GDDR5 memory has become something of a standard for high-end graphics cards. Some more economical options use DDR3, and we have seen many advances in HBM2, but the people at SK Hynix announced a new GDDR6 chip of eight gigabits. Its speed is 16 Gbps per pin, which translates to a chilling 768 gigabytes per second with a 384-bit bus. The first product to use GDDR6 will be Nvidia's Volta architecture, whose launch is scheduled for 2018.
Thinking about buying a graphics card? We are definitely going through a good time. AMD and Nvidia keep trading blows, and the so-called “mid-range” is nothing mid, to be honest. All you have to do is find the best price-to-performance ratio, prepare your computer, and take the plunge. If your budget has no limits, there are always things like the Nvidia Titan Xp, and there are many AMD fans out there waiting to see what RX Vega offers. It gets a little hard to think about the future of graphics cards when their present has such high-performance examples, but we already know how this works. 4K has fully established itself; however, several companies have begun early experiments in 8K, and the graphics cards that must process that content will have to chew pixels like candy. Optimizing architectures is always useful, but what about memory...?
At this moment, what is best on the market is based on HBM2, GDDR5 and GDDR5X, but SK Hynix announced a new GDDR6 memory chip. The chip has a capacity of eight gigabits, with a maximum speed of 16 Gbps per pin. The final parameters depend on the manufacturer's needs, although as an example, SK Hynix said it equates to a bandwidth of 768 gigabytes per second over a 384-bit bus. If we compare that number with the GDDR5X memory on the Titan Xp, the new GDDR6 has a 220 GB/s advantage. As if that were not enough, the operating voltage will be ten percent lower.
The strongest candidates to use Hynix's GDDR6 memories first are the cards of the Volta architecture that Nvidia will launch in 2018. Volta lowers its manufacturing process to 12 nanometers, and adoption of HBM2 is also expected in some models.
