SK Hynix has announced the introduction of the world’s first GDDR6 modules, with a data rate of up to 16Gbps and a theoretical bandwidth of 768GB/s of bandwidth if paired with a 384-bit I/O bus. SK Hynix’ own press release suggests such a product is coming within the next 12 months, in fact, when it states “SK Hynix has been planning to mass produce the product for a client to release high-end graphics card by early 2018 equipped with high performance GDDR6 DRAMs.”
This is an interesting statement, given that HBM2, not GDDR5X/6 has been broadly positioned as the future of next-generation memory. But having watched the HBM situation evolve — or not-evolve, depending on your viewpoint — I think we might have a plausible explanation in play. The simplest and most straightforward explanation is that Nvidia is tapping GDDR6 for Volta, which it intends to debut next year, and I think there could be plenty of truth to that. Whether it’s all of the truth, however, is open to interpretation.
Flash back to 2015, and AMD’s stated public plan was to introduce HBM with its Fury X family. Polaris, which it first discussed in December of that year, would be the midrange refresh based on GDDR5. An up-sized Polaris was also expected back then, with a new high-end GPU architecture, Vega, following at a later date. AMD killed fat Polaris and Vega has been delayed — and HBM2, it must be noted, has scarcely taken the world by storm. GDDR6 is expected to offer up to 16Gbps of bandwidth per pin, doubling the bandwidth available via GDDR5.
In the past, we’ve talked about how AMD has often led Nvidia when it comes to adopting new memory technology (AMD was the only company to use GDDR4 and the first to adopt GDDR5). But it’s worth noting that HBM2 has, so far, followed a very different path. One might have reasonably expected an initial HBM introduction in 2015 from AMD, followed by widespread adoption in 2016. That didn’t happen. Instead, we’ve had two extremely high-end products from Nvidia that both use HBM2, and nothing from AMD to date. Is this simply a function of cost, since new memory technologies are expensive, and Nvidia may have chosen to stick with GDDR5X before it knew what kind of shape HBM2 was going to be in? Definitely. But it’s also possible that AMD’s long absence from the high-end GPU market — and I’d like to point out that AMD has literally never gone this long without a high-end refresh in the past 15 years — is partly related to the difficulty of rolling HBM2 out.
Prior to launching HBM, AMD had a regular habit of refreshing its high-end architectures every 9-12 months, even if that refresh cycle was a re-badge effort with tweaked clock speeds. The company has kept to this cycle timing for its midrange cards; it just launched its RX 5xx family some 10 months after the RX 4xx family debuted. The Radeon R9 290X launched on October 24 2013, but the revamped 300 series and the Fury X didn’t debut until June 2015, 19 months later.
We’re now closing in on two years’ since Fury X launched and affordable HBM2 is nowhere to be found. The lack of buy-in from Nvidia is significant if only because Nvidia, as the dominant player in the GPU business, has a critical role to play in pushing the adoption of any new memory standard. Memory, after all, needs to ship before it can begin to take advantages of economies of scale and move towards affordability.
If Nvidia can roll a 384-bit 768GB/s memory interface with Volta starting next year, it would dwarf the 512GB/s bus that Vega is expected to debut. Of course, memory bandwidth efficiency matters just as much as raw figures, but HBM/HBM2 were once considered the obvious long-term replacements for GDDR5. If Nvidia sticks with GDDR6, it’ll throw a bone in that entire plan.