News The Nvidia GeForce RTX 3070: RTX 2080 Ti Performance at $499

[InvalidError]

The guys still selling their Pascal cards for outrageous sums also need to be adjusting their prices downwards.

They aren't going to lower their prices for used cards until they stop selling and until the new stuff launches, less informed people shopping for a new used card only have pre-RTX3000 benchmarks to go on. I'm not surprised at all that old GPUs still resell at prices that look outrageous to people looking into next-gen rumors and announcements.

The smart high-end GPU flippers buy their GPUs after the prices settled down from launch and re-sell them long before the next gen arrives so they can get most of the MSRP back.

 

[animekenji]

Not really. As a good example, look at AMD's RX 580 compared to the GTX 1060, two cards that typically offer rather similar gaming performance on average. However, the RX 580 is around a 6.1 Tflop card in terms of FP32 performance, while the 1060 is around a 4.3 Tflop card. The RX 580 offers around 40% more FP32 compute performance, roughly on par with a GTX 1070, but that does not directly translate to typical gaming performance. That's a huge margin of difference, not at all what I would consider "close enough".

Likewise, look at what Nvidia had to say about the performance of these new cards. Going by quotes from the other article here...




Hmm... Let's look again at those Tflop numbers provided for the cards.

RTX 2080 = 10 Tflops
RTX 3080 = 30 Tflops

Titan RTX = 16 Tflops
RTX 3090 = 36 Tflops

So apparently, going from 10 to 30 Tflops only makes the 3080 "up to double the performance" of the 2080. And going from 16 to 36 Tflops only makes the 3090 "up to 50% faster" than the Titan RTX. If the "Tflop to gaming performance ratio" were similar, then the 3080 would be considered a 20 Tflop card, and the 3090 would be a 24 Tflop card, at least going by Nvidia's performance claims.

What this means, is that the actual gaming performance is likely around 33% lower than what these high Tflop numbers might imply. That would make the 3070's "20 Tflops" roughly comparable to the 13.5 Tflops of a 2080 Ti. That's still a good generational improvement, likely offering over 40% more performance than the 2070 SUPER currently available at that price point, but certainly not the more-than-doubling of performance that the Tflop number alone would suggest.

Tflops is a compute benchmark, not a gaming benchmark. A tripling of Tflops performance does not equate to a tripling of gaming performance. Tflops are better for miners than for gamers.

 

[InvalidError]

Tflops is a compute benchmark, not a gaming benchmark.

The main problem with TFlops numbers is the same as CPU TFlops: software needs to get written to actually leverage them and the dominant player in the CPU/GPU market gets to dictate what software developers will be designing around and optimizing for. If you have a different architecture with a different balance between resources than the dominant player, you automatically get a massive handicap from being too different from what everything currently in existence is written for - first-gen Ryzen performance was all over the place for the better part of a year while developers re-wrote and re-compiled their games to fix a plethora of performance issues and crashes caused by Ryzen being too different.

 

[San Pedro]

In Australia the price of the GPU is at 1139AUD. If I translate 699 USD to AUD and mutiply that by 1.1 for the tax, it comes to 1050. So charging 90 AUD more is not much of a difference. US has taxes too.

So I guess it's mainly Japan having to pay a $300 premium for these cards. If I get one it'll be cheaper for me to buy from amazon in the states and have it shipped here...

 

[vinay2070]

@JarredWaltonGPU Please Benchmarks When? Why nobody answer this question

 

[Schlachtwolf]

@JarredWaltonGPU Please Benchmarks When? Why nobody answer this question

The first 3080 cards come out on the 17th of september, the 3090 I think aweek after and the 3070 in October, even if they had the cards to bechmark the embargo is not yet lifted by Nvidia so until that happens benchmarks from Tom's, GN etc are still not available

 

[JarredWaltonGPU]

@JarredWaltonGPU Please Benchmarks When? Why nobody answer this question

We cannot answer because that information is under embargo. But reviews of the RTX 3080 will be live for sure by September 17 when the cards go on sale, and September 24 for the 3090. We still don't have an exact launch date for RTX 3070, other than "October," but I would guess one of the first two Thursdays will be the launch date.

 

[TUFOM_]

I'm quite certain AMD will not compete in high-end. I know terfalops is far from perfect measurement but its great approximate where card will land. RTX 3070 20 terafops, 5700XT has 10 terfalops and big navi best estimates put it x2 so that would be roughly in RTX 3070 territory. RTX 3080 30 teraflops and RTX 3090 35 teraflops, I imagine AMD simply can't touch that.

