Personal View site logo
Make sure to join PV on Telegram or Facebook! Perfect to keep up with community on your smartphone.
Please, support PV!
It allows to keep PV going, with more focus towards AI, but keeping be one of the few truly independent places.
NVidia Ampere GPUs from RTX 3050 to RTX 3090 Ti
  • Topic for all news related to new GPUs.

  • 161 Replies sorted by
  • GA100 largest chip

    • 800mm2 die
    • 55 Billion transistors
    • 2304 Gbps bandwidth speed
    • 6 HMB memory stacks

    GA101 half performance chip

    • 400mm2 die
    • Around 30 Billion transistors
    • Half memory speed
    • 3 HMB memory stacks
    • Cancelled
  • Nvidia made bet on Samsung 5nm EUV process

    image

    Now it is absolutely not ready for anything.

    Even in 2020 it is expected only small test manufacturing with level of defects reaching 70-80%.

    We can see Nvidia chips delays as much as 2 years (original release must had been late 2019), and they can see light in 2021-2022.

    Rumors are that original plans had been to use 7nm by Samsung, but results had been so horrible that Nvidia decided to wait for mythical 5nm EUV.

    sa10801.jpg
    800 x 477 - 44K
  • According to Chris Caso via HKEPC, an analyst at Raymond James, NVIDIA is aiming to unveil high-end Ampere GPUs at Computex 2020. Hosted in Taipei from June 2 through June 6, Computex 2020 will probably mark the arrival on the RTX 3070, RTX 3080 and RTX 3080 Ti, with midrange cards appearing in Q3.

    We may see NVIDIA unveil the Ampere architecture early than June though, albeit not as consumer GPUs. Instead, Caso claims that the company will use GTC 2020 in March to demonstrate Ampere-based Tesla GPUs.

    I think it won't happen in 2020. And instead we'll see anther rehash of existing cards with memory upgrades.

  • Some first leaks

    • GA103: 60 blocks, 3480 CUDA cores, 320-bit video memory bus, 10 or 20GB GDDR6 video memory capacity.
    • GA104: 48 blocks, 3072 CUDA cores, 256-bit video memory bus, 8 or 16GB GDDR6 video memory capacity.
    • 7nm architecture
    • Card names will be RTX 3080、RTX 3070
    • RTX 3080 Ti and Titan will have GP102 chip.
  • image

    image

    sa11929.jpg
    800 x 781 - 180K
    sa11930.jpg
    767 x 916 - 193K
  • GA100 will support up to 8,192 CUDA cores. Additionally, the GA100 will feature up to 48 GB of HBM2e VRAM and a TDP of 300 W, while its core can apparently clock up to 2.2 GHz.

    GA102 will support 5,376 CUDA cores and 12 GB of VRAM. The GA102 will apparently feature 7 graphics processing clusters (GPCs) too, all the while achieving 40% better performance than the RTX 2080 Ti.

  • Graphics chip maker NVIDIA in May will release a new generation of 7nm graphics chip products, produced by TSMC.

    It is strong rumors that delay had been caused by catastrophic Samsung 7nm process issues.

  • Ampere expected to offer a 10-20% IPC increase, 4x the ray tracing performance, and clocks that can be easily pushed to 2 GHz.

    image

    image

    sa13008.jpg
    800 x 514 - 67K
    sa13009.jpg
    800 x 469 - 64K
  • image

    image

    image

    sa13063.jpg
    715 x 822 - 111K
    sa13064.jpg
    800 x 391 - 102K
    sa13065.jpg
    722 x 960 - 189K
  • Energy efficiency... gone, first photos

    image

    image

    image

    sa13331.jpg
    800 x 594 - 65K
    sa13332.jpg
    800 x 361 - 57K
    sa13333.jpg
    800 x 360 - 41K
  • image

    sa13348.jpg
    800 x 586 - 67K
  • Rumored initial lineup

    image

    Look at TDP figures.

    It is already clear that thermal transfer limits the GPU capabilities and with move to finer 5nm or even 3nm it'll become much worse (as leaks also will rise).

    In 2-3 years we'll see big tragedy on GPU market.

    Rumored prices will be +25-35% compared to present RTX lineup, performance will be +20-25%.

    sa13358.jpg
    742 x 164 - 28K
  • image

    Present rumors are that cooling design is product of management and marketing departments, engineers had been told to shut up.

    In reality all good GPU cards manufacturers will offer water block based versions as best performing, air cards themselfs will be very long 3 deep biig coolers.

    sa13370.jpg
    800 x 620 - 58K
  • image

    Cooler uses 4 heatpipes.

    Rumors are that GPU will also use very dynamic frequency regulation, as card can have big issues keeping 320-350w consumption long term, so GPu can start behaving more like modern CPUs with short time bursts and instant frequency drop.

    sa13373.jpg
    800 x 596 - 78K
  • Possible benchmark results

    image

    sa13468.jpg
    749 x 352 - 52K
  • There are rumors saying that it will have a second GPU accelerator in tht second fan. Only tensor core

  • Traversal co-processor that’s the name !!!

  • image

    sa13480.jpg
    794 x 866 - 125K
  • ASUS card photo

    image

    sa13620.jpg
    748 x 445 - 53K
  • Another interesting info

    image

    Nvidia try to make Samsung 8nm work for almost 2 years now.

    We can expect new lineup to be most problematic in recent Nvidia history.

    As they tried to use TSMC AMD blocked the contract.

    sa13629.jpg
    734 x 292 - 37K
  • image

    sa13687.jpg
    701 x 282 - 32K
  • image

    RTX 3050 also will come, 128bit memory and most probably only 4-6GB RAM.

    sa13727.jpg
    738 x 866 - 131K
  • image

    sa13813.jpg
    480 x 150 - 12K