3 ms·

No, the next generation of HBM is expected for Vega, which should be released by the end of 2016 or more likely beginning of 2017. It seems AMD and Nvidias firs

No, the next generation of HBM is expected for Vega, which should be released by the end of 2016 or more likely beginning of 2017. It seems AMD and Nvidias first 14nm GPUs will only use gddr5 and gddr5x.

dogma1138 11y ago
Pascal will definitely use HBM in it's initial release, the cut down mid range versions later might come with GDDR5/GDDR5X. Anyhow atm HBM doesn't seem to offer much other than increase the cost and thermal envelope of the package. Nvidia for 3 generations now have been continuously shipping cards with lower memory bandwidth (with heavy focus on compression) than AMD even without HBM and continuously not only remaining competitive but managing in most cases to be the top performer. Memory bandwidth is already not being saturated even with GDDR5/4 @ 256bits, HBM has some other benefits than raw bandwidth but those for the most part aren't very noticeable as far as gaming goes. For compute the memory performance is almost completely irrelevant as you are working with very small batches and the biggest latency is in the compute operations them selves most people that build compute setups tend to under clock the memory considerably in order to save power (which can be used to further overclock the GPU) and to reduce the thermal output by as much as possible as they usually stack multiple cards into a single machine.
- redtuesday 11y ago
  > Pascal will definitely use HBM in it's initial release, the cut down mid range versions later might come with GDDR5/GDDR5X. The newest rumors I'm aware of (from around GDC) suggest Nvidia will realease the samller GP104 with GDDR5/GDDR5X to consumers before the bigger GP100 with HBM2. This sounds reasonable since Sk Hynx (which developed HBM together with AMD) will only start mass production in Q2/Q3 of 2016 - but maybe Nvidia has more luck with Samsung. Do you know of newer rumors? Maybe we know more after GTCs opening keynote from Nvidias CEO. [1] > Anyhow atm HBM doesn't seem to offer much other than increase the cost and thermal envelope of the package. Yeah, and HBM allows for much smaller high end gpus (like AMD Nano) and is much more power efficient, which means AMD and Nvidia could use the energy savings to clock the cards higher (but I have my doubts, AMDs Fury had not the best clocks, and overclocking was bad compared to cards with gddr). > Nvidia for 3 generations now have been continuously shipping cards with lower memory bandwidth (with heavy focus on compression) than AMD even without HBM and continuously not only remaining competitive but managing in most cases to be the top performer. That's why I'm looking forward to DirectX12 and Vulkan. AMDs drivers couldn't bring out the full power of their cards but that seems to change thanks to the new APIs if newest benchmarks are an indicator. Nvidia on the other hand doesn't gain much because their Dx11 drivers were already really good. [1] https://mygtc.gputechconf.com/form/session-listing&do-search=true&filter-type=complete&schedule-day-filters=Tuesday&session-keyword-filter=&session-keyword-filter-field=&session-level-filters=&session-tag-filters=&session-type-filters=Keynote https://mygtc.gputechconf.com/form/session-listing&do-search...
  - dogma1138 11y ago
    Nvidia tends not to release cut down versions first when they already have a mid-range card with the 960 and 970 in play there is very little reason to release a new core (like they did with the 770ti). They do sometimes tend to release a new gen as a mobile GPU fist like they've done with the GTX800 series but it looks now that they are no longer will make dedicated GPU SKU's but rather let OEM's use desktop cards like they've done with the 980. As far as HBM goes it doesn't really reduces power consumption, the R9 Fury Nano has a higher power consumption than a GTX 980 and the Nano is quite a bit underclocked compared to the Fury and the Fury X. Also Vulkan and DX12 performance boosts for AMD are well kinda meh, Vulkan is currently a performance loss and DX12 more or less depends on the game Ashens which is less DX12 and more Mantle spec allows them to remain competitive against Nvidia however the Fable benchmarks showed a different picture. The easiest way to sum it up currently is that Ashens uses compute shaders which are more or less designed around the specs for Mantle so they are optimized for lots of small batches rather than fewer bigger ones, the whole idea that DX12 some how takes out optimization out of the picture is nonsense it only shifts the optimization from driver to the game developer. Both Vulkan and DX12 eventually will get driver optimizations, game developers are incapable for shipping functional code even with DX10/11 managed profiles which have quite good validation. But in any case both in any DX12 benchmark overclocking the memory does absolutely squat as far as performance goes while GPU overclocking gives almost 1 to 1 increase in performance, on most modern GPU's you can underclock the memory by upto 30% before you notice any performance drops as far as most games go.

