Newly Ratified JEDEC GDDR5X Specification Doubles Bandwidth To 14 Gbps

We have been hearing about the next iteration of DDR for graphics for the past handful of months, and now, the folks at JEDEC have made it official. Called GDDR5X (the full standard is called JESD232 GDDDR5X SGRAM), this memory is designed to be twice as fast as GDDR5 per clock, boasting total throughput rates of 14Gb/s per pin.

GDDR5X Example Clock Diagram

With GDDR5, peak bandwidth was 8Gb/s, which required the memory to be run at 2GHz. NVIDIA's current top-end cards peak at 7Gb/s, with clock speeds of 1.75GHz. With GDDR5X, we can see performance of 10Gb/s - 2Gb/s better than the previous max - at 1.25GHz. That's 70% of the frequency for 125% of the performance - not too shabby.

At the top end, GDDR5X can push 14Gb/s at 1.75GHz, which compared to NVIDIA's current top-end offerings, is in effect exactly double the performance (since we haven't seen 2GHz/8Gbps cards up to this point.)

Micro GDDR5X Performance

One thing that's important to note is that GDDR5X isn't going to be a replacement for HBM "3D" memory, as it's able to push far better performance than even GDDR5X (especially once second-generation hardware gets here.) Instead, GDDR5X could act as a direct replacement for the memory in any graphics card that ranges from the absolute bottom up to the ~$300-ish mainstream area. That would leave HBM2 exclusive to top-tier cards, where it could likely be taken better advantage of.

It's anticipated that Micron will be making a formal announcement at some point this year about shipping GDDR5X to customers, although it's unclear at this point when we'll see products hit the shelves with it under the hood. Rumor has had it that we'd see such parts in mid-2016, although that seems unlikely at this point that we'll see it quite so soon.

For those wanting to rock the high-end parts of the next generation, GDDR5X isn't likely to affect you at all. But if you're looking forward to a solid mid-range part, it seems you can rest assured that your memory bandwidth will not be holding you back.