GDDR5X vs. HBM Gen2: It’s No Contest
NVIDIA’S NEW Titan X-beating Pascal graphics cards aren’t rocking the advanced stacked memory shenanigans of AMD’s top cards, or even its own pro-level Tesla GPUs. So, with the second generation of HBM hitting soon, that’s got to be a disadvantage for the
At the start of April, Nvidia unveiled its latest Tesla graphics processing behemoth, based on the brand new 16nm Pascal architecture. And there was much rejoicing. We were shown a 15.3-billion-transistor beast, with a huge number of stream processors, and vast swathes of second-generation highbandwidth memory (HBM), with unprecedented levels of memory bandwidth.
Fast-forward to May, and Jen-Hsun was back on stage showing off the very first consumer versions of Nvidia’s much-vaunted Pascal GPU architecture. And, despite some stunning performance claims, they’re very much second-tier GPUs compared with the GP100 chip at the heart of the top-tier Teslas.
That’s not a huge surprise, given that Nvidia has always liked to keep something back to fill the ultraenthusiast segment. But what is maybe a little more surprising is that the GTX 1080 has eschewed the HBM we know Pascal is capable of supporting—less we forget, AMD’s last generation GPUs supported it.
While you might be sad you’re not getting a 4,096bit memory interface or 720GB/s bandwidth in your GTX 1080, the use of GDDR5X memory in conjunction with the GP104 silicon makes sense, and is one of the reasons we’ll see a card that’s 25 percent faster than a Titan X, but for 40 percent less cash. GDDR5X is a new standard in graphics memory, but is still built on the same foundations as GDDR5. That’s important, because it makes it the perfect match for a secondtier GPU set to power multiple cards throughout Nvidia’s 1000-series stack. It means Nvidia’s chip designers didn’t need to find some convoluted way to ensure the GP104 GPU was cross-compatible with the more complex HBM memory as well as the older GDDR5 standard.
GDDR5X introduces the holy trinity of higher performance, greater capacity, and improved efficiency, but as well as that, because it follows the traditional GDDR5 design, one GPU is capable of supporting both the new and older standards. Hence the GTX 1080 comes with GDDR5X, and the slightly cut-down GP104 of the GTX 1070 rolls with old-school GDDR5.
It’s also far cheaper for Nvidia to implement a 256-bit memory interface with GDDR5X, knowing it can rock higher density 10Gb/s memory chips, rather than the more complex HBM standard. And it still gets to post impressive memory bandwidth scores without the cost of HBM or a 384 or 512-bit bus.
But none of this means GDDR5X is going toe-to-toe with HBM; it genuinely is no contest. The secondgen HBM standard will offer bandwidth of 1TB/s, while even the top GDDR5X spec’s 14Gb/s modules can only offer 448GB/s. But HBM is expensive and only going to be used on the top cards for the foreseeable future. Lower down the stack, Nvidia gets to offer significantly improved capacities and bandwidth without a huge bump in pricing.
HBM and GDDR5X, then, are likely to remain in two distinct camps of graphics card, with neither liable to impinge on the other’s turf. Well, until Polaris tips up, maybe....
The new GTX 1080 has eschewed the HBM we know Pascal is capable of.