Feeding Pascal: GDDR5X

An ongoing problem for every generation of GPUs is the matter of memory bandwidth. As graphics is an embarrassingly parallel problem, it scales out with additional ALUs – and consequently Moore’s Law – relatively well. Each successive generation of GPUs are wider and higher clocked, consuming more data than ever before.

The problem for GPUs is that while their performance tracks Moore’s Law well, the same cannot be said for DRAM. To be sure, DRAM has gotten faster over the years as well, but it hasn’t improved at nearly the same pace as GPUs, and physical limitations ensure that this will continue to be the case. So with every generation, GPU vendors need to be craftier and craftier about how they get more memory bandwidth, and in turn how they use that memory bandwidth.

To help address this problem, Pascal brings to the table two new memory-centric features. The first of which is support for the newer GDDR5X memory standard, which looks to address the memory bandwidth problem from the supply side.

By this point GDDR5 has been with us for a surprisingly long period of time – AMD first implemented it on the Radeon HD 4870 in 2008 – and it has been taken to higher clockspeeds than originally intended. Today’s GeForce GTX 1070 and Radeon RX 480 cards ship with 8Gbps GDDR5, a faster transfer rate than the originally envisioned limit of 7Gbps. That GPU manufacturers and DRAM makers have been able to push GDDR5 so high is a testament to their abilities, but at the same time the technology is clearly reaching its apex (at least for reasonable levels of power consumption)

As a result there has been a great deal of interest in the memory technologies that would succeed GDDR5. At the high end, last year AMD became the first vendor to implement version 1 of High Bandwidth Memory, a technology that is a significant departure from traditional DRAM and uses an ultra-wide 4096-bit memory bus to provide enormous amounts of bandwidth. Not to be outdone, NVIDIA has adopted HBM2 for their HPC-centric GP100 GPU, using it to deliver 720GB/sec of bandwidth for Pascal P100.

While from a technical level HBM is truly fantastic next-generation technology – it uses cutting edge technology throughput, from TSV die-stacking to silicon interposers that connect the DRAM stacks to the processor – its downside is that all of this next-generation technology is still expensive to implement. Precise figures aren’t publicly available, but the silicon interposer is more expensive than a relatively simple PCB, and connecting DRAM dies through TSVs and stacking them is more complex than laying down BGA DRAM packages on a PCB. For NVIDIA, a more cost-effective solution was desired for GP104.

That solution came from Micron and the JEDEC in the form of GDDR5X. A sort of half-generation extension of traditional GDDR5, GDDR5X further increases the amount of memory bandwidth available from GDDR5 through a combination of a faster memory bus coupled with wider memory operations to read and write more data from DRAM per clock. And though it’s not without its own costs such as designing new memory controllers and boards that can accommodate the tighter requirements of the GDDR5X memory bus, GDDR5X offers a step in performance between the relatively cheap and slow GDDR5, and relatively fast and expensive HBM2.

Relative to GDDR5, the significant breakthrough on GDDR5X is the implementation of Quad Data Rate (QDR) signaling on the memory bus. Whereas GDDR5’s memory bus would transfer data twice per write clock (WCK) via DDR, GDDR5X extends this to four transfers per clock. All other things held equal, this allows GDDR5X to transfer twice as much data per clock as GDDR5.

QDR itself is not a new innovation – Intel implemented a quad pumped bus 15 years ago for the Pentium 4 with AGTL+ – but this is the first time it has been implemented in a common JEDEC memory standard. The history of PC memory standards is itself quite a tale, and I suspect that the fact we’re only seeing a form of QDR now is related to patents. But regardless, here we are.

Going hand-in-hand with the improved transfer rate of the GDDR5X memory bus, GDDR5X also once again increases the size of read/write operations, as the core clockspeed of GDDR5X chips is only a fraction of the bus speed. GDDR5X uses a 16n prefetch, which is twice the size of GDDR5’s 8n prefetch. This translates to 64B reads/writes, meaning that GDDR5X memory chips are actually fetching (or writing) data in blocks of 64 bytes, and then transmitting it over multiple cycles of the memory bus. As discussed earlier, this change in the prefetch size is why the memory controller organization of GP104 is 8x32b instead of 4x64b like GM204, as each memory controller can now read and write 64B segments of data via a single memory channel.