dogma1138 11y ago

Pascal will definitely use HBM in it's initial release, the cut down mid range versions later might come with GDDR5/GDDR5X. Anyhow atm HBM doesn't seem to offer much other than increase the cost and thermal envelope of the package. Nvidia for 3 generations now have been continuously shipping cards with lower memory bandwidth (with heavy focus on compression) than AMD even without HBM and continuously not only remaining competitive but managing in most cases to be the top performer. Memory bandwidth is already not being saturated even with GDDR5/4 @ 256bits, HBM has some other benefits than raw bandwidth but those for the most part aren't very noticeable as far as gaming goes. For compute the memory performance is almost completely irrelevant as you are working with very small batches and the biggest latency is in the compute operations them selves most people that build compute setups tend to under clock the memory considerably in order to save power (which can be used to further overclock the GPU) and to reduce the thermal output by as much as possible as they usually stack multiple cards into a single machine.

redtuesday 11y ago
> Pascal will definitely use HBM in it's initial release, the cut down mid range versions later might come with GDDR5/GDDR5X. The newest rumors I'm aware of (from around GDC) suggest Nvidia will realease the samller GP104 with GDDR5/GDDR5X to consumers before the bigger GP100 with HBM2. This sounds reasonable since Sk Hynx (which developed HBM together with AMD) will only start mass production in Q2/Q3 of 2016 - but maybe Nvidia has more luck with Samsung. Do you know of newer rumors? Maybe we know more after GTCs opening keynote from Nvidias CEO. [1] > Anyhow atm HBM doesn't seem to offer much other than increase the cost and thermal envelope of the package. Yeah, and HBM allows for much smaller high end gpus (like AMD Nano) and is much more power efficient, which means AMD and Nvidia could use the energy savings to clock the cards higher (but I have my doubts, AMDs Fury had not the best clocks, and overclocking was bad compared to cards with gddr). > Nvidia for 3 generations now have been continuously shipping cards with lower memory bandwidth (with heavy focus on compression) than AMD even without HBM and continuously not only remaining competitive but managing in most cases to be the top performer. That's why I'm looking forward to DirectX12 and Vulkan. AMDs drivers couldn't bring out the full power of their cards but that seems to change thanks to the new APIs if newest benchmarks are an indicator. Nvidia on the other hand doesn't gain much because their Dx11 drivers were already really good. [1] https://mygtc.gputechconf.com/form/session-listing&do-search=true&filter-type=complete&schedule-day-filters=Tuesday&session-keyword-filter=&session-keyword-filter-field=&session-level-filters=&session-tag-filters=&session-type-filters=Keynote https://mygtc.gputechconf.com/form/session-listing&do-search...
- dogma1138 11y ago
  Nvidia tends not to release cut down versions first when they already have a mid-range card with the 960 and 970 in play there is very little reason to release a new core (like they did with the 770ti). They do sometimes tend to release a new gen as a mobile GPU fist like they've done with the GTX800 series but it looks now that they are no longer will make dedicated GPU SKU's but rather let OEM's use desktop cards like they've done with the 980. As far as HBM goes it doesn't really reduces power consumption, the R9 Fury Nano has a higher power consumption than a GTX 980 and the Nano is quite a bit underclocked compared to the Fury and the Fury X. Also Vulkan and DX12 performance boosts for AMD are well kinda meh, Vulkan is currently a performance loss and DX12 more or less depends on the game Ashens which is less DX12 and more Mantle spec allows them to remain competitive against Nvidia however the Fable benchmarks showed a different picture. The easiest way to sum it up currently is that Ashens uses compute shaders which are more or less designed around the specs for Mantle so they are optimized for lots of small batches rather than fewer bigger ones, the whole idea that DX12 some how takes out optimization out of the picture is nonsense it only shifts the optimization from driver to the game developer. Both Vulkan and DX12 eventually will get driver optimizations, game developers are incapable for shipping functional code even with DX10/11 managed profiles which have quite good validation. But in any case both in any DX12 benchmark overclocking the memory does absolutely squat as far as performance goes while GPU overclocking gives almost 1 to 1 increase in performance, on most modern GPU's you can underclock the memory by upto 30% before you notice any performance drops as far as most games go.