 

[JarredWaltonGPU]

I'm quite certain AMD will not compete in high-end. I know terfalops is far from perfect measurement but its great approximate where card will land. RTX 3070 20 terafops, 5700XT has 10 terfalops and big navi best estimates put it x2 so that would be roughly in RTX 3070 territory. RTX 3080 30 teraflops and RTX 3090 35 teraflops, I imagine AMD simply can't touch that.

I explained this in more detail in the Ampere Architecture deep dive, but you have to look carefully at what Nvidia has done.

Turing could do concurrent FP32 + INT. Ampere can do FP32 + INT or FP32 + FP32, but not FP32 + FP32 + INT. So if a game uses 35% of the compute power of the GPU on INT operations, that basically means 2/3 of the second FP32+INT core is just doing INT. Which means relative to Turing, the speedup is far lower than the TFLOPS would suggest. Let's do the math:

Ampere GA102 (3080) has 30 TFLOPS + 15 TOPS theoretical, but it can really only do 15 FP32 TFLOPS + 15 INT TOPS, or 30 FP32 TFLOPS.
For a hypothetical 65% FP32 + 35% INT workload, that means 50% (total) of the FP32 gets sent to the FP32-only cores and 15% to the FP32+INT cores, and the remaining 35% INT goes to the FP32+INT as well.

Turing TU104 (2080) has 10 TFLOPS + 10 TOPS theoretical. It does the full FP32 workload while additionally using 35% of the INT cores. Note that this is inefficient -- there are INT cores sitting idle 65% of the time (possibly more in some games)!

Interestingly, 0.65 * 30 = 19.5, which is basically double the 2080's 10 TFLOPS. You won't always see double the performance I don't think, but in GPU-limited situations it will probably get close.

 

[nofanneeded]

I am little upset about the new cards from mobile point of view ... Ampere is power hungry and I cant see any 100-150 watts GPU this time that can be made a Mobile chip.

dark future for Mobile gaming ...

 

[InvalidError]

Ampere is power hungry and I cant see any 100-150 watts GPU this time that can be made a Mobile chip.

Ampere has better performance per watt than Turing, mobile is only a question of scaling it down. That will probably come with the 3050 and 3060 later.

 

[cryoburner]

I'm quite certain AMD will not compete in high-end. I know terfalops is far from perfect measurement but its great approximate where card will land. RTX 3070 20 terafops, 5700XT has 10 terfalops and big navi best estimates put it x2 so that would be roughly in RTX 3070 territory. RTX 3080 30 teraflops and RTX 3090 35 teraflops, I imagine AMD simply can't touch that.

All of Nvidia's marketing materials that they have posted so far showing relative frame rates put the 3070's performance roughly around the level of a 2080 Ti, a card that's only around 50% faster than a 5700 XT. So, if AMD delivers a card with double the performance of a 5700 XT, it would likely be around 33% faster than a 3070, and closer to the performance level of a 3080. As Jarred pointed out, Nvidia changed the way FP32 works on the 30-series cards, which can make those Tflop numbers a little misleading when comparing them with Turing or AMD's existing hardware, and not really a great way to approximate where a card's performance will land, at least in today's games.

Of course, we don't know exactly how the performance of AMD's top-end RDNA2 cards will compare, or how their implementation of raytracing will perform, so it's really hard to say whether they might be a bit faster or a bit slower than that in practice. Seeing as Nvidia priced this generation of cards more competitively than the last, I imagine they expect more competition this time around though. I also suspect they have more inside knowledge about what their competitors are planning compared to the general public.

I am little upset about the new cards from mobile point of view ... Ampere is power hungry and I cant see any 100-150 watts GPU this time that can be made a Mobile chip.

dark future for Mobile gaming ...

For mobile chips, they tend to just turn the clock rates way down to improve efficiency. An RTX 2080 Max-Q, for example, has the same number of cores as a desktop RTX 2080. However, it's clocked significantly lower to get it down around 80 watts, and as a result, only performs about on par with a desktop 2060. They will probably be able to underclock a 3080 to get it down under 100 watts for a Max-Q version as well, but again, it will likely perform more like a desktop card costing half as much.

 

[nofanneeded]

Ampere has better performance per watt than Turing, mobile is only a question of scaling it down. That will probably come with the 3050 and 3060 later.

I know , but Gone are the days when mobile GTX 1080 was same performance of Desktop GTX 1080 ..

