 32||96||2x 64|
|Graphics Processor||Hawaii||GK110||GM204||Hawaii||GK110||GM204||2x GK104||GM200||2x Hawaii|
|Transistors||6200M||7100M||5200M||6200M||7100M||5200M||2x 3500M||8000M||2x 6200M|
|Memory Size||4096 MB||6144 MB||4096 MB||4096 MB||3072 MB||4096 MB||2x 2048 MB||12288 MB||2x 4096 MB|
|Memory Bus Width||512 bit||384 bit||256 bit||512 bit||384 bit||256 bit||2x 256 bit||384 bit||2x 512 bit|
|Core Clock||947 MHz||837 MHz+||1051 MHz+||1000 MHz||876 MHz+||1126 MHz+||915 MHz+||1000 MHz+||1018 MHz|
|Memory Clock||1250 MHz||1502 MHz||1750 MHz||1250 MHz||1750 MHz||1750 MHz||1502 MHz||1750 MHz||1250 MHz|