Overall GDDR5X is planned to offer enough bandwidth for at least the next couple of years. The current sole supplier of GDDR5X, Micron, is initially developing GDDR5X from 10 to 12Gbps, and the JEDEC has been talking about taking that to 14Gbps. Longer term, Micron thinks the technology can hit 16Gbps, which would be a true doubling of GDDR5’s current top speed of 8Gbps. With that said, even with a larger 384-bit memory bus (ala GM200) this would only slightly surpass what kind of bandwidth HBM2 offers today, reinforcing the fact that GDDR5X will fill the gap between traditional GDDR5 and HBM2.

Meanwhile when it comes to power consumption and power efficiency, GDDR5X will turn back the clock, at least a bit. Thanks in large part to a lower operating voltage of 1.35v, circuit design changes, and a smaller manufacturing node for the DRAM itself, 10Gbps GDDR5X only requires as much power as 7Gbps GDDR5. This means that relative to GTX 980, GTX 1080’s faster GDDR5X is essentially “free” from a power perspective, not consuming any more power than before, according to NVIDIA.

That said, while this gets NVIDIA more bandwidth for the same power – 43% more, in fact – NVIDIA has now put themselves back to where they were with GTX 980. GDDR5X can scale higher in frequency, but doing so will almost certainly further increase power consumption. As a result they are still going to have to carefully work around growing memory power consumption if they continue down the GDDR5X path for future, faster cards.

On a final specification note, GDDR5X also introduces non-power-of-two memory chip capacities such as 12Gb. These aren’t being used for GTX 1080 – which uses 8Gb chips – but I wouldn’t be surprised if we see these used down the line. The atypical sizing would allow NVIDIA to offer additional memory capacities without resorting to asymmetrical memory configurations as is currently the case, all the while avoiding the bandwidth limitations that can result from that.

Moving on to implementation details, GP104 brings with it a new memory controller design to support GDDR5X. As intended with the specification, this controller design is backwards compatible with traditional GDDR5, and will allow NVIDIA to support both memory standards. At this point NVIDIA hasn’t talked about what kinds of memory speeds their new controller can ultimately hit, but the cropped signal analysis diagram published in their slide deck shows a very tight eye. Given the fact that NVIDIA’s new memory controller can operate at 8Gbps in GDDR5 mode, I would be surprised if we don’t see at least 12Gbps GDDR5X by the tail-end of Pascal’s lifecycle.

But perhaps the bigger challenge is on the board side of matters, where NVIDIA and their partners needed to develop PCBs capable of handling the tighter signaling requirements of the GDDR5X memory bus. At this point video cards are moving 10Gbps/pin over a non-differential bus, which is itself a significant accomplishment.  And keep in mind that in the long run, the JEDEC and Micron want to push this higher still.

To that end it somewhat undersells the whole process to just say that GDDR5X required “tighter signaling requirements”, but it’s an apt description. There is no single technology in place on the physical trace side to make this happen; it’s just a lot of precision, intensive work into ensuring that the traces and the junctions between the PCB, the chip, and the die all retain the required signal integrity. With a 256-bit wide bus we’re not looking at something too wide compared to the 384 and 512-bit buses uses on larger GPUs, so the task is somewhat simpler in that respect, but it’s still quite a bit of effort to minimize the crosstalk and other phenomena that degrade the signal, and which GDDR5X has little tolerance for.

As it stands I suspect we have not yet seen the full ramifications of the tighter bus requirements, and we probably won’t for cards that use the reference board or the memory design lifted from the reference board. For stability reasons, data buses are usually overengineered, and it’s likely the GDDR5X memory itself that’s holding back overclocking. Things will likely get more interesting if and when GDDR5X filters its way down to cheaper cards, where keeping costs in check and eking out higher margins becomes more important. Alternatively, as NVIDIA’s partners get more comfortable with the tech and its requirements, it’ll be interesting to see where we end up with the ultra-high-end overclocking cards –the Kingpins, Lightnings, Matrices, etc – and whether all of the major partners can keep up in that race.

Designing GP104: Running Up the Clocks Feeding Pascal, Cont: 4th Gen Delta Color Compression


View All Comments

  • TestKing123 - Wednesday, July 20, 2016 - link

    Sorry, too little too late. Waited this long, and the first review was Tomb Raider DX11?! Not 12?

    This review is both late AND rushed at the same time.
  • Mat3 - Wednesday, July 20, 2016 - link

    Testing Tomb Raider in DX11 is inexcusable.

