SK Hynix has shown their first GDDR6 wafers and specifications during the GTC 2017 event. The new standard will allow for twice the bandwidth per pin as well as 10% lower power consumption. So far SK Hynix appears to be ready to produce GDDR6 memory chips using 20nm process technology. These chips offer 16Gbps of bandwidth per pin.


The GDDR6 chips shown at GTC come with a capacities of 8Gb (1GB) per chip and 16Gb (2GB) per chip. Therefore it’s possible that GPUs combined using GDDR6 memory will offer higher memory capacities. Other than that the memory bandwidth is also going to be increased by a factor of two. This will allow GPUs with a 128-bit memory bus and GDDR6 memory to offer the same bandwidth like a GPU with 256-bit bus that’s been combined with GDDR5 memory.

These changes will allow mid-range GPUs to be created with a reduced power draw and cost due to lower power requirements and component count. It looks like at launch GDDR6 will offer maximum bandwidth of 16Gbps while lower-end models might feature 12-15Gbps. Just for reference, the NVIDIA Titan X Pascal, which is based on GDDR5X chips, offers 11.4Gbps memory bandwidth.

GDDR6 will be a big step forward for the industry. Nevertheless there is also HBM2, if it should really become widely available for the high-end market. Still HBM2 will have the significant downside that it’s much more complex to implement than GDDR6, requiring direct to chip communication using an interposer. The upside on the other hand is even higher bandwidth due to lower latencies.




Source: OC3D