Micron’s overclocked GDDR6 graphics memory at 20 Gbps surpasses HBM

During the year, it’s time for a new generation of graphics cards, but those who have been waiting for something good for Computex 2018 will be disappointed. Once Nvidia launches Pascal’s successor, it will not only have a new architecture but also the next generation of graphics memory, which builds on the long-running GDDR5 and refined GDDR5X.

At the end of May, it was revealed that Nvidia had entered into an agreement with SK Hynix to supply GDDR6 type graphics memory. However, the South Korean memory manufacturer is not alone in the GDDR6, Samsung and Micron are also ready with their variants which also appear to be higher performing.

While SK Hynix’s product portfolio today stops at memories of 14 Gbps (14,000 MHz effective clock frequency), Samsung can handle up to 18 Gbps, while Micron previously stated that they performed tests of up to 16.5 Gbps. Now Micron is talking about the company’s GDDR6 memory can go even higher with a “small, but helpful” increase in voltage.

While the preceding results demonstrate full DRAM functionality up to as high as 16,5 Gbps, it is possible for the overall performance of an architecture to be capped by timing limitations in the memory array itself. To determine if this GDDR6 interface could extend beyond the 16,5 Gbps range, the device was placed into a mode of operation which exercises only the I/O while bypassing the memory array.

Without taking into account the memory controller in, for example, a graphics circuit, Micron managed to squeeze the memory up to a full 20 Gbps. Both AMD and Nvidia have previously talked about the difficulties in developing a memory controller that can handle high speeds with GDDR5 and GDDR5X, where Nvidia at most delivers speeds of 11.4 Gbps with the latter technology.

Assuming that the graphics card players can master GDDR6, 20 Gbps would mean substantial bandwidth with a narrow memory bus. A smaller memory bus means that the graphics circuit becomes smaller and thus cheaper to manufacture, which is why AMD and Nvidia normally only go over 256 bits for graphics cards in the enthusiast class.

As an example, a graphics circuit with a 256-bit memory bus and GDDR6 at 20 Gbps would provide a bandwidth of 640 GB / s – higher than all AMD’s HBM-equipped graphics cards and just behind Nvidia’s expensive Titan V. A not too unusual 384- bit bus width would mean 960 GB / s, that is almost 1 TB / s.

As GDDR6 is an evolutionary development of GDDR5X, the challenges of AMD and Nvidia remain with a memory controller that can handle high speeds. It is therefore likely that the first wave of GDDR6-equipped graphics cards will be much lower than 20 Gbps and more in line with SK Hynix’s product portfolio.

Source: Wccftech