Comments Locked

59 Comments

Back to Article

  • Eden-K121D - Thursday, May 12, 2016 - link

    SO i Guess Polaris 10/11 will have GDDR5 as GDDR5X is out of reach of AMD
  • casperes1996 - Thursday, May 12, 2016 - link

    I'm betting on one of them coming with HBM. The majority will probably be GDDR5, yeah. But I'm expecting one or two products to be announced with HBM. I have a feeling AMD wants to slowly increase the amount of products in their stack that uses the memory, and the more it is used (by AMD and others) the cheaper it'll become too.
  • StevoLincolnite - Thursday, May 12, 2016 - link

    Not going to happen. Polaris isn't high-end. HBM will not be coming until Vega late this year/early next year. Then Navi follows that up with a next generation memory. We know this because AMD has told us.
  • bug77 - Friday, May 13, 2016 - link

    Do you honestly believe that if we're getting GDDR5X and HBM2 this year, we'll be getting a next generation memory in 2017 or 2018? That slide from AMD was probably referring to some improved HBM2 memory. HBM3 at best.
  • vladx - Friday, May 13, 2016 - link

    Next generation is probably HMC, or just HBM3 we all know how much AMD likes to hype everything.
  • StevoLincolnite - Saturday, May 14, 2016 - link

    I'm just the messenger. :P
    Just remember we had HBM 1 last year. HBM 2 this year, why is it hard to believe that we will get a next gen memory/HBM 3 in 2 years time?
  • slickr - Saturday, May 14, 2016 - link

    They can couple the higher end of it, one card with HBM1, you do understand they still have HBM1, so if they have some card that goes for around $280-330 they could probably make it with HBM1.

    Personally I hate that they are not going to use G5x for any of their Polaris, it feels like we are getting just half of the innovations that we should by having old GDDR5 memory. I mean its not like 10 or 11 gbps are going to make much difference from 7gbps but lower end gpu with less shaders and processing units, its more often going to be bottlenecked by the core, but it would have been still nice to have g5x where in some cases the memory would be important, especially the more expensive $200+ parts.
  • BurntMyBacon - Friday, May 13, 2016 - link

    It'll probably be a Vega chip with HBM. Rumor has it that there are two version of (Vega 10/Vega 11). There seems to be enough space in the lineup for two chips, so the rumor at least makes sense. I've got no idea how accurate the naming might be, but given no solid information, I'll use these references for now. If the rumors are ture, it is possible that Vega 10 will be designed for HBM and Vega 11 will use GDDR5X (which would jive with casperes1996's assertion, though not with Polaris).

    The question on my mind is whether or not the Polaris memory controller is designed to use GDDR5X as an option. Contrary to Eden's assertion, GDDR5X is not out of reach for AMD. Micron said that they are the only supplier for the GTX1080. They never said that they weren't making any for other cards or other manufacturers. Furthermore, there are several other memory manufacturers on the market that are working on their own GDDR5X and we have no idea how close they are to mass production. It is interesting that AMD announced first, had a working silicon demonstration first, and yet they will reach market second. It's possible that something other than the silicon is holding them up. Perhaps they picked the wrong memory manufacturer to partner with.
  • Kalelovil - Thursday, May 12, 2016 - link

    Polaris 10 is only 2/3rds the size of GP104. The fastest Polaris 10 card will probably be slightly slower than the GTX 1070, which makes do with GDDR5.
  • dragonsqrrl - Thursday, May 12, 2016 - link

    Yes, but you have to remember that AMD has traditionally required far higher memory bandwidth to achieve that level of performance. I suspect this will change with Polaris, but I still wouldn't be surprised to see GDDR5X on the higher-end Polaris 10 based card.
  • FlorisR - Monday, May 16, 2016 - link

    Yes, that's was true in the past, but not with Polaris. They implemented memory compression, to save on required memory bandwidth.
  • BurntMyBacon - Friday, May 13, 2016 - link

    @Kalelovil: "Polaris 10 is only 2/3rds the size of GP104. The fastest Polaris 10 card will probably be slightly slower than the GTX 1070, which makes do with GDDR5."

