RTX 3xxx performance speculation

[Snowdog]

Its going to be a beast even in a cut down consumer version: https://devblogs.nvidia.com/nvidia-ampere-architecture-in-depth/

There will be no consumer version of GA100. It doesn't even have RT cores. That is a pure data-center part.

 

[5150Joker]

There will be no consumer version of GA100. It doesn't even have RT cores. That is a pure data-center part.

Yes it doesn't have RT but you can still get an idea of how the SM will look since both will have the tensor cores, fp/int 32 and the transistor count with that die size is gigantic.




Are fp64 needed at all in consumer gaming parts? I bet consumer Ampere will have fp64 replaced with fp32 and RT cores tacked on but look otherwise very similar to this. But 826 mm2 with 54.2B transistors vs Tesla T4 at 545 mm2 and 13.6B transistors is a gigantic leap. Just using that metric alone is enough to figure out that Ampere is going to be a significant jump from Turing.

 

[Armenius]

Yes it doesn't have RT but you can still get an idea of how the SM will look since both will have the tensor cores, fp/int 32 and the transistor count with that die size is gigantic.

View attachment 245435
View attachment 245434

Are fp64 needed at all in consumer gaming parts? I bet consumer Ampere will have fp64 replaced with fp32 and RT cores tacked on but look otherwise very similar to this. But 826 mm2 with 54.2B transistors vs Tesla T4 at 545 mm2 and 13.6B transistors is a gigantic leap. Just using that metric alone is enough to figure out that Ampere is going to be a significant jump from Turing.

FP64 isn't needed in games, no. How did Volta compare to Turing? Was there any similarity between the two? Trying to remember because if NVIDIA is differentiating architectures between product lines than the forthcoming GeForce products could not even be based on or called Ampere.

 

[5150Joker]

FP64 isn't needed in games, no. How did Volta compare to Turing? Was there any similarity between the two? Trying to remember because if NVIDIA is differentiating architectures between product lines than the forthcoming GeForce products could not even be based on or called Ampere.

This is Volta:


So definitely different than Turing. The interesting thing is Marketwatch has Jensen saying the following:

Ampere will eventually replace Nvidia’s Turing and Volta chips with a single platform that streamlines Nvidia's GPU lineup, Huang said in a pre-briefing with media members Wednesday.

..

Nvidia did not release any information Wednesday about consumer GPUs using Ampere, but when asked by a reporter in the briefing about the difference between enterprise and consumer approaches to Ampere, Huang said “there’s great overlap in the architecture, but not in the configuration.”

 

[Snowdog]

So definitely different than Turing. The interesting thing is Marketwatch has Jensen saying the following:

So they share a name this time. I don't think we are seeing much that will be in the gaming cards. Same Tensor cores?

 

[5150Joker]

So they share a name this time. I don't think we are seeing much that will be in the gaming cards. Same Tensor cores?

If they gut the fp64 and replace it with fp32 and also tack on RT then it will technically be the same architecture but a different "configuration". I have a feeling NVIDIA is going to push DLSS 3.0 very hard this generation given how much die space tensor cores are taking up and the 2x performance improvement they claim in their whitepaper. They'll probably emphasize it as much as RT if not more.

A100 has four Tensor Cores per SM, which together deliver 1024 dense FP16/FP32 FMA operations per clock, a 2x increase in computation horsepower per SM compared to Volta and Turing.

 

[Dayaks]

Think about AMD's ray tracing video then watch the one from nVidia running on Turing at 9:50.



Ampere is going to absolutely crush AMD.

 

[5150Joker]

Think about AMD's ray tracing video then watch the one from nVidia running on Turing at 9:50.

Ampere is going to absolutely crush AMD.

What really impresses me is the transistor density NVIDIAs squeezed out with TSMC 7nm EUV at the top end. AMD will really have to do something magical with Navi 2x to match that and maintain reasonable power consumption. I'm still waiting to see how AMD does in RT, they might end up matching NVIDIA in that area but what will really separates the two is the tensor cores and dlss 3.0. AMD really needs a dlss equivalent this generation otherwise NVIDIA will use it's marketing muscle + dollars to make sure DLSS gets used in the latest AAA titles, especially Cyberpunk.

Also just think about this fact: With 826 mm2 die and 54 billion transistors, A100 is clocking to 1410 MHz in a datacenter configuration, if they even get out a top end consumer part with say 25-30 billion transistors for consumers on a smaller die, they'd probably be able to easily scale that boost clock well passed 2 GHz. I think Moores Law Is Dead was pretty accurate in his analysis and we're going to see something pretty great from the top end parts of Ampere. I'm still curious about what NVIDIA will do with the more mainstream Samsung manufactured gpus and how competitive they will be vs RDNA 2 which will be TSMC top down. Who knows, maybe Ampere will shine at the very high end but RDNA 2 might keep up in the low/med/mid high range because all their parts will be built at TSMC.

I don't think 3080 Ti is going to be cheap though, we're probably looking at $1300+ again this generation.

 

[sethk]

I believe TSMC has three different 'flavors' of '7nm' at various stages of maturity, and I assume average feature size (or density) and transistor cost is not the same on each. There were also big bets made on node allocation early on which forced Nvidia to choose from remaining availability, but the newer nodes may have more differences than just speed and power draw.

