Earlier this year, AMD became the first GPU vendor to roll out graphics cards that utilized HBM, or High Bandwidth Memory. We’ve discussed this and other new memory technologies extensively over the past 10 months, and there’s wide agreement that HBM is the GPU RAM of Tomorrow for both AMD and Nvidia. What’s less clear, however, is how quickly HBM and its successor, HBM2, will waterfall down the product lines. Now there’s a rumor that Nvidia could adopt a new type of GDDR5, GDDR5X, for some of its Pascal GPUs in 2016.
Back on September 1, Micron made a pair of announcements concerning the future of its GDDR5 development. First, the company stated that it had developed new 8Gbps GDDR5 modules, which would allow companies to double RAM densities without adding more RAM chips. That’s important relative to scaling concerns, since one of the problems with GDDR5 scaling is that you can’t just keep embedding more memory around the GPU. There’s a minimal amount of room between traces to keep the electrical signaling clean.
Second, and more interesting for our purposes, was Micron’s announcement that it had created a new type of GDDR5, dubbed GDDR5X, that would offer 10-14Gbps signaling rates. Existing transfer rates for GDDR5 are 6-8Gbps, which is why this is being spun in some circles as a “doubling” of GDDR5 performance. According to micron, it can do this by doubling the amount of data that GDDR5X prefetches, thereby allowing for devices with as much as 16Gbps worth of throughput.
Could Micron use GDDR5X to slow HBM adoption — and will Nvidia adopt it as a solution for next-generation hardware? We’re dubious in both cases.
The GDDR5 Dilemma
Micron’s month-old announcement is getting press-time because there’s a kernel of truth to all the GDDR vs. HBM rumors. For all of HBM / HBM2’s benefits, these are new memory technologies, and new technologies typically command a premium compared to long-established alternatives. AMD confined its first-generation of HBM products to the upper end of the market, where it could command a commensurate price for the technology. Next year, HBM2 is expected to boost capacities and cut costs, but it’s not realistic to expect the technology to roll out at every price point.
Looking back across the years, there’s always been a transition phase between old and new memory standards. Back in the HD 3000 and 4000 period, AMD regularly rolled out its newest memory standards on upper-range cards, while midrange and budget models relied on older technology. GDDR5 debuted on the HD 4870 in 2008, but wasn’t deployed on anything but the flagship until 2009. As late as 2012, AMD was still using GDDR3 for the lowest-end GCN 1.0 GPUs.
We can assume, therefore, that AMD and Nvidia will continue to use GDDR5 in modern GPUs, even as those chips roll out on 14/16nm. This is partly because the benefits of HBM, like increased RAM densities, smaller PCBs, and decreased power consumption grow as the amount of RAM on the GPU grows. If you only need a 2GB frame buffer, HBM doesn’t offer much of an advantage compared to a conventional solution. Lower-end GPUs often don’t pack enough firepower to saturate a huge amount of RAM bandwidth — a two-wide HBM configuration would still offer 256GB/s of RAM bandwidth. If your $150 GPU doesn’t need that much capability, paying a premium for an HBM-based design makes no sense.
Thus far, we’ve only discussed GDDR5. What about GDDR5X? There are two problems with this technology. First, it’s proprietary to Micron and not expected before the end of 2016, when both AMD and Nvidia will hopefully have shipped next-generation chips from top to bottom. Second, however, it only addresses the bandwidth part of GDDR5’s problem. There’s always a power cost associated with increasing memory bandwidth, and simply doubling prefetch isn’t going to magically reduce GDDR5’s power consumption. One of the advantages of HBM that we covered long before AMD began talking it up this year is that HBM significantly improves GPU power efficiency. Micron will be pushing its technology in the opposite direction.
If rumors are true, AMD and Nvidia have both already taped out their next-generation designs — and that means they won’t be retrofitting them for GDDR5X at the last second. The change from GDDR5 to HBM requires a completely different memory controller, and it’s a non-trivial switch to move from one to the other.
I’m not going to say that GDDR5X won’t come to market, but it would surprise me to see AMD or NV use it. HBM and HBM2 have too many advantages in too many categories over GDDR5 as a whole, and bandwidth is only one part of that equation. HBM allows for products like theRadeon R9 Nano, which packs far more performance into a smaller space than was possible before. We expect to see that trend spread to lower-end cards as well, and when every millimeter counts, HBM/HBM2 is always going to have an advantage over GDDR5/5X.