Gadgetory

and video detailed more of its Volta architecture yesterday a name that we first covered in 2015 where Volta was shown as targeted for 2018 delivery Nvidia seems to be on target for 2018 with its initial server grade Tesla V 100 accelerators shipping and dgx servers by quarter three of this year V 100 is a major architecture milestone for Nvidia and marks a divergence from architectures that more closely resemble the likes of Kepler Maxwell and Pascal today we're detailing the initial V 100 layout before that this video is brought to you by courses at new Vengeance RGB LED Ram which ships with custom screens ICS for better overclocking performance and stability given that memory is highly relevant for performance with new rise in CPUs now is a good time to do research on high performance kits start with the Vengeance RGB LED kit at the link in the description below the 100 is not a consumer products just like P 100 was not but eventually we'll probably have something like V 102 which will be consumer targeted and would effectively be an 1180 TI if we were to take today's now in clay Chur and it looks like nvidia is on target for 2018 delivery of potential consumer grade Volta architecture implementation on video cards that would be in the gaming market for today we're looking at the initial Volta architecture that includes things like the new tensor cores which are not really something that gamers will be able to make use of but are critical to business operations deep learning HPC all kinds of things that involve data heavy crunching and compute whereas later graphics versions of Volta will change in their organization somewhat just like with Pascal P 100 was 64 cores for SM whereas the one of the two chips and all the other consumer chips were at 128 cores per SM so there's a bit of difference between compute and graphics products but the base architecture is the same so we can start with V 100 and see what it looks like today starting with the block diagram this is the full-size V 100 block diagram as both the top one of the largest pieces of silicon we've seen on a GPU the eye size at 4 V 100 is 815 millimeters squared around 30% larger than P 1 hundreds already large 610 millimeters squared for comparison consumer grade at 1080 tig through one of two chips are 471 a millimeters just give some perspective fold v100 hosts 53 76 FP 32 CUDA cores across 84 SM is using similar organizational hierarchy to the GP 100 at 64 cores per SM with a 1 to 2 FP 64 ratio and 2 to 1 FP 16 ratio the result is 15 Dara flops of single-precision compute about 8 teraflops of double-precision and about 30 teraflops of FP 16 or half precision compute for deep learning FP 16 is useful as the precision is not needed when working with such large data matrices despite similar organizational layout the P 100 the actual architecture is vastly different scheduling and thread execution rings may have been reinvented with V 100 but we don't really know how just yet and the introduction of tensor cores complicates things so alongside the 53 76 CUDA cores which were fairly familiar with at this point there are also 672 tensor chords these are the new ones each SM holds 8 tensor cores just like each SM holds 64 CUDA cores and a tensor core is targeted at HPC deep learning machine typed neural net applications whereas CUDA cores are more known they can work and apply to all these things especially with FP 16 towards deep learning but tensor cores are special that's because they are a cluster of al use in a single cores with a single unit that specializes in FMA operations refused multiply add and it can execute those operations across 2 4 by 4 matrices what does that mean well it means that each tensor core can execute 64 FMA s per clock so 64 fuse multiply add operations per clock of a single tensor core which means that you're roughly quadrupling what a previous coup de coeur could do in the same amount of time one clock cycle in terms of FMA operations so that is helpful for compute work it's helpful again for deep learning it's a huge deal in that regard for us gaming not really something we care about right now anyway nothing that we can see an immediate use for but that means that doesn't mean it's it's irrelevant it's obviously a huge deal for the Volta architecture tensor quadrupling coup de coeur execution throughput in terms of flops per clock for those server-side very expensive money driving businesses that's a big deal so that's the main thing with v100 that end video was talking about at their event tensor cores are something to pay attention to but not something that you need to worry about as a consumer anytime soon here's a shot of the SM layout HSM it contains the two tee pcs that contain blocks of FP 64 FP 32 int and tensor core units just like always except for the additional tensor course this is just a block diagram and not necessarily representative of actual diet space allocation to the tensor cores but it's looking like a large part