  • TheJian - Friday, July 22, 2016 - link

    Furyx still loses to 980ti until 4K at which point the avg for both cards is under 30fps, and the mins are both below 20fps. IE, neither is playable. Even in AMD's case here we're looking at 7% gain (75.3 to 80.9). Looking at NV's new cards shows dx12 netting NV cards ~6% while AMD gets ~12% (time spy). This is pretty much a sneeze and will as noted here and elsewhere, it will depend on the game and how the gpu works. It won't be a blanket win for either side. Async won't be saving AMD, they'll have to actually make faster stuff. There is no point in even reporting victory at under 30fps...LOL.

    Also note in that link, while they are saying maxwell gained nothing, it's not exactly true. Only avg gained nothing (suggesting maybe limited by something else?), while min fps jumped pretty much exactly what AMD did. IE Nv 980ti min went from 56fps to 65fps. So while avg didn't jump, the min went way up giving a much smoother experience (amd gained 11fps on mins from 51 to 62). I'm more worried about mins than avgs. Tomb on AMD still loses by more than 10% so who cares? Sort of blows a hole in the theory that AMD will be faster in all dx12 stuff...LOL. Well maybe when you force the cards into territory nobody can play at (4k in Tomb Raiders case).

    It would appear NV isn't spending much time yet on dx12, and they shouldn't. Even with 10-20% on windows 10 (I don't believe netmarketshare's numbers as they are a msft partner), most of those are NOT gamers. You can count dx12 games on ONE hand. Most of those OS's are either forced upgrades due to incorrect update settings (waking up to win10...LOL), or FREE on machine's under $200 etc. Even if 1/4 of them are dx12 capable gpus, that would be NV programming for 2.5%-5% of the PC market. Unlike AMD they were not forced to move on to dx12 due to lack of funding. AMD placed a bet that we'd move on, be forced by MSFT or get console help from xbox1 (didn't work, ps4 winning 2-1) so they could ignore dx11. Nvidia will move when needed, until then they're dominating where most of us are, which is 1080p or less, and DX11. It's comic when people point to AMD winning at 4k when it is usually a case where both sides can't hit 30fps even before maxing details. AMD management keeps aiming at stuff we are either not doing at all (4k less than 2%), or won't be doing for ages such as dx12 games being more than dx11 in your OS+your GPU being dx12 capable.

    What is more important? Testing the use case that describes 99.9% of the current games (dx11 or below, win7/8/vista/xp/etc), or games that can be counted on ONE hand and run in an OS most of us hate. No hate isn't a strong word here when the OS has been FREE for a freaking year and still can't hit 20% even by a microsoft partner's likely BS numbers...LOL. Testing dx12 is a waste of time. I'd rather see 3-4 more dx11 games tested for a wider variety although I just read a dozen reviews to see 30+ games tested anyway.
  • ajlueke - Friday, July 22, 2016 - link

    That would be fine if it was only dx12. Doesn't look like Nvidia is investing much time in Vulkan either, especially not on older hardware.

  • Cygni - Wednesday, July 20, 2016 - link

    Cool attention troll. Nobody cares what free reviews you choose to read or why. Reply
  • AndrewJacksonZA - Wednesday, July 20, 2016 - link

    Typo on page 18: "The Test"
    "Core i7-4960X hosed in an NZXT Phantom 630 Windowed Edition" Hosed -> Housed
  • Michael Bay - Thursday, July 21, 2016 - link

    I`d sure hose me a Core i7-4960X. Reply
  • AndrewJacksonZA - Wednesday, July 20, 2016 - link

    @Ryan & team: What was your reasoning for not including the new Doom in your 2016 GPU Bench game list? AFAIK it's the first indication of Vulkan performance for graphics cards.

    Thank you! :-)
  • Ryan Smith - Wednesday, July 20, 2016 - link

    We cooked up the list and locked in the games before Doom came out. It wasn't out until May 13th. GTX 1080 came out May 14th, by which point we had already started this article (and had published the preview). Reply
  • AndrewJacksonZA - Wednesday, July 20, 2016 - link

    OK, thank you. Any chance of adding it to the list please?

    I'm a Windows gamer, so my personal interest in the cross-platform Vulkan is pretty meh right now (only one title right now, hooray! /s) but there are probably going to be some devs are going to choose it over DX12 for that very reason, plus I'm sure that you have readers who are quite interested in it.

Log in

Don't have an account? Sign up now