Mobile GTX 1080 was 150W TDP while Desktop was 180W TDP , not huge difference ...

We had real Desktop flagship performance in notebooks.

 

[nofanneeded]

For mobile chips, they tend to just turn the clock rates way down to improve efficiency. An RTX 2080 Max-Q, for example, has the same number of cores as a desktop RTX 2080. However, it's clocked significantly lower to get it down around 80 watts, and as a result, only performs about on par with a desktop 2060. They will probably be able to underclock a 3080 to get it down under 100 watts for a Max-Q version as well, but again, it will likely perform more like a desktop card costing half as much.

That was not the case at the GTX 1080 time , desktop GTX 1080 was 180 watts , while Mobile MXM GTX 1080 was 150 watts .. not big difference ...

Nvidia GTX 10xx generation allowed very near desktop performance for Mobile.

Max-Q is not a good option for Flagships , I wouldnt bother to pay for mobile gaming laptop that is 80 watts max. I allways avoided MaxQ.

 

[King_V]

Here's a kind of a tangential question. Any idea as to whether there's going to be lower GTX versions? Say, a GTX 17- series, without the ray-tracing?

Or is the assumption that they're all-in now with ray-tracing?

 

[JarredWaltonGPU]

Here's a kind of a tangential question. Any idea as to whether there's going to be lower GTX versions? Say, a GTX 17- series, without the ray-tracing?

Or is the assumption that they're all-in now with ray-tracing?

My assumption is that Nvidia's goes all-in on RTX. There will probably be a continuation of the MX350/MX450 stuff for laptops, but I think there's a good chance that a hypothetical RTX 3050 could cost $250 and still deliver ~2070 levels of performance. Or at least 2060 Super? And if you have RTX at $250, there's not much lower left to go -- not with the next generation of integrated GPUs basically killing everything at the $100 mark.

But Nvidia might do another GTX line with Ampere. I just don't know what it would be called. GTX 1800 series? GTX 1700? Where did we even get 1600 from!? We could do the +1000 and have GTX 2600 series. That would be just as weird as GTX 16-series, but we've come to accept it now (maybe).

 

[nofanneeded]

Here's a kind of a tangential question. Any idea as to whether there's going to be lower GTX versions? Say, a GTX 17- series, without the ray-tracing?

Or is the assumption that they're all-in now with ray-tracing?

IMO GTX or non raytracing cards will stay for a little longer ... I think they are important for mobile gaming where low TDP is more important than raytracing altogether.

I think that GTX 1080 ti performance will appear in the form of GTX 3060/50

 

[cryoburner]

My assumption is that Nvidia's goes all-in on RTX. There will probably be a continuation of the MX350/MX450 stuff for laptops, but I think there's a good chance that a hypothetical RTX 3050 could cost $250 and still deliver ~2070 levels of performance. Or at least 2060 Super? And if you have RTX at $250, there's not much lower left to go -- not with the next generation of integrated GPUs basically killing everything at the $100 mark.

But Nvidia might do another GTX line with Ampere. I just don't know what it would be called. GTX 1800 series? GTX 1700? Where did we even get 1600 from!? We could do the +1000 and have GTX 2600 series. That would be just as weird as GTX 16-series, but we've come to accept it now (maybe).

Yeah, I kind of think they would make RTX available across the lineup too, at least through the "mid-range". If a $500 RTX 3070 ends up performing roughly similar to a 2080 Ti, then 3060-class cards in the sub-$400 range could potentially offer 2080 to 2080 SUPER levels of performance. Cards in the sub-$300 range might then offer around 2070 level performance, which would in turn mean that 2060 level performance might be available for around $200. Does it make sense to cut RTX from a card similar in performance to a 2060 or 2070? I wouldn't really think so.

It does seem more likely that a 75 watt card around the $150 range might give up the feature though, as might anything lower. I don't think better integrated graphics will necessarily eliminate $100 cards. Those will still be marketable to people upgrading older and non-gaming systems, especially if they can offer performance near that of a 1650, which I don't think integrated graphics will be touching for a while still.

 

[InvalidError]

IMO GTX or non raytracing cards will stay for a little longer ... I think they are important for mobile gaming where low TDP is more important than raytracing altogether.

Jensen wants RTX for everyone, so I'm not expecting non-RTX models this time around. Mobile will be fine with under-volted and down-sized RTX - still going to perform better than anything else previously available for the same power.

