Question 'Ampere'/Next-gen gaming uarch speculation thread

[Ottonomous]

How much is the Samsung 7nm EUV process expected to provide in terms of gains?
How will the RTX components be scaled/developed?
Any major architectural enhancements expected?
Will VRAM be bumped to 16/12/12 for the top three?
Will there be further fragmentation in the lineup? (Keeping turing at cheaper prices, while offering 'beefed up RTX' options at the top?)
Will the top card be capable of >4K60, at least 90?
Would Nvidia ever consider an HBM implementation in the gaming lineup?
Will Nvidia introduce new proprietary technologies again?

Sorry if imprudent/uncalled for, just interested in the forum member's thoughts.

 

[MrTeal]

And if it is like 1080ti, that is still serving a ton of people, it will be even better deal. If anything i expect future games to support DLSS upscaling to help extend lifetime at the cost of some image quality.

The halo card is never about value even if you keep it for awhile. I'll wait on reviews, but if the 3090 is 20-25% faster for more than twice the money it's going to be a small subset that considers that value vs buying a 3080 now and 4080/3080 Ti when they come out.

 

[Saylick]

Unless you absolutely need an RTX 30 series card today, I think it's prudent to wait for AMD to drop their Big Navi line-up and hopefully Nvidia will drop a RTX 3080 or possibly 3080 Ti with 20 GB to counter. MSRP $899? I just can't imagine any serious PC enthusiast would be willing to settle for 10 GB of VRAM on the 3080 as it will make you regret your purchase within 6 months.

 

[MrTeal]

For a minute I thought you wrote release the Karen’s!

$1499! We need to talk with Nvidia's manager.

 

[Hitman928]

1,9x its measured at 60fps, at top performance the perf/watt is 37% higher.

How do you get the 37% higher perf/watt? Comparing the 3080 to the 2080 I see roughly 80% higher performance across RTX and non-RTX titles for a 48.8% increase in TDP. That would make it 1.8/1.488 = 1.21 or 21% higher perf/watt.

Edit: This is obviously based on TDP listed for each card and not measured game power use which we don't have at this time.

 

[blckgrffn]

Unless you absolutely need an RTX 30 series card today, I think it's prudent to wait for AMD to drop their Big Navi line-up and hopefully Nvidia will drop a RTX 3080 or possibly 3080 Ti with 20 GB to counter. MSRP $899? I just can't imagine any serious PC enthusiast would be willing to settle for 10 GB of VRAM on the 3080 as it will make you regret your purchase within 6 months.

It will be interesting to see how far down the stack AMD pushes more than 8GB of ram. Are we going to see it at the $399 spot?

Whatever the performance numbers are, for the armchair enthusiasts a $400 AMD card having 16GB of ram vs everything less than $1,500 on the nvidia front having "last gen" ram quantities is going to be some poor optics.

How many 1080ti owners are going to see that buffer downgrade and take a pause on the 3080? I would, but I already sold my 1080 ti.

Can't wait for the DLSS vs Frame Buffer debates to ensue for people to hypothesize how soon each purchase will be regrettable 😂

I am certain people equate bigger is better with vram out there and buy accordingly. How many large vram cards with weak GPUs have been built for suckers in the past? I am sure we have all seen them.

 

[AtenRa]

How do you get the 37% higher perf/watt? Comparing the 3080 to the 2080 I see roughly 80% higher performance across RTX and non-RTX titles for a 48.8% increase in TDP. That would make it 1.8/1.488 = 1.21 or 21% higher perf/watt.

Edit: This is obviously based on TDP listed for each card and not measured game power use which we don't have at this time.

RTX3080 110fps / 320W

RTX2080 60fps / 240W

 

[Saylick]

It will be interesting to see how far down the stack AMD pushes more than 8GB of ram. Are we going to see it at the $399 spot?

Whatever the performance numbers are, for the armchair enthusiasts a $400 AMD card having 16GB of ram vs everything less than $1,500 on the nvidia front having "last gen" ram quantities is going to be some poor optics.

How many 1080ti owners are going to see that buffer downgrade and take a pause on the 3080? I would, but I already sold my 1080 ti.

Can't wait for the DLSS vs Frame Buffer debates to ensue for people to hypothesize how soon each purchase will be regrettable 😂

I am certain people equate bigger is better with vram out there and buy accordingly. How many large vram cards with weak GPUs have been built for suckers in the past? I am sure we have all seen them.

