This announcement might seem a bit anticlimactic now that Nvidia has already stated that the GeForce GTX 1080 will use GDDR5X RAM, but Micron has announced that GDDR5X is now in mainstream production, months ahead of its originally expected schedule.
One interesting thing about the GTX 1080 is that the RAM it uses is at the lower-end of the memory Micron said it could ship. GDDR5X has previously been touted as a solution capable of providing up to 12.0 Gbps of bandwidth in quad-data rate (QDR) mode. Nvidia, in contrast, has stated that the GTX 1080 uses just 10Gbps memory — towards the low-end of the scale. One potential advantage of starting at the lower end, however, is that the company has plenty of headroom to scale in the future.
Absolute bandwidth on the GTX 1080 is 320GB/s, which nearly matches the GTX 980 Ti while retaining the GTX 980’s 256-bit memory bus. Keeping the memory bus small is one way to control die size and lower production costs; large ring buses eat up die space and consume a significant amount of power.
We suspect that GDDR5X will be similar to GDDR4. GDDR4 was introduced by AMD in 2006 as a way to improve memory bandwidth and GPU performance. Nvidia never switched to GDDR4, preferring instead to retain GDDR3 until GDDR5 became available. AMD used GDDR4 in various products from 2006 – 2009, at which point it moved to GDDR5.
AMD hasn’t said whether or not it will use GDDR5X in upcoming products, but the fact that Nvidia is only using GDDR5X for the GTX 1080 suggests that the memory is still rather expensive; the GTX 1070 uses GDDR5. AMD’s Polaris is expected to target the mainstream PC market, which means it’ll likely use GDDR5 as well. Meanwhile, both AMD and Nvidia are planning late-year launches for next-generation products, both of which are confirmed to use HBM2.
The implication here is that GDDR5X’s market position long-term will depend on whether or not AMD and Nvidia see a need for the memory to offset higher costs in HBM2. If HBM2 costs come down relatively quickly, there may not be much room left for GDDR5X in the market, given that GDDR5 is produced by multiple companies and is relatively cheap (Micron, thus far, is the only company building GDDR5X). If, on the other hand, interposer costs stay high, we may still see all three standards in-market for some period of time.
What’s highly unlikely is that we’ll see any single GPU supporting both HBM2 and GDDR5X. Sources we’ve spoken to have indicated that supporting both memory standards simultaneously in the same silicon would be extremely difficult for minimal benefits. Supporting GDDR5 and 5X in the same silicon is much simpler than an HBM2 + GDDR5X configuration.