of the die space of the 100 is allocated to this new type of core the usual 40 mus are present on each SM totaling 336 TM use across 84 screen multi processors to P 100 to 240 TM use across 60 SMS these are indicated by the text marks on the bottom of the diagram and that's again an SM layout diagram whereas the block diagram shows them all working together the architecture will be built on a new 12 a nanometer FinFET process which Nvidia and tsmc they stick the name and video on it so it's thin set Nvidia which is just special optimizations that we don't know about made to the process by the SMC so 12 nanometers it's a completely new architecture this isn't Pascal base it's not Maxwell or Kepler base or anything like that and the die size again 815 millimeters squared totally in 20 1.1 billion transistors that is at the reticle limit for what TSMC can manufacture that is the absolute hard limit of their manufacturing tools the physical tools in the factory the fab I should say you can't really get bigger than that right now and when the eye size increases on GPUs if you're not aware it generally means that yield goes down this is just part of why the bigger GPUs like something you'd find on a 1080p eye are more expensive than the smaller ones which are physically smaller yield goes down to Costco up because there's greater risk involved you lose more of your silicon each run you're doing of the wafers so that drives up cost but for consumers it eventually comes down our way because as these large enterprise b2b type operations pay for the new technology because they can really benefit from it the cost goes down the yields go up the die sighs eventually shrinks a bit for something like an eleven atti or whatever it may be in the future that's why you normally see these GPUs on gaming grade cards smaller than the accelerated cards the GB 100 accelerator is the first product shipping with Volta architecture it is a cut-down version of the v1 100 block diagram that we just went through not back cut down it's 50 120 CUDA cores rather than the 53 76 that the full of uLTA block is capable of from what we've seen so far so it's a little cut down and not much and then it is HP m2 supported just like the GP 100 accelerator was that would be the previous tesla card high in tesla card and HBM 2 here is limited to 16 gigabytes they can't go higher until someone can manufacture HP m with a greater count of dies per stack and because nvidia is sticking to four stacks for this new GPU just like for GP 100 accelerators that means they're limited to 16 gigabytes of HP m - it's still a 4096 bit bus just like GP 100 was but the efficiency has gone up from what they've said and we don't really have a whole lot of information on volt at a low level just yet we've got a decent out of starter information but efficiency should have gone up that also means that there could be a change in perk or efficiency which is something we've seen the past architectures for example Kepler - Maxwell there's something like a 30% increase in performance per watt out of the cores or just efficiency overall which means that you can't always compare just raw core count across architecture even from the same manufacturer because those cores might be physically capable of a different amount of throughput or different operations or whatever depending on which art you're looking at how old it is things like that so our optimizations of course we don't know the details just yet other than top level things like threat execution and organization have changed but no no real details to speak of right now that said they should be coming out pretty soon because white papers for these new architectures normally get pushed pretty shortly after the GTC event which is where this was announced and is in the process of concluding this weekend so a consumer target looks like next year is pretty likely 2018 that's what NVIDIA has said all along for the most part so that's not changed if you're in the server business and for some reason watching us and I guess the thing for you to know would be that this GV 100 accelerator cost $18,000 standalone or you can buy a server a dgx server which has it's got the full specs online on their site that's $150,000 and means you get them earlier quarter three but if you're in the business of making money off these things I guess getting it earlier is probably something that would make you money but that's not really Howard Territory so we'll wait for Volta on consumer products but as more information on GV 100 comes out we'll deep dive it just like we did with GPU 100 and of course Vega is theoretically happening at some point so we'll stay tuned for that as well thank you for watching as always you can go to patreon.com/scishow and ice themselves out with this type of reporting or you can go to store that gamers nexus dotnet where we've got shirts like this one or the graphed logo shirt now restock Singh cotton subscribe for more I'll see you all next time you