Given how next-gen consoles will have 10 GB+ of VRAM with 10+ TFLOPS of compute, I would hope that any discrete GPU with at least 10 TFLOPS of compute has at least that much VRAM, if not more. I personally think 16 GB is the sweet-spot for high-end discrete GPUs targeting 4K resolution. 8 GB is too little and 24 GB seems excessive.

 

[Hitman928]

RTX3080 110fps / 320W

RTX2080 60fps / 240W

First, that's a single RTX game. Second, the RTX2080 has a TDP of 215W, so you're not comparing 3080 to 2080 there, more likely the 2080 Super at 250W. Third, that does not look like 110 fps to me, more like 105 fps for Ampere. Lastly, If you look at Digital Foundry's video across multiple games comparing 2080 to 3080, it tells a different story.

Edit: even if we are very generous to the 3080 and say it is 90% faster than the 2080 which I doubt is the case across an average of many games, you're looking at 27.65% higher performance/watt.

 

[GoodRevrnd]

Are Tensor cores used for anything other than DLSS? And if not, doesn't DLSS barely tax them? What is the point, just design consistency from the A series?

Is 10GB really that insufficient for these cards?

 

[Asterox]

How do you get the 37% higher perf/watt? Comparing the 3080 to the 2080 I see roughly 80% higher performance across RTX and non-RTX titles for a 48.8% increase in TDP. That would make it 1.8/1.488 = 1.21 or 21% higher perf/watt.

Edit: This is obviously based on TDP listed for each card and not measured game power use which we don't have at this time.

Or simple, do we need to calculate Ampere IPC increase?

- RTX 2080 TI, 4352 Cuda cores, 1500mhz

- RTX 3080, 8704 Cuda Cores, 1700mhz

 

[Konan]

Given how next-gen consoles will have 10 GB+ of VRAM with 10+ TFLOPS of compute, I would hope that any discrete GPU with at least 10 TFLOPS of compute has at least that much VRAM, if not more. I personally think 16 GB is the sweet-spot for high-end discrete GPUs targeting 4K resolution. 8 GB is too little and 24 GB seems excessive.

G6X memory > GDDR6 though??

 

[GoodRevrnd]

Or simple, do we need to calculate Ampere IPC increase?

- RTX 2080 TI, 4352 Cuda cores, 1500mhz

- RTX 3080, 8704 Cuda Cores, 1700mhz

Aren't those potentially apples/oranges cuda cores though?

 

[jpiniero]

As one of the threads mentioned, it's possible that the integer core that was part of the SM can now do both int and fp32.

eg: here's the Turing block diagram:

 

[Hitman928]

Aren't those potentially apples/oranges cuda cores though?

Yes. If you actually want to do that comparison that way then Ampere is a severe reduction in "IPC".

3080/2080 performance = 1.8
3080/2080 cuda cores * freq = 3.32
3080/2080 "IPC" = 0.54 or in other words a 46% reduction in IPC for Ampere.

Clearly this is not the proper comparison.

 

[blckgrffn]

G6X memory > GDDR6 though??

Won't help if you are running out of memory. Running out of vram is like hitting wall historically.

We'll see if more powerful, lower latency storage subsystems can help mitigate that, but it's still easier just to have the right amount of memory for what you are doing.

 

[Hitman928]

As one of the threads mentioned, it's possible that the integer core that was part of the SM can now do both int and fp32.

eg: here's the Turing block diagram:

Turing could already do both int & fp32. The theory is that now Ampere can do int & fp32 OR 2xfp32. This won't be 2x performance in games though as you aren't doubling the rest of the pipeline to feed 2xfp32 and games don't do just fp32 calculations, they need a significant amount of int calculations as well.

 

[Saylick]

Are Tensor cores used for anything other than DLSS? And if not, doesn't DLSS barely tax them? What is the point, just design consistency from the A series?

Is 10GB really that insufficient for these cards?

Tensor cores are used for denoising for the ray-tracing and also for AI-based upscaling. In other words, you can run RTX at native 4k without upscaling, as an example, and it will use the tensor cores, or you can just run DLSS without ray-tracing and it will also use the tensor cores, and a game with RTX and DLSS will both tax them.

 

[MrTeal]

As one of the threads mentioned, it's possible that the integer core that was part of the SM can now do both int and fp32.

eg: here's the Turing block diagram:

Nvidia itself is quoting the 36TFlop FP32 number, so that would seem to be the case. I'm actually quite surprised, but that is a lot of compute horsepower there if there's no gotchas. 29.8TF FP32 for $700 is a crazy value compared to the 13.4TF @ $1200 for the 2080 Ti or 13.8TF @ $700 for the Radeon VII.