    I agree that Polaris probably won't need the kind of bandwidth that a 256bit 10Gbps GDDR5X array can provide. Hoever, if GDDR5X ends up being as cost effective as Micron seems to suggest, then AMD could save significant die area and silicon cost by reducing the bus width even on mid-high end parts. Some of that savings would be eaten by the initial premium for GDDR5X, but it could make sense. Reducing the bus width generally means a reduction in the overall memory capacity of the cards which wouldn't bode well for high end cards like the GTX1080/GTX1070, but in the mainstream that Polaris is targeting, this wouldn't be much of a problem.
  • fanofanand - Friday, May 13, 2016 - link

    AMD is desperately working on reducing their TDP, shaving off 10+ watts simply by using a different memory type must be appealing to them.
  • SunLord - Thursday, May 12, 2016 - link

    We really don't know at what stage Samsung and Hynix are at when it comes to mass production for gddr5x. When GDDR5X was announced it was assumed all the manufactures were 6 months out though Samsung might of focus on HBM 2 first
  • extide - Friday, May 13, 2016 - link

    Lol, I love nvidiots. I mean wow, just the world some of you guys live in is absolutely ridiculous.
  • Michael Bay - Monday, May 16, 2016 - link

    Not worrying about what the next driver will break is certainly nice.
    Then again, you guys save on heating big time.
  • xthetenth - Tuesday, May 17, 2016 - link

    Agreed on not worrying about driver issues. I'm so glad I got rid of my 970. I'd already had two different driver issues, getting the joys of the 364 drivers would've driven me nuts. My 290's been vastly better in terms of drivers. It's also been nice not having to run NV inspector to keep my card from heating the room on the desktop because apparently clocking down from full 3D clocks when someone's running two screens is hard, you guys. The only reason NV has mindshare is because of a reputation that people like you don't care to check.
  • mfearer - Thursday, May 12, 2016 - link

    Is this happening at Micron's Manassas, Virginia, USA location?
  • casperes1996 - Thursday, May 12, 2016 - link

    Why isn't it called GQDR? If it runs at quad-data rate, why do we name it double-data rate?
  • LukaP - Thursday, May 12, 2016 - link

    Because it slips off the tongue nicer? plus DDR is already known, so GDDR makes the logical connection of it being memory. GQDR is just some weird acronym noone would connect to memory at first.
  • SleepyFE - Thursday, May 12, 2016 - link

    So we can go from SDR to DDR but QDR is too much? Anyone who knows any of those acronyms knows them all. Anyone who doesn't doesn't care. I care because the acronyms stand for something descriptive. An all terrain four wheeler is not called a bikex after all.
  • qap - Friday, May 13, 2016 - link

    Actually no one used "SDR" (SDRAM has different meaning if anyone wonders). It appeared only as afterthought when DDR arrived. So really - industry massively adopted only "DDR" so far.
  • extide - Friday, May 13, 2016 - link

    Yeah, technically it would be SDR SDRAM and DDR SDRAM, etc.
  • dusty5683 - Thursday, May 12, 2016 - link

    My guess is that QDR is just another mode, it can still operate in DDR mode. Also the addressing scheme is the same as DDR. Its probably also a little to do with not wanting to use a new acronym.
  • Despoiler - Thursday, May 12, 2016 - link

    Correct. It has multiple modes it can operate in.
  • CaedenV - Thursday, May 12, 2016 - link

    Let's Review!

    SDRRAM was old-school ram that could send one bit of information per clock
    DDR SDRAM is 'current' RAM that sends 2 bits per clock (one bit on the tick, and one of the tock)
    QDR SDRAM would mean that you can somehow send 4 bits per clock... which has no end of issues, and is not done.

