Nvidia Ampere Discussion [2020-05-14]

For all the criticism of Ampere's efficiency, RTX3070 gives pretty much 2080ti performance, including path traced games like Quake2 rtx and minecraft rtx, and it does so with less power, fewer tensor cores, fewer RT cores, fewer TMUs and less memory bandwidth. It has more shader cores, more clock speed and more ROPs.

Leads me to believe, as most people have said, that RT performance is not bottlenecked by the actual RT core functions, but most likely by pure shader performance, where the 3070 has an advantage.

For a "non gaming" architecture, ampere is showing a very nice generational improvement with the 3070.

 

Interesting to see the variance in the 3070 reviews. TechPowerUp has the 3070 in a dead heat with the 2080Ti across all resolutions, within +/- 1% on average.
On the other hand, in the review from Igor's Lab the 3070 is leaving the 2080Ti in the dust, by ~10% on average at lower resolutions, with the gap narrowing somewhat at 4k to maybe something like a 5% performance uplift.

I wonder if some sites are using heavily overclocked (out of the box) AIB variants of the 2080Ti with higher power limits to compare to, as opposed to FE vs FE which seems to be the fairest comparison. Given that the 3070 seems heavily power limited with lots of OC potential on the VRAM, I'd expect to see some good OC results from the AIB variants too.

EDIT: Computerbase explicitly said they used the 2080Ti FE for their testing, and comes to the same conclusion as TechPowerUp - that for the most part the two cards are within +/- 1 or 2 percent of each other. I wonder why igorslab.de has the 3070 winning by a landslide?

A good example is Assassin's Creed:Odyssey at 1440p. TechPowerUp has the two cards dead even at 74.x fps. In the same title at the same resolution, IgorsLab has their 2080Ti sample slightly quicker at 76.7fps, but their 3070 results are 87.4fps - basically an entire product tier higher, and halfway to the 3080!

 

Very impressive, the 20TF 3070 actually seems to perform as one. The OC'd 2080Ti is actually avg around 16TF.

 

I'm not sure how that makes sense. By your statement the 3070 should be 40% faster or more than a 2080ti, which it's not and no one was expecting it to be.

The 3070 seems to be exactly where Nvidia touted it would be, which is great. And best of all the price is $700 less than the previously anchored $1200.

 

So far, the only thing that I can think of that makes sense: IgorsLab specifically says they're using an X570 motherboard and a Ryzen 3900XT, vs TechPowerUp's system is Intel. I wonder if PCIe 4.0 is the difference maker?

And when I said an entire product tier faster in my earlier post, I meant it! The performance uplift in AC:O at 1440p in IgorsLab's test between the 3070 and the 2080Ti is about the same as between the 2060 Super and 2080 Super! It's not the only title to show this either. Somewhat more impressive are the minimum frame times, which are a huge improvement from 2080Ti to 3070. Far Cry New Dawn shows this well, comparing minimum frame times between 2060 Super and 2080 Super or even 2080Ti and then again between 2080Ti and 3070 shows an enormous difference.

 

If the 3070 was that much faster than the 2080 Ti Nvidia would’ve said so. Igor’s numbers are probably wrong.

 

Maybe Igor accidently benched this thing on Zen3 lol.

 

The thing is, the 2080Ti is a well matured product by now. The 3070 is going to climb more and more above the 2080Ti as time moves on. It performs very close to what i would expect from a 20TF product.
Then i have no idea how relevant older games are, how does this perfom in next generation games?

What we should not forget is RT and DLSS performance either, i think those are atleast or even more important going forward. Throw in some competitive AMD products and we are in a good position.

 

Hardware Unboxed:

https://www.techspot.com/review/2124-geforce-rtx-3070/

shows a large gain for 3070 over 2080 Ti in Assassin's Creed Odyssey, but only at 4K: 50/37 versus 46/32 = 9/16% (average/1% low). Benched on Ryzen 3950X.

Igor isn't testing at "maximum" settings. Why even bother?

In my opinion the extra FLOPs and fillrate, despite the bandwidth deficit, are working very nicely in GA104.

 

~25% better perf/W than Turing; with a full-node jump. But yea, still better than those overstretched GA102s.

https://www.computerbase.de/2020-10...st/3/#diagramm-performance-pro-watt-1920-1080

 

Yes they were not lying. Was abit sceptical too in the beginning but then saw 100% perf inceases already right of the bat.

 

Once again memory overclocks don’t do nearly as much to improve performance as core overclocks. Nvidia GPUs don’t require much bandwidth at all even with the huge amt of FP32 cores. Wonder what utilization looks like and where in the chip the bottlenecks are most frequently occurring.

 

From the Q3 earnings call today:-

Hans Mosesmann

Congrats guys good stuff here. A question on capacity, Lisa you mentioned that it got better here in the back half of 2020. But as you look at 2021, 7-nanometer and I assume 5-nanometers specifically towards the end, I suppose? How is that capacity looking like and then I have a follow-up? Thanks.

Lisa Su

Yes thanks, Hans. So, look our second half has certainly been very strong and it was stronger than we originally planned. And so, we've worked closely with our suppliers to improve the supply availability. And I would say that even with that demand still exceeds supply in certain segments.

As we go into 2021, I think we are planning for success. And so, we're working very closely across the supply chain to ensure that we have enough wafer capacity, as well as back-end capacity. And we're going to continue to work on that, but certainly there are areas where we would like to supply to be higher and we're working on that.

With the caveat that it has less VRAM than the 2080Ti though. It should be fine in most current games, but time will tell how it will hold up in a year or two.

 

It's high time review sites learn to use profiling tools that give more insight than raw fps. It would be great to know how the 3070 fares vs 2080 Ti in specific parts of the frame.

 

Damn I miss hardware.fr for "advanced" reviews

 

And Anandtech's Beyond3D Suite of synthetics.

 

Are you able to attach a profiler to most games? I haven't tried. I would have thought the series profiling tools would require an executable compiled with debug code. Or is it as simple as turning on performance counters in drivers, or using developer drivers?

 

Dont forget that nVidia increased L1 cache by 33%, doubled L1 bandwidth within a SM and doubled triangle intersection within the RT Cores. So they did a few changes with Ampere which seems to help GA104 more than GA102.

 

Yes, I agree. People were criticizing the architecture as not having sufficient gains over Turing because the power consumption on the 3080 is so high. But now we can see because of smart improvements they've managed to equal a 2080ti with a lower power part that's on paper more limited in many ways.