In a brief news post made to their GeForce website last night, NVIDIA has announced that they have delayed the launch of the upcoming GeForce RTX 3070 video card. The high-end video card, which was set to launch on October 15th for $499, has been pushed back by two weeks. It will now be launching on October 29th.

Indirectly referencing the launch-day availability concerns for the RTX 3080 and RTX 3090 last month, NVIDIA is citing a desire to have “more cards available on launch day” for the delay. NVIDIA does not disclose their launch supply numbers, so it’s not clear just how many more cards another two weeks’ worth of stockpiling will net them – it likely still won’t be enough to meet all demand – but it should at least improve the odds.

NVIDIA GeForce Specification Comparison
  RTX 3070 RTX 3080 RTX 3090 RTX 2070
CUDA Cores 5888 8704 10496 2304
ROPs 96 96 112 64
Boost Clock 1.725GHz 1.71GHz 1.7GHz 1.62GHz
Memory Clock 14Gbps GDDR6 19Gbps GDDR6X 19.5Gbps GDDR6X 14Gbps GDDR6
Memory Bus Width 256-bit 320-bit 384-bit 256-bit
VRAM 8GB 10GB 24GB 8GB
Single Precision Perf. 20.4 TFLOPs 29.8 TFLOPs 35.7 TFLOPs 7.5 TFLOPs
Tensor Perf. (FP16) 81.3 TFLOPs 119 TFLOPs 143 TFLOPs 59.8 TFLOPs
Tensor Perf. (FP16-Sparse) 163 TFLOPs 238 TFLOPs 285 TFLOPs 59.8 TFLOPs
TDP 220W 320W 350W 175W
GPU GA104 GA102 GA102 TU106
Transistor Count 17.4B 28B 28B 10.8B
Architecture Ampere Ampere Ampere Turing
Manufacturing Process Samsung 8nm Samsung 8nm Samsung 8nm TSMC 12nm "FFN"
Launch Date 10/15/2020
10/29/2020
09/17/2020 09/24/2020 10/17/2018
Launch Price MSRP: $499 MSRP: $699 MSRP: $1499 MSRP: $499
Founders $599

Interestingly, this delay also means that the RTX 3070 will now launch after AMD’s planned Radeon product briefing, which is scheduled for October 28th. NVIDIA has already shown their hand with respect to specifications and pricing, so the 3070’s price and performance are presumably locked in. But this does give NVIDIA one last chance to react – or at least, distract – should they need it.

Source: NVIDIA

Comments Locked

105 Comments

View All Comments

  • MrVibrato - Friday, October 2, 2020 - link

    Oi. I just noticed that there is actullay the wrong transistor count for the TU106 in the table. Oops...
  • MrVibrato - Friday, October 2, 2020 - link

    (The TU102 has 18.6 bln transistors. TU102 is in the 2080 Ti)
  • drexnx - Friday, October 2, 2020 - link

    Mainly because you can't compare the CUDA core counts vs. Turing, but also because the chart is wrong, that's the TU102 transistor count.

    for Turing they added an INT core for each CUDA core, but didn't count that core anywhere. If you look at the block diagrams, sure, they SEEM like they took half of the cores in each SM and turned them into INT cores, but they actually doubled the SM count at the same time.
    ie. 1080 = 2560 CUDA cores across 20 SMs, 2070S = 2560 CUDA cores across 40 SMs, plus another 2560 INT32 cores.

    For Ampere they also added FP32 capability to that secondary INT core, so now they're counting it as a full CUDA core. Makes sense, but as you can tell by the game performance, it's really not working like that. Part of the advantage of Turing vs. Pascal was that those "hidden" INT cores added concurrency to 36:100 game instructions (INT:FP) per the Turing architecture whitepaper. Now that's gone and each Ampere CUDA core is more comparable to a Pascal CUDA core for performance scaling purposes - granted there's much more granularity on Ampere, as Pascal could only do FP or INT operations on a group of CUDA cores - i.e. all 128/SM had to be INT for that clock, even if it was 1 INT operation, but those previously uncounted INT32 cores are now counted.

    as for marketing, of course the FLOPs number looks great compared to Turing because a million INT32 cores can't even do 1 FLOP. For that reason it looks 2.x+ as powerful at pure FP math because it is. But that doesn't matter in applications.
  • Meteor2 - Saturday, October 3, 2020 - link

    That's a great explanation. Thanks.
  • Spunjji - Monday, October 5, 2020 - link

    Didn't see you'd answered this already before putting in my own answer... and yours was better. 😅
  • Gigaplex - Saturday, October 3, 2020 - link

    Also don't forget the architecture is different, so the definition of what makes up a CUDA core is different.
  • Spunjji - Monday, October 5, 2020 - link

    In Turing they had parity between FP32 and INT32 resources, which resulted in some inefficiency (those workloads are not equal). With Ampere the INT32 resources can also perform FP32 calculations, so if all you're doing is FP32 then there's double the capacity - but for graphics workloads it doesn't work that way.

    In short: some things that were "invisible" in their quoted specs have now been made visible in a manner that, for gamers at least, is very misleading.
  • Sychonut - Friday, October 2, 2020 - link

    Right on time for winter to keep your house warm and cozy.
  • Gigaplex - Saturday, October 3, 2020 - link

    Halloween isn't winter. Depending on the hemisphere, it's either spring or autumn.
  • catavalon21 - Sunday, October 4, 2020 - link

    By the time many are available at retail, it will no longer be Spring/Fall. Unless it's the one six months from now.

Log in

Don't have an account? Sign up now