    Dual DDR SDRAM is a configuration where you have DDR with 2 chanels of access to it (Think RAID0 with 2 drives)
    Quad DDR SDRAM (which is essentially what this is) is DDR with 4 channels of access to it (Think RAID0 with 4 drives)

    Anywho that is why it is not QDR memory; GQDR5 would not make any sense. GQDDR5 is a mouthful. GDDR5x... that is just marketing gold right there.
  • Cygni - Thursday, May 12, 2016 - link

    The G and 5 are not really meaningful, so you could easily just call it QDDR. It's really just dumb marketing jargon from an industry full of comically bad marketing jargon, so I guess that shouldn't be a surprise.
  • willis936 - Thursday, May 12, 2016 - link

    ... except it is 4 bits per clock and is QDR and it is done in the real world and this isn't the first technology to do it.
  • MrSpadge - Friday, May 13, 2016 - link

    Yep, Pentium 4 FSB and some version of AGP used QDR.
  • SleepyFE - Friday, May 13, 2016 - link

    I think it's 2 bits per clock with 2 clocks.
  • gurok - Thursday, May 12, 2016 - link

    What's SDRRAM? I think you mean just SDRAM -- synchronous dynamic RAM.
  • extide - Friday, May 13, 2016 - link

    It's not the tick and the tock, it's the rise and the fall of the signal. With Single Data Rate, you send on every rise or every fall, with DDR you send on the rise and the fall.
  • willis936 - Thursday, May 12, 2016 - link

    My guess would be because the implementer can choose to run GDDR5X memory in either DDR or QDR mode.
  • Yojimbo - Thursday, May 12, 2016 - link

    I wonder how the 1080 gets by with such relatively low RAM bandwidth (in terms of ratio to performance as compared to Maxwell GPUs)? Did they improve their compression algorithms significantly again or is it something else? NVIDIA also seems to be claiming a lower real-work performance gain compared to peak FLOPS gain with Pascal over Maxwell (25%(So I've seen claimed somewhere) for performance and >50% for FLOPS), which is odd (if true). Is that because of an aggressive boost clock? Otherwise it seems to imply an architecture that isn't fed as efficiently. What would be the beneficial trade-off that could lead to that? Overall power efficiency? Die size efficiency?
  • jasonelmore - Thursday, May 12, 2016 - link

    Double the Register size and much higher clocks
  • Yojimbo - Thursday, May 12, 2016 - link

    What question were you trying to answer there? I asked more than one. Were you referring to the RAM bandwidth? What does core clock have to do with that? Does register size help alleviate need for RAM bandwidth? If so, how?
  • dragonsqrrl - Thursday, May 12, 2016 - link

    Larger registers and caches can help reduce memory bottlenecks (the GPU doesn't have to access main memory as often).
  • Yojimbo - Sunday, May 15, 2016 - link

    I just found this on videocardz: http://cdn.videocardz.com/1/2016/05/NVIDIA-GeForce... NVIDIA is claiming a 1.7 times bandwidth improvement with Pascal versus Maxwell due to fourth generation delta color compression.
  • Yojimbo - Sunday, May 15, 2016 - link

    http://cdn.videocardz.com/1/2016/05/NVIDIA-GeForce...
  • Yojimbo - Sunday, May 15, 2016 - link

    Sorry, the color compression only accounts for 20% improvement in the bandwidth. The other ~40% is from faster DRAM.
  • Schecter1989 - Thursday, May 12, 2016 - link

    Small question, why is it that recently the ratings of memory speed have changed to Gb/s instead of Ghz. When I see my memory clock using MSI afterburner or Precision X they both show 3500Mhz. Which of course with DDR equals out to the 7Ghz we used to advertise memory as. Whats the change to the Gb/s signify? And if the GDDR5X chip runs at 10Gb/s wouldnt that mean that if you take bit to byte of that 10Gbps you would only get a bandwidth of 1.25GB/s.