 

[DXDiag]

All RTX Turing and Ampere GPUs will be able to accelerate decompression data from SSDs in accordance with Microsoft DirectStorage API. This will put PC on equal footings with consoles.

NVIDIA Developer

developer.nvidia.com

 

[Glo.]

Nvidia itself is quoting the 36TFlop FP32 number, so that would seem to be the case. I'm actually quite surprised, but that is a lot of compute horsepower there if there's no gotchas. 29.8TF FP32 for $700 is a crazy value compared to the 13.4TF @ $1200 for the 2080 Ti or 13.8TF @ $700 for the Radeon VII.

They quote SHADER Flops.

Not FP32 . Its similar to what they claimed that FP16 performs "the same", enhanced by all that GEMM stuff, in their A100 chip to native FP32.

In reality native FP32 will be exactly that, native FP32.

 

[AtenRa]

First, that's a single RTX game. Second, the RTX2080 has a TDP of 215W, so you're not comparing 3080 to 2080 there, more likely the 2080 Super at 250W. Third, that does not look like 110 fps to me, more like 105 fps for Ampere. Lastly, If you look at Digital Foundry's video across multiple games comparing 2080 to 3080, it tells a different story.

Edit: even if we are very generous to the 3080 and say it is 90% faster than the 2080 which I doubt is the case across an average of many games, you're looking at 27.65% higher performance/watt.

Its the slide they mentioned 1.9x perf/watt and according to official NVIDIA slide,

Turing 60fps / 240w = 0.25
vs
Ampere 105fps / 320W = 0.33

This is 32% higher perf/watt , again this is according to official NVIDIA slide.

 

[MrTeal]

They quote SHADER Flops.

Not FP32 . Its similar to what they claimed that FP16 performs "the same", enhanced by all that GEMM stuff, in their A100 chip to native FP32.

In reality native FP32 will be exactly that, native FP32.

Maybe? The nvidia website itself says 2x FP32 in the SM, though they don't directly list the TF number there.


Edit: You're correct though and I was wrong. They don't directly say 36TFlops, they just say 10496/8704 CUDA cores and 2x FP32 throughput

 

[lobz]

3080 looks good.Probably best card in whole stack.70-80% faster than 2080 non super at same price.Thanks to transition from middle 104die to BIG 102die.Also more vram and 320bit.
3070 is not that good.Only 50-60% faster than 2070 non super and still only 8Gb Vram and 256bit.It is still middle 104die.Should be priced at 399.
3080 actually have better price/performance than 3070 this time.
3080 have +50% more SP, 320bit and more Vram and it cost only 40% more.

Don't forget guys, that Turing's horrible pricing is one of the reasons why the 2080 - 3080 'for the same price' sounds that good. The 2080 was on par with the 1080ti that released for $699. People said that's because of Ray Tracing, but all I saw was the same performance for roughly the same price with an added feature that's unusable in 99,99% of the time.

Let's put this into perspective: The 3080 is between 70% and 80% faster, and in some cases 100% faster than the GTX 1080ti that was released 3.5 years ago for the same price.
Now FINALLY we actually get Ray Tracing as a bonus this time, with a handful of worthy games coming with the feature this year.
We get all this for a TDP increase of 75W on a smaller process, and a 1GB decrease in RAM.

Sounds much less impressive. I've said all along that Turing as an improvement is a dud, compared to other NVIDIA launches.

 

[Hitman928]

Its the slide they mentioned 1.9x perf/watt and according to official NVIDIA slide,

Turing 60fps / 240w = 0.25
vs
Ampere 105fps / 320W = 0.33

This is 32% higher perf/watt , again this is according to official NVIDIA slide.

Again, that is in 1 game which is currently the most RTX intense game available, look at the fine print. It is also comparing a 3080 versus an essentially overclocked 2080 so that will also tilt the scale in favor of the 3080 in terms of perf/w.

If you watch the Digital Foundry video you get more like 20-25% perf/w improvement compared to Nvidia's media slide.

 

[Glo.]

Maybe? The nvidia website itself says 2x FP32 in the SM, though they don't directly list the TF number there.
View attachment 29056

Its 2xFP32 THROUGHPUT.

Its the same marketing gibberish they used for A100 chips.

In reality, when tested actual, native FP32 performance, it was the same, as V100.