Farewell 28nm, It's Been a Long Journey
Things used to move further and faster back in the early days of GPUs. Looking back to the year 2000 we had the GeForce 2 Ultra, an epic product of its time built using the 180nm manufacturing process. The following year the GeForce 3 series moved down to 150nm. It would be just 2 more years before the move to 130nm was made and the 110nm process a year later.
In fact, things continued to move pretty quickly until just four years ago, when we made the transition from the 40nm design process used by Nvidia’s Fermi architecture to the 28nm process, still used today. This extended development cycle has had AMD and Nvidia squeezing the absolute most out of the 28nm design process.
As good as the Titan X and Fury X are, it is time to move on in the quest for greater efficiency and even greater performance. First to market with a true next generation GPU is Nvidia. Codenamed Pascal, this latest architecture promises big things and could very well be the biggest step we've seen in recent years.
Leading the charge is the new GeForce GTX 1080. Built using the 16nm design process and packed with faster GDDR5X memory, it promises to put away the Titan X while consuming less power than the 980 Ti. Nvidia claims the GTX 1080 is three times more power efficient than Maxwell.
Following a few weeks later will be the more affordable GeForce GTX 1070. Though only certain details are known about that card (update: full specs here), for now we've learned its cut down core will still be good for 6.5 TFLOPS, it'll come armed with an 8GB VRAM buffer running on more typical GDDR5 memory. So in spite of the downgrades, the GTX 1070 is poised to become the hotter commodity considering its more attainable $380 price tag.
Actually, pricing is something to consider with these new Pascal graphics cards. For the first time consumers will be asked to pay a price premium for Nvidia’s reference card, now referred to as the "Founders Edition." Previously reference cards could be had for the base MSRP, while custom partner boards often came at a price premium.
Built using the 16nm design process and packed with faster GDDR5X memory, the new GeForce GTX 1080 promises to put away the Titan X while consuming less power than the 980 Ti.
This time around the GeForce GTX 1070 "Founders Edition" will be available for no less than $450, going on sale June 10th. Custom boards from partners are expected to start at $380 and will become available at a later date.
The same is true for the flagship GeForce GTX 1080, with initial sales limited to the Founders Edition at $700 from May 27th onwards. Eventually the GTX 1080 will become available from board partners for as little as $600.
Nvidia feels their reference design is the best version possible of the GTX 1080 and GTX 1070 and they intend to have their design on the market for the life of the card, which is why you'll be able to purchase Founders Edition cards from Nvidia directly.
GeForce GTX 1080 in Detail
The GeForce GTX 1080, codenamed GP104-400, features a surprisingly small GPU - the die size is just 314mm2, roughly half that of the Titan X and 980 Ti. There are still 7.2 billion transistors crammed inside, just 10% fewer than the most complex Maxwell GPUs.
Nvidia has been able to squeeze so much into this small space thanks to the use of the 16nm FinFET process. This means Nvidia is now able to produce more complex GPUs with greater core counts, though that isn’t exactly what they have done here with the GTX 1080.
As with CPUs, one of the key defining aspects of GPU performance is core count. The GeForce GTX 680 had 1536 cores, or CUDA cores, as Nvidia calls them. The original Titan featured 2688 CUDA cores, while the Titan X received a whopping 3072 CUDA cores.
Nvidia tells us that the GTX 1080 is a good bit faster than the Titan X, so you might expect a similar or larger core count. To everyone's surprise at the heart of the GTX 1080 are just 2560 CUDA cores, 17% less than the Titan X and 9% less than the 980 Ti. Thus Pascal is not just more efficient, but thanks to the newer process, it's received a massive boost in frequency speed at which those cores can operate.
Compared to the Titan X and 980 Ti, the GTX 1080 runs its cores up to 61% faster, with a base clock speed of 1607MHz and boost as high as 1733MHz.
Further, the GTX 1080 only requires a single 8-pin PCI Express power connector with a TDP rating of just 180 watts, so on paper its power consumption is just 9% higher than the super-efficient GTX 980.
There was just one more ingredient that these high frequency cores required in order to get the job done: very fast memory. The solution here isn’t the hotly anticipated HBM2, but new GDDR5X memory. GDDR5X is designed to run at speeds of 10 - 14Gbps, which is twice that of GDDR5. In order to achieve this, the chip interface has been widened from eight nibbles (32 bit) to sixteen nibbles (64 bit).
The GTX 1080 Founders Edition we received to test came equipped with Micron's GDDR5X memory running at 10 Gbps, which is essentially 5GHz! That is a 43% increase over the GTX Titan X and GTX 980 Ti, both of which use GDDR5 memory running at 7 Gbps (1.7GHz).
Although memory is considerably faster, the 384-bit memory bus of the Titan X and 980 Ti has been downgraded to the same 256-bit bus used by the GTX 980 and thus limited to a 320GB/s memory bandwidth.
With performance expected to exceed even that of the Titan X, we can assume that the 1080 is going to be a bit starved for memory bandwidth, which means overclocking could be very beneficial assuming GDDR5X memory can be pushed higher (you can bet we'll try). So let's move on into testing.