    What exactly am I missing here? Lol
  • zepi - Thursday, May 12, 2016 - link

    10Gbps is unambiguous, whereas 1GHz can be either 1, 2, 4 or even 8 Gbps depending on the modulation used. Maybe that is the reason?
  • SunnyNW - Thursday, May 12, 2016 - link

    I'm not a memory expert but from what I unerstand... the 10 Gb/s is the per pin speed rating of the memory. So with a 256 bit bus (like the GTX 1080 has) thats 256 pins multiplied by 10 Gb/pin or 1.25 GB giving you the 320 GB/s of memory bandwith.
  • SunnyNW - Thursday, May 12, 2016 - link

    And to answer your question your other question I believe the change to Gb/s from of Ghz is to make it easier to compare between different memory technologies, ie GDDR, HBM, etc.
    Also these gpus use 32-bit memory controllers and as per the article each of these 8 Gb (1GB) GDDR5x chips has a 32-bit interface therefore giving you 320 Gb/s of bandwidth per IC or 40 GB/s. Hope that helps at least a little and if anyone would like to correct anything I said or add to it feel free.
  • BurntMyBacon - Friday, May 13, 2016 - link

    I'll just add that there are 8 ICs on the GTX1080. At 40 GB/s per IC you get a total of 320GB/s (as advertised).
  • Gigaplex - Thursday, May 12, 2016 - link

    Because they're trying to increase (or at least maintain) bandwidth while lowering clock speeds. Lower clock speeds means less power usage. It also means it looks slower than older generation hardware when you simply quote the GHz rating.
  • BurntMyBacon - Friday, May 13, 2016 - link

    @Schecter1989: "why is it that recently the ratings of memory speed have changed to Gb/s instead of Ghz."

    You need more information for the GHz rating to mean anything. For instance. I can have a 100GHz clock rate, but if I can only transmit 1 bit every 100 cycles, then I only have a transfer rate of 1Gbps. Gbps is more meaningful for transfers. 1Gbps is will give me the same amount of data over the same amount of time whether I use a 100GHz clock or a 500MHz clock. Given that GDDR5X can transmit 2 bits per cycle (DDR mode) or 4 bits per cycle (QDR mode) but doesn't change its frequency, it makes a lot of sense to start listing transfer speed in terms of Gbps per pin.

    @Schecter1989: "When I see my memory clock using MSI afterburner or Precision X they both show 3500Mhz."

    This is the actual clock frequency.

    @Schecter1989: "Which of course with DDR equals out to the 7Ghz we used to advertise memory as."

    This is a misnomer. There is no clock frequency of 7GHz with DDR. DDR simply transfers 2 bits per cycle. One on each edge (rising edge / falling edge) of the clock. SDR is an acronym invented after the introduction of DDR for the purposes of distinguishing it from the older memory that can only transfer on one of the edges of the clock. Marketing likes to throw the 7GHz number around because it is far more impressive to show the larger "effective frequency" to explain the doubling of bandwidth given the same bus width than it is to explain how the same frequency and bus width can given you double the bandwidth through an more clever clock scheme. BTW, QDR tranfers on both edges and both levels, but the technical details would be better left to a reputable tech site do a full writeup on than a random post in the comments section.

    @Schecter1989: "if the GDDR5X chip runs at 10Gb/s wouldnt that mean that if you take bit to byte of that 10Gbps you would only get a bandwidth of 1.25GB/s."

    Correct. You are almost done. This is the pin speed. The bus width on the GTX1080 is 256 bits wide. Multiply that by the bandwidth you show (1.25GB/s) and you get the advertised 320GB/s.
  • Schecter1989 - Saturday, May 14, 2016 - link

    Thanks @BurntMyBacon Your response covered my issues. I did get the 320GB/s when using the pin speed multipled by each actual pin count. I did get the same result I just was unsure if that was the correct method to find your actual full bandwidth potential.

