Announcement

Collapse
No announcement yet.

The Tesla P100 Is NVIDIA's New & Most Powerful Accelerator

Collapse
X
 
  • Filter
  • Time
  • Show
Clear All
new posts

  • #11
    This GP100 is a real beast. For those who are concerns about FP16/32/64 mix can have a look at https://devblogs.nvidia.com/parallel...inside-pascal/ where clearly the GP100 is a real good beast at 16, 32 as well as 64 FP compute. And all of this fits in the 300W range which is really nice.

    Comment


    • #12
      Maybe somebody should read the announcment correctly, possible OEM cards like a successor to current GTX 980 Ti will not be available this year. They might sell a Titan upgrade at premium price later this year, depending on the yield i would guess. Tesla is extremely expensive. So you can soon expect some GDDR5X cards with Pascal chips for average gamers.
      Last edited by Kano; 06 April 2016, 06:56 AM.

      Comment


      • #13
        Originally posted by jaxxed View Post
        That thing is a beast. It's impressive that they got the HBM2 in it. I thought it wasn't certain if there was going to be stable production of HBM2 in time. NVIDIA gets what they want I guess.
        Not really, Samsung announced in January that they started mass production. Also, GP100 is nowhere near buyable, they weren't even able to show a running prototype.

        Originally posted by adakite View Post
        For those who are concerns about FP16/32/64 mix can have a look at https://devblogs.nvidia.com/parallel...inside-pascal/ where clearly the GP100 is a real good beast at 16, 32 as well as 64 FP compute. And all of this fits in the 300W range which is really nice.
        It was not a concern about performance. It's actually a really good thing, first-time double the throughput for FP16 and utilizing the same ALUs. It's just that they haven't been able to implement mixed precision all the way from 16 up to 64 bits float for the same ALUs. That costs transistors and die space and 600 sqmm is quite huge for what it offers, imho. Desktop Pascals will be interesting, I just hope they also put HBM on GP104, leave FP64 and NVLink out for a really great consumer product.

        Originally posted by Kano View Post
        Maybe somebody should read the announcment correctly, possible OEM cards like a successor to current GTX 980 Ti will not be available this year. They might sell a Titan upgrade at premium price later this year, depending on the yield i would guess. Tesla is extremely expensive. So you can soon expect some GDDR5X cards with Pascal chips for average gamers.
        I'm not even sure there will ever be a Titan or xxx Ti card with GP100. Rumors have a GP102, too this time (out of driver entries). They did not draw raster engines in the block diagram. Maybe this is a pure computing chip only intended for Tesla usage.
        If GP102 or GP104 only have GDDR5(X) and not HBM, that would be really lame.
        Last edited by juno; 06 April 2016, 07:03 AM.

        Comment


        • #14
          Mass production is an elastic term, do you expect good yields in the beginning? Usually it will get cheaper over time. AMD used 4 GB HBM and from benchmarks it was less impressive than from the specs. GDDR5X supports the double data rate over GDDR5, so memory speed should not be so problematic and should be at least 25 % faster than the HBM from the Fury X.

          Comment


          • #15
            Of course, I don't know how Samsung's HBM yields are. The low clocked DRAM itself should be easy to do. Samsung also has some years of experience with TSVs. There are still some things that could go wrong, obviously.
            Fiji's problem was the limited capacity of HBM at this time, not the technology itself. And it is not drowning in insufficient memory bandwidth, so what should have benchmarks been showing in your opinion? I also see Fiji as some sort of a test drive vehicle for HBM and 2.5D stacking. Of course it has drawbacks due to the old 28 nm process.

            I also don't get what you are trying to say about memory bandwidth. GP100 seems to utilise 1.4 Gbps (720 GiB/s), Fiji had 1.0 Gbps (512 GiB/s), both at 4096 bits wide interfaces. GDDR5X will provide 10 Gbps from the start, using a sane 256 bit interface, resulting in 320 GiB/s.

            Comment


            • #16
              Originally posted by jaxxed View Post
              That thing is a beast. It's impressive that they got the HBM2 in it. I thought it wasn't certain if there was going to be stable production of HBM2 in time. NVIDIA gets what they want I guess.
              There's HBM2 to be had, it's just that the production is still ramping up, thus the limited availability.

              Originally posted by Kano View Post
              Maybe somebody should read the announcment correctly, possible OEM cards like a successor to current GTX 980 Ti will not be available this year. They might sell a Titan upgrade at premium price later this year, depending on the yield i would guess. Tesla is extremely expensive. So you can soon expect some GDDR5X cards with Pascal chips for average gamers.
              I thought it was common knowledge that for this round HBM(2) will be reserved to the top tier cards (meaning Tesla, Titan and maybe the successor of the 980Ti). Because of the above.

              Edit: Also, I'm thoroughly disappointed in the announcement, as I was expecting something consumer oriented to be announced as well. As it is now, we still know nothing about Pascal, besides some SM and FPxx units count.
              Last edited by bug77; 06 April 2016, 09:58 AM.

              Comment


              • #17
                It appears that GP102 will be the top GPU for GeForce products this year, which might be similar to GP100 with some of FP64 and the NVLink support stripped out. Hopefully they will add more CUDA cores to push the FP32 performance.

                Originally posted by Kano View Post
                Mass production is an elastic term, do you expect good yields in the beginning? Usually it will get cheaper over time. AMD used 4 GB HBM and from benchmarks it was less impressive than from the specs. GDDR5X supports the double data rate over GDDR5, so memory speed should not be so problematic and should be at least 25 % faster than the HBM from the Fury X.
                Yes, but we know the chips are in mass production. But then again it takes ~5 months from an empty wafer to a GPU on a PCB in a nice box.

                Comment

                Working...
                X