SK Hynix’s first GDDR6 RAM will initially top out at 14 Gbps

A few weeks ago, SK Hynix announced that it would start mass-producing 16 Gbps GDDR6 memory, and teased its plans to have the high-speed graphics memory show up in graphics cards by early 2018. Imagine our surprise, then, when GDDR6 made it first entry into SK Hynix's databook—but not quite at the blistering speed the company initially claimed.

The two GDDR6 products that SK Hynix lists in its databook operate at 12 Gbps and 14 Gbps, lower than the 16 Gbps that the company talked about in its initial news. Those are still respectable speeds, especially given the chips' 1.35V power requirements. It's worth noting that Micron has touted GDDR5X running at 13 Gbps, but the fastest shipping version of that RAM (in Nvidia's GeForce GTX 1080 Ti) only runs at 11 Gbps.

Although the seemingly-low speed of the new GDDR6 chips might be a little concerning, it's worth pointing out that neither GDDR5 or GDDR5X started their life at the blistering speeds we're seeing now. We expect that Hynix will tune its processes and ramp up clock speeds on GDDR6 as time moves on. After all, the GDDR5X spec has a theoretical ceiling of 14 Gbps, but we've yet to see any chips at that speed.

SK Hynix's updated catalog also lists upcoming GDDR5 variations with speeds ranging from 6 Gbps all the way up to 10 Gbps. The fastest memory chips in this lineup require a little more juice, drawing 1.55V. Those speeds are fast enough to blur the line between GDDR5 and GDDR5X, and they might bring GDDR5X levels of performance to more mainstream products in the future.

The databook from SK Hynix declares that all of these products will make their debut in the fourth quarter of 2017. Presumably, the company will release more information about its announced 16 Gbps GDDR6 with time. In any case, consumers will have to continue to pay close attention to the fine print when looking at graphics memory, because the distinctions between these various types of RAM may not end up as clear as one might hope.

 

Comments closed
    • Shobai
    • 3 years ago

    [quote<]when GDDR6 made it first entry [/quote<] This should probably be 'its'

      • quaz0r
      • 3 years ago

      s/probably//

    • techguy
    • 3 years ago

    The fastest shipping variant of GDDR5x runs at 11.4GT/s on the Titan Xp. Not a huge difference from the 11GT/s stated in the article, certainly, but also not technically correct.

      • ImSpartacus
      • 3 years ago

      That’s underclocked 12 Gbps GDDR5X.

      Micron changed their GDDR5X product catalog to include 11 &12 Gbps GDDR5X in like January and sure enough, we see both get used in Nvidia’s lineup a couple months after that.

      [url<]https://www.micron.com/products/dram/gddr?306=GDDR5X&show=true[/url<]

    • ImSpartacus
    • 3 years ago

    Is anyone surprised about the GDDR6 news?

    16 Gbps was the long term target.

    Micron did the same thing with GDDR5X and its “theoretical” 16 Gbps max and 14 Gbps long term target. We got 10 Gbps, then 11-12 Gbps before attention shifted to more long term tech.

    I will say that 10 Gbps GDDR5 surprises me. I saw it mentioned on somebody’s slides earlier this week, but I wasn’t sure if it was a typo or a long term goal.

    Hopefully next year’s 2018 600 series refresh brings 9-10 Gbps GDDR5 to the Polaris rebrands as they need the bandwidth (No need to worry about Nvidia as they will surely be utilizing it asap).

      • techguy
      • 3 years ago

      I think it’s significant because they announced a certain graphics IHV with intentions to ship high-end consumer graphics product in 2018 which will feature 16GT/s GDDR6 connected via a 384-bit bus yielding 768GB/s bandwidth. 14GT/s GDDR6 would bring that total down to 672GB/s.

        • ImSpartacus
        • 3 years ago

        Remember Micron did not say their client was getting 16 Gbps GDDR6. The tech press reported it that way, but they didn’t say those words.

        The original press release:

        [url<]https://www.skhynix.com/eng/pr/pressReleaseView.do?seq=2086&offset=1[/url<] "The product operates with an I/O data rate of 16Gbps(Gigabits per second) per pin, which is the industry’s fastest. With a high-end graphics card, this DRAM processes up to 768GB(Gigabytes) of graphics data per second. 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." They very clearly say "up to", which is 100% true. It's the fastest that would be possible. They are probably talking about GV102. If we assume GV102 has 5376 cores like GV100 (just like GP102 and GP100), then that's 40% more cores than GP102's 3840. 768 GB/s (i.e. 16 Gbps on 384-bit) is 60% more than the 480 GB/s on the 2016 Titan X (GP102's debut product). Way too much, a waste of power that could be spent elsewhere. But 14 Gbps on that same bus yields a more modest 672 GB/s, as you pointed out. That's conveniently 40% more than the Titan X's 480 GB/s. That's perfectly in line with the increase in compute resources. Maybe Volta is capable of slightly higher clocks (though I doubt it, Pascal was built to clock high), but new architectures generally need slightly less bandwidth all else being equal, so it would probably even out.

          • Beahmont
          • 3 years ago

          Can you please stop with the assumptions of linear scaling across µArches as a given?

          You have no way of knowing if 60% more bandwidth is too much for Volta. No one has even seen Volta in action yet and here you are claiming as fact that it has the same bandwidth use as Pascal. For all you or anyone other than Nvidia knows, Volta could make do with Pascal bandwidth, be choked to death on it, run even better than Pascal because it doesn’t need as much bandwidth, or make do with Pascal bandwidth but run even better with more bandwidth.

            • ImSpartacus
            • 3 years ago

            We can only use history and history has pretty much been nearly linear in that relationship (with a slight adjustment as architectures get more and more efficient with their bandwidth, as I mentioned).

            Overall, firms like AMD and Nvidia aren’t producing these architectures in a vacuum. They have expectations for memory availability (and tons of other stuff from all of their vendors).

            And when you’re setting expectations, you don’t assume that your memory vendors are going to hit their long term performance targets in the first year. It doesn’t happen.

            We’re seeing that with SK Hynix and we’ll see it with Micron and Samsung. There’s just a reality that these people are very used to working in and we can’t ignore it just because we get excited over some clickbait headlines.

        • the
        • 3 years ago

        Still 672 GB is nothing to sneeze at. That’s more than the HBM1 could product and in the neighborhood of HBM2.

Pin It on Pinterest

Share This