    Ok so if the old style of cards being advertised in the GHZ range, then what was the faster MHz rating really giving us? Did pin speeds ever change throughout the past 10 years?

    Basically like with the change from the 6xx series to the 7xx. I purchased a GTX 770 back when I first built my machine, now this card advertised a 7Ghz speed on the memory, and when the GTX 680 arrived it had 6Ghz. So if how youre saying the pin speed is the actual measure of performance, does that mean that the so called +1Ghz in speed really did not do anything for me? If of course their pin speeds were the same?
  • jasonelmore - Thursday, May 12, 2016 - link

    It is starting to look like Pascal is Memory Bandwidth Starved and that any increase in bandwidth over the 10Gbps rate, will increase performance quite a bit. Nvidia is purposely holding back pascal because they know they are going to be on 16nm for 2-3 years.

    We shall see when HBM finds it's way onto pascal. the P100 is already doing 21 TFLOPS half precision compared to 9 TFLOPS on the 1080
  • Yojimbo - Thursday, May 12, 2016 - link

    Purposely holding it back? I don't think so. Seems like a silly "conspiracy theory". They have to worry too much about competition for that. Their gross margins would plummet if they spend billions of dollars developing and manufacturing a highly capable chip and then don't take full advantage of it by pairing it with sub-optimal memory configurations, because their selling prices would be lower.

    I also don't think they'll be at 16nm for 2 to 3 years. The data center segment is becoming more important than the mobile segment. The 20nm process catered to mobile chips. I doubt the foundries will do the same with 10nm because mobile growth has stalled and they won't want to miss out on the opportunity to make data center chips.

    Finally, HBM already has found its way to Pascal in the Tesla P100. I assume you meant consumer Pascal. I doubt HBM will find its way onto a consumer graphics-oriented GPU that has a die size close to that of GP104. GDDR5X still has headroom above its implementation in the 1080. A consumer variant of the GP100 would use HBM2, I assume, but it has a die size of ~600 mm^2 compared to GP104's ~330 mm^2.

    As far as half-precision, does it not also use half bandwidth (per operation)? If they allow the 1080 to do half-precision then the bandwidth shouldn't be any more of an issue than it is for single precision usage. The P100 has the bandwidth it has because its beneficial for the compute tasks it's designed to be used for. The 1080 has the bandwidth it does presumably because that bandwidth is well-balanced for graphics workloads, and that shouldn't change if its a half-precision graphics workload (but where is that going to be used?). Some people may want to use the 1080 for half-precision compute and in that scenario the card may be bandwidth starved in some instances, but if it is it should also be bandwidth starved in similar single precision compute workloads.
  • BurntMyBacon - Friday, May 13, 2016 - link

    @Yojimbo: "Purposely holding it back? ... They have to worry too much about competition for that."

    Current evidence to the contrary. AMD and nVidia seem to be targeting two completely different market spaces. AMD states that Polaris 10 will be R9-390(X) level performance where nVidia is targeting the space above the GTX980Ti.

    @Yojimbo: "Their gross margins would plummet if they spend billions of dollars developing and manufacturing a highly capable chip and then don't take full advantage of it by pairing it with sub-optimal memory configurations, because their selling prices would be lower."

    Why? They are selling these things for $600/$700. Using an inferior memory pairing doesn't magically lower that price. The selling price won't need to drop until competition enters the field so the only thing an inferior memory pairing might do at this point is lower costs and increase their gross margins.

    That all said, I don't think nVidia is "purposely holding back" either. Add-In-Board partners pairing the GPU with higher speed GDDR5X is entirely likely (Micron can go up to 12Gbps). This feels more like Kepler (600 / 700 series) where nVidia had to wait until yields were high enough for such large chips to be cost effective in the consumer market. We'll no doubt see higher bandwidth memory configurations when they get to market.

