It’s no secret that AMD and Nvidia are planning to launch new GPUs in the next few months, but there are still questions about configurations and technology. Initially, we expected these new cards to use the HBM2 memory standard, but the persistent rumor has been that AMD’s upcoming Polaris would be GDDR5 based. That now appears to be true for Nvidia as well.
Leaked photos of the GP104 die show the chip with a standard memory interface — either GDDR5 or GDDR5X (it’s difficult to tell). PC Perspective is reporting that there’s an ongoing rumor the GTX 1080 will use GDDR5X, which could be possible, but would honestly seem to be a bit early — when we spoke to Micron about the memory, the company implied that the production ramp would be later in the year. With JEDEC qualification and good initial production yields, however, it’s possible that Nvidia could deploy GDDR5X on high-end GPUs, and it wouldn’t be the first time that a GPU manufacturer used a limited memory run. AMD was the only company to make significant use of GDDR4 or HBM, after all.
As for the lack of HBM2, it’s not surprising at this juncture. All current data suggests that HBM2 will arrive with the launch of the high-end “big” GPUs — Nvidia’s GP100 and AMD’s new Vega architecture, currently expected in the Q4 2016 / Q1 2017 timeframe. We don’t expect AMD to tap HBM for Polaris this time out, especially not if they’ve focused on scaling 1-2 GPU cores for desktop and mobile. Remember, HBM requires an entirely different memory controller architecture than GDDR5, and neither AMD nor Nvidia is going to want to build two different memory controllers for the same GPU core unless absolutely necessary. It’s much easier to believe Nvidia might do a GDDR5/GDDR5X mixture than an HBM2 / GDDR5 cross, and the same is true for AMD.
We don’t expect Polaris or Pascal to ship with HBM support in two months (if we’re wrong on this, we’ll have us a bite of crow), and we’re genuinely not concerned about what that means for the GPUs themselves. Over the last ten years, both AMD and Nvidia have adopted a wide variety of memory bus sizes and memory technologies to match the GPUs they shipped. It’s not uncommon to see both companies adopt multiple memory standards for multiple products — in past years, DDR2, GDDR3, and GDDR5 all shared space in Nvidia’s product stack.
HBM2 is the long-term future for GPU memory and that’s still going to be true if the next-generation cards from AMD and Nvidia use GDDR5(X). Both companies have decades of experience in matching bandwidth to graphics processor; the last time we remember a GPU being terminally memory bandwidth-bound was the GeForce 2 era.