Comments Locked

16 Comments

Back to Article

  • StevoLincolnite - Saturday, May 20, 2017 - link

    Wonder what will be more price competitive, HBM2 or GDDR6? Both seem to hit around the 400-600GB/s barrier at the same capacity for GPU's.
  • Yojimbo - Saturday, May 20, 2017 - link

    I think GDDR6 will be more price competitive. HBM2, and seemingly all TSV memories, don't seem to have lived up to all the promises made for them.

    HBM2, however, is not hitting a 400-600 GB/s barrier. It can achieve 900 GB/s while still maintaining reasonable power draw, as shown by NVIDIA's GV100, but it seems to be very expensive when doing so.
  • Tchamber - Saturday, May 20, 2017 - link

    I think HBM was a great idea, it just took too long to get to the market. Definitely too long to help AMD in any meaningful way. On the other hand, GDDR5 scaled very well, it more than doubled its speed, and if GDDR6 does as well, HBM will be sidelined.
  • alwayssts - Saturday, May 20, 2017 - link

    GDDR6.

    Not only do you have Micron added to the mix of Samsung and Hynix, but it's also lacking the stacking requirements and tighter voltage tolerances (HBM2 is 1.2v), more-or-less requiring the newer sub-20nm nodes to really make sense.

    As you can see with Hynix, the planned speeds of GDDR6 are going to be much more easily achievable. Odds are the first generation of GDDR6 from Hynix (and Micron) is ~21nm, just like (older) GDDR5. It's very possible the newer (9/10gbps) versions of GDDR5 are Hynix's 18nm, which is just starting to ramp. As we see a transition of GDDR6 to 18nm, it would only make sense a faster bin (16gbps) would become available. That said, it's also quite likely we'll finally meet the 2ghz mark with HBM2 as well.

    FWIW the conventional wisdom is that Samsung is roughly six months ahead (think how much earlier their 9gbps GDDR5 was available; currently in nvidia products) wrt DRAM tech and will have all production converted to ~18nm shortly. You can more-or-less map when these ram techs will reach the (required) speeds, and then commodity status based on that. IOW, Samsung will probably have 16gbps chips first, while 12/14 gbps will be cheap(er). This will likely also be true wrt to HBM2 (faster speeds earlier as Samsung starts at ~18 and then transititions to ~15nm), but to a lesser extent.
  • nunya112 - Sunday, May 21, 2017 - link

    gddr6 will b cheaper as it uses standard mounting and interconnects. HBM requires the GPU TSV's etc and they are stupid expensive and unreliable AMD found that out when it had to secure 2 or 3 companies just to barely keep up with a slow moving R9 Fury sales.
  • fanofanand - Wednesday, May 24, 2017 - link

    Didn't the article say it doesn't use the standard interconnects and that it requires a new interface?
  • edcoolio - Saturday, May 20, 2017 - link

    Thanks to SK Hynix and others there will be two paths towards higher performance. In regards to cost, I believe that both HBM2 and GDDR6 will naturally take their separate paths according to plan:

    HBM2 will likely remain a high end solution due to its finicky nature while GDDR6 will become the natural successor to GDDR5 in mainstream graphics cards.

    What could be exciting is a huge drop in price and mass production of their 8 Gb GDDR5 chips @9 and 10 GT/s while still drawing only around 1.5V. I think in the short to mid-term, that will have the greatest impact on mainstream gaming cards. Good stuff!
  • philehidiot - Saturday, May 20, 2017 - link

    This is my instinct. I think GDDR6 will be the mainstream path but HBM2 will provide competition and potentially higher bandwidth for those willing to put the R&D into the far trickier implementation.

    I think it'll be one of those situations where mainstream can provide 90-95% of the real world performance with a relatively simple "drop-in" solution whereas for those looking for an extra bump or headline figures HBM2 will enable another path. Hopefully this will push GDDR6 development along a little faster - as you say this will hopefully enable better memory for us real world consumers than we might otherwise have so it can only be a good thing.
  • haukionkannel - Sunday, May 21, 2017 - link

    Most propably Nvidia xx80 series and xx80ti will use gddr6 and Titan X versio could use HBM2. So cards up to 700-1200$ will use gddr memories and card from 1500-2500$ will use HBM memory types. Sounds very plausible.
    Hard to say what AMD will do, but They will lose battle in price if They don't sell Also gddr versions in the upper part of the gpu gategory. The Vega needs less memory band than Fury because of better memory compression, so maybe Navi will offer both options!
  • DanNeely - Saturday, May 20, 2017 - link

    "The insertion of the new chips into the roadmap indicates that GDDR5 will remain on the market for at least a couple of years down the road and will celebrate a decade on the market next year."

    Unless this is just talking about long tail products I'm a bit skeptical. While HBM has stalled out due to the high complexity of TSV manufacturing, previous generations of GDDR always did rapid change overs.

    eg AMD launched its first GGDR5 cards in June 2009, after that the only GDDR3 they did was a handful of sub $100 cards in Jan 2010 and a single card in 2011 that I suspect may have been OEM only (Because there is no retail price listed in Wikipedia). NVidia's switchover was a bit messier with low end GGDR5 parts showing up in June 09, but no high end ones until Sept 2010 (due to the problems they had with big Fermi delaying the high end 4xx cards); but their last new GDDR3 part was also released in Jan 2010.

    Is there some reason that GDDR6 is expected to be a slow switchover? If not, I'd expect to see GDDR6 displace GDDR5 from everything except the bottom end of the market sometime in the next 6-15 months as the next generation of cards launch and to see GDDR5 sales fall of the cliff as it happens with it vanishing from newly produced cards around the 1 decade anniversary of its initial consumer availability in about 2 years.
  • haukionkannel - Sunday, May 21, 2017 - link

    The cheap cards could use DDR memory instead of gddr memory. Like it has been decades in very low budget cards. Also older type of memory has been done in older factories because of production costs. So gddr5 may remain with gddr6 for some years.
  • DanNeely - Sunday, May 21, 2017 - link

    Yeah, the lowest end cards will probably linger with standard DDR4 (which to be fair will be a decent upgrade over the DDR3 currently being used); and legacy products are rarely moved to new fabs while their volume is declining to end of life. My point is that old versions of GDDR have always stopped being used very shortly after the new one came out; and to ask if there's any reason why we shouldn't expect it to happen this time around.
  • Alexvrb - Wednesday, May 24, 2017 - link

    I don't think so. DDR-equipped cards will struggle to compete with APUs at this point. Even a lowly RX 550 or GT 1030 uses GDDR5. My suspicion is that entry-level GDDR5 will fill the low-end until GDDR6 becomes as cheap or cheaper. Higher-speed GDDR5/6 will fill in the rest of the lineup until you get to top cards, which will be HBM of some variety.

    The high density and bandwidth per pin may also allow HBM to see some on-package use as VRAM or VRAM cache on high-end APUs... eventually. I don't think we'll see that this generation, though. DDR4 plus more efficient architectures will give Raven Ridge enough breathing room that it won't be necessary yet.
  • Stochastic - Sunday, May 21, 2017 - link

    So when can be expect HBM3?
  • Stochastic - Sunday, May 21, 2017 - link

    Hmmm, according to Wikipedia "No release date has been announced, though Samsung expects volume production by 2020."
  • fanofanand - Wednesday, May 24, 2017 - link

    GDDR5X is curiously missing from the table, should we assume that GDDR5X was merely a stopgap until GDDR6 arrived?

Log in

Don't have an account? Sign up now