    @Yojimbo: "I doubt HBM will find its way onto a consumer graphics-oriented GPU that has a die size close to that of GP104. GDDR5X still has headroom above its implementation in the 1080."

    Agreed. Perhaps with later iterations and lower costs, but not as the technology currently stands.
  • Yojimbo - Friday, May 13, 2016 - link

    "Current evidence to the contrary. AMD and nVidia seem to be targeting two completely different market spaces. AMD states that Polaris 10 will be R9-390(X) level performance where nVidia is targeting the space above the GTX980Ti."

    First of all that isn't "evidence to the contrary", that's just an opening of opportunity. It's like saying you have "evidence" that John will rob a bank because the vault door will be left unlocked. Secondly, the money spent on R&D on Polaris is a sunk cost. The amount it costs to manufacture doesn't change significantly by purposely gimping the chip with not enough memory bandwidth. The higher the performance of the GPU relative to what else is out there the more they can charge for the GPU. Since the costs are mostly fixed, what do they have to gain by gimping the performance and selling the card for lower margins? As an aside, AMD isn't targeting the 390(X) level of performance with their flagship Polaris card, they are targeting the market segment that the 390(X) currently occupies. NVIDIA is not targeting the the 980Ti market segment with the 1080, it's targeting the 980 market segment. The new x80 market segment will have a performance greater than the 980Ti. The new x90(X) market segment will have a performance greater than that of the 390(X).

    "Why? They are selling these things for $600/$700. Using an inferior memory pairing doesn't magically lower that price. The selling price won't need to drop until competition enters the field so the only thing an inferior memory pairing might do at this point is lower costs and increase their gross margins."

    You're right. Pairing the cards with debilitating memory doesn't magically lower the price they can charge for the GPUs, it does so in a quite logical and understandable way. Even without AMD competing directly in the segment (although they will be with Fury), the greater the performance delta between the 1080 and Polaris 10, the more NVIDIA can charge for the 1080. Also consider that AMD will eventually release Vega, which NVIDIA will eventually have to compete with. If NVIDIA has to re-engineer a new card and worry about properly clearing the inventory of the 1080 in order to compete with the Vega that's going to be enormously inefficient. It just makes a whole lot more sense to engineer and sell a non-gimped card to begin with.

    "That all said, I don't think nVidia is "purposely holding back" either."

    You also said:
    "@Yojimbo: "Purposely holding it back? ... They have to worry too much about competition for that."

    Current evidence to the contrary. AMD and nVidia seem to be targeting two completely different market spaces. AMD states that Polaris 10 will be R9-390(X) level performance where nVidia is targeting the space above the GTX980Ti."

    Which is it? Do you think they are purposely holding it back or not? Why are you bothering to try to provide "evidence" for something you don't believe?
  • xthetenth - Tuesday, May 17, 2016 - link

    1080 is priced much less on price/performance and more on being the fastest card right now. Taking extra time to come out with faster memory to make it faster is directly counterproductive to that appeal.
  • xthetenth - Tuesday, May 17, 2016 - link

    They're not purposely holding back Pascal, they're building it to get it out now, rather than in 2017, which is when HBM2 is going to be available in enough supply to sell to customers, or however long it takes to get better clocks out of GDDR5X.
  • aznchum - Thursday, May 12, 2016 - link

    I believe the GK 110 was a 384-bit bus with 288 GB/s of VRAM bandwidth.
  • LarsonP - Friday, May 13, 2016 - link

    GDDR5X is right here... right now.
    See you in Q2 2018 HBM 2 !
  • Yojimbo - Friday, May 13, 2016 - link

    You mean Q2 2017?
  • willis936 - Saturday, May 14, 2016 - link

    We're in Q2 2017 right now :p
  • Yojimbo - Saturday, May 14, 2016 - link

    Only if you're operating on NVIDIA fiscal year time.

Log in

Don't have an account? Sign up now