 

[hotaru.hino]

My assumption is that Nvidia's goes all-in on RTX. There will probably be a continuation of the MX350/MX450 stuff for laptops, but I think there's a good chance that a hypothetical RTX 3050 could cost $250 and still deliver ~2070 levels of performance. Or at least 2060 Super? And if you have RTX at $250, there's not much lower left to go -- not with the next generation of integrated GPUs basically killing everything at the $100 mark.

But Nvidia might do another GTX line with Ampere. I just don't know what it would be called. GTX 1800 series? GTX 1700? Where did we even get 1600 from!? We could do the +1000 and have GTX 2600 series. That would be just as weird as GTX 16-series, but we've come to accept it now (maybe).

I'm hoping that NVIDIA goes all RTX if only to prevent further market segmentation. But I'm open to the possibility of the 50 or even the 30 series cards being RT-less GPUs to keep costs down.

Or they could bring Turing to 8nm. It wouldn't cost NVIDIA that much more to spin the architecture to a new node and sell off 8nm 2060s as 3050s or something.

 

[InvalidError]

Or they could bring Turing to 8nm. It wouldn't cost NVIDIA that much more to spin the architecture to a new node and sell off 8nm 2060s as 3050s or something.

There is no point in bringing RTX and DLSS to the 3050 by shrinking the 2060 when the performance is going to be so bad as to be unusable as it is on the RTX2060. For the RTX3050 to bring sufficient DLSS and RTX performance to bring Jensen's RTX supremacy to the masses, it needs to be Ampere.

Also, if you are going to have to go through the trouble of redoing layout a whole existing design to shrink it, it is likely simpler and cheaper to just go with the current-gen design where all of the function blocks have already been optimized for the new process. Shrinking an existing design mainly makes sense when you have nothing new ready to go on a new process.

Another option would be for the 3050 to be a 1/4-1/3 cut-down 3060, in which case there is practically no additional design work.

 

[hotaru.hino]

There is no point in bringing RTX and DLSS to the 3050 by shrinking the 2060 when the performance is going to be so bad as to be unusable as it is on the RTX2060. For the RTX3050 to bring sufficient DLSS and RTX performance to bring Jensen's RTX supremacy to the masses, it needs to be Ampere.

Also, if you are going to have to go through the trouble of redoing layout a whole existing design to shrink it, it is likely simpler and cheaper to just go with the current-gen design where all of the function blocks have already been optimized for the new process. Shrinking an existing design mainly makes sense when you have nothing new ready to go on a new process.

Another option would be for the 3050 to be a 1/4-1/3 cut-down 3060, in which case there is practically no additional design work.

I don't think the 2060 (at least the Super variant) is a terrible performer for the price category it would be targeting. i.e., the $120-$150 market, even with RT . With regards to using what is already done, NVIDIA was willing to spin off a completely different version of Turing for the GeForce 16 and the GTX 1050 was its own version of Pascal.

But I'll give it that using Ampere is more likely. However, I don't think it'll be a cut-down 3060. Especially if for some reason NVIDIA makes it a 6GB card. The 50 SKU has mostly been its own thing since NVIDIA redid the numbering.

 

[InvalidError]

I don't think the 2060 (at least the Super variant) is a terrible performer for the price category it would be targeting. i.e., the $120-$150 market, even with RT.

The GTX1650S was a ~$150 part, $120 would be the GT1030 and at that point you are better off using an IGP.

With the RTX3000 lineup performing two to two and a half model numbers above the previous generation, an RTX3050 would be performing in the neighborhood of an RTX2070S which will definitely require at least 6GB of VRAM to keep reasonably well fed, so I'm expecting the RTX3050 to be a $170-200 part.

 

[hotaru.hino]

so I'm expecting the RTX3050 to be a $170-200 part.

If that's the case, I'm going to enjoy all of the groaning about how the 50 SKU increased in price despite what you get for it.

 

[InvalidError]

If that's the case, I'm going to enjoy all of the groaning about how the 50 SKU increased in price despite what you get for it.

The groaning with most of the previous generation is because you pay 20-30% more for 20-30% more performance so you get near-zero net gain in performance per dollar instead of the usual ~30%.

If the RTX3050 pans out like I predicted, I think most people won't mind paying 20-30% extra for 50% more VRAM, 50-80% more VRAM bandwidth and ~150% more GPU-power too much.

I don't particularly cherish the idea of entry-level rising to $200 (I'm still using a GTX1050) but $150 GPUs make little sense when IGPs are catching up and dGPUs require more VRAM to deliver the details they are capable of delivering.