“Particularly exciting is the GTX 1080, the most advanced gaming card ever created and the first card in the world to use Micron’s GDDR5X memory,” Kido said in his blog. “This card pumps out 9 TFLOPS with 2,560 cores offering an unprecedented level of graphics performance and efficiency.”
The upcoming card will feature a 256-bit memory bus and quad data rate GDDR5X memory moving at 10Gb/s, providing 320GB/s of bandwidth. It’s slated as the world’s fastest video card memory to date, and is the result of “thousands of hours” of teamwork between Micron’s Graphics memory engineers in Munich, and Nvidia’s own hardware team.
GDDR5X is essentially an extension of the current GDDR5 standard. It improves on the older memory by using a quad data rate bus instead of a double data rate bus, thus doubling data transfer amounts without raising the memory clock speeds. GDDR5X also prefetches two times more data than GDDR5, up to 16n (64 bytes) from 8n (32 bytes), and has better energy efficiency due to lower I/O and supply voltages.
Micron reportedly began sampling GDDR5X chips at the beginning of the year, which operated at 10Gb per second, 11Gb per second, and 12Gb per second (although GDDR5X can theoretically do 16Gb per second) per pin. Nvidia, it seems, is taking the safe route by using the slower GDDR5X chips in its upcoming GTX 1080 card. The chips also offer more breathing room on the graphics card, as they come on 190-ball FBGA packages measuring a mere 14mm x 10mm.
The first GDDR5X memory chips seen supplied by Micron had an 8Gb (1GB) capacity, a 1.8V pump voltage, supply and I/O voltages measuring 1.35V, and was manufactured using 20 nanometer process technology. The overall estimated DRAM power consumption used by the 8GB of GDDR5X memory on Nvidia’s card will supposedly be around 20 watts.
“GDDR5X represents a significant leap forward for high end GPU design,” said Mian Quddus, chairman of JEDEC’s board of directors. “Its performance improvements over the prior standard will help enable the next generation of graphics and other high-performance applications.”
During Nvidia’s Pascal reveal event last week, CEO Jen-Hsun Huang showed what data looked like as it moved from the GPU to the GDDR5X memory (or G5X as he calls it). He said that just one signal is 100 picoseconds wide, which equals to the length of time light takes to travel one inch. He added that in that little bit of time, Nvidia’s engineers have to make sure the data is indeed transferred between the GPU and the memory. When you look at all of that data in an oscilloscope, it’s all “just noise,” he said.
The world’s first graphics card to use GDDR5X memory, the GeForce GTX 1080, will be packed with 7.2 billion transistors, 2,560 Nvidia CUDA cores, and feature a base clock speed of 1,607MHz and a turbo clock speed of 1,733MHz. Other technical details supposedly include 2,560 unified shaders, a pixel fill rate of 102.8 GPixels per second, and a texture fill rate of 257.1 GTexels per second. The card is slated to hit the streets on May 27, 2016.