 

[imsirovic5]

If they gut the fp64 and replace it with fp32 and also tack on RT then it will technically be the same architecture but a different "configuration". I have a feeling NVIDIA is going to push DLSS 3.0 very hard this generation given how much die space tensor cores are taking up and the 2x performance improvement they claim in their whitepaper. They'll probably emphasize it as much as RT if not more.

Is there any info on DLSS 3.0 and how it may differ from 2.0? Or is it same as DLSS 2.0 but on steroids (for those that have more tensor cores / Ampere GPU)

 

[noko]

A very aggressive Nvidia is good for us. AMD can compete with price vs performance. Lisa Sue previous goal was to make an overall 45% profit margin which they were able to do.

I really like what Nvidia is doing with HPC, I don't think anyone will be close to them. Nvidia could be the next Apple. With there incredible AI performance, I can see some major potential shifts coming especially if Mvidia creates an AI driven OS that works with their hardware.

Ampere is looking to be the break away product, looking forward to their gaming cards and surprises coming with more AI driven software.

 

[Snowdog]

Is there any info on DLSS 3.0 and how it may differ from 2.0? Or is it same as DLSS 2.0 but on steroids (for those that have more tensor cores / Ampere GPU)

DLSS 3.0 "info" is just BS clickbait.

 

[imsirovic5]

DLSS 3.0 "info" is just BS clickbait.

That makes sense, I just see references to it and am not sure if I there is any big piece of information that I may have missed.

 

[5150Joker]

Is there any info on DLSS 3.0 and how it may differ from 2.0? Or is it same as DLSS 2.0 but on steroids (for those that have more tensor cores / Ampere GPU)

If dlss 3.0 is even a thing, it'll probably just be marketing to push Ampere. I used that term based on Moores laws dead video, there's no confirmation it's real (yet).

 

[melk]

Hope the 3060 is one of the first cards announced and not just the 70/80

 

[Snowdog]

Hope the 3060 is one of the first cards announced and not just the 70/80

This is almost never the case. They always start at the top and work there way down. It makes for a more orderly clearance of old stock.

 

[imsirovic5]

This is almost never the case. They always start at the top and work there way down. It makes for a more orderly clearance of old stock.

Do you think NVDA may start with 3080 Ti or do you see 3080 Ti as 2021 product? To me starting with 3080 Ti would make logical sense as they are very low volume product and can generate halo / excitement for lower end tiers as NVDA ramps up for mass production.

 

[Snowdog]

Do you think NVDA may start with 3080 Ti or do you see 3080 Ti as 2021 product? To me starting with 3080 Ti would make logical sense as they are very low volume product and can generate halo / excitement for lower end tiers as NVDA ramps up for mass production.

From my memory they always launch with the top part, whether it called x80Ti or Titan.

 

[Nebell]

Well, I hope it's not coming later than October.
Becausr it's been 3 weeks without a GPU (only integrated) and I'm getting slowly pissed off.
I have tons to do this summer so I can love without it, but when the weather turns to shit in Sweden (October), that's when I want VR working at home...

 

[Nenu]

Something makes me think there might not be a 3080ti, I'm not sure what it was I read.
It was along the lines of there is likely to be a Titan then 3080 down, to try and keep the pricing up.
Most likely rubbish but it made me uneasy.

 

[imsirovic5]

Something makes me think there might not be a 3080ti, I'm not sure what it was I read.
It was along the lines of there is likely to be a Titan then 3080 down, to try and keep the pricing up.
Most likely rubbish but it made me uneasy.

You could be correct, they could do Titan first, then mass production for lower end tiers then 3080 Ti sometimes in 2021 which would really suck for all of us stuck at 1080 Ti and waiting for a worthy successor. I do not think I will go for Titan as the price / performance ratio for those (in gaming) is beyond ridiculous.

Although if there ever was a time I would be tempted to pay such a huge premium and go with Titan, this would be it (given upcoming Zen 3, LG OLED CX 48" 4K 2.1 HDMI "monitor", Cyberpunk 2077 etc.)

 

[Snowdog]

You could be correct, they could do Titan first, then mass production for lower end tiers then 3080 Ti sometimes in 2021 which would really suck for all of us stuck at 1080 Ti and waiting for a worthy successor. I do not think I will go for Titan as the price / performance ratio for those (in gaming) is beyond ridiculous.

GTX 1000 Launched with Titan instead of 1080ti.
RTX 2000 Launched with 2080ti instead of Titan, but it was priced like Pascal Titan.

So really launching with 3080ti or Titan is more like a naming/marketing decision than a pricing difference. Though calling a 2080ti instead of Titan did make a lot of people completely lose it.

To me the naming is irrelevant. They will launch with their Biggest die Ampere gaming part (GA 102?), and it will be expensive (> $1000) regardless of whether they call lit Titan or 3080Ti.

 

[imsirovic5]

GTX 1000 Launched with Titan instead of 1080ti.
RTX 2000 Launched with 2080ti instead of Titan, but it was priced like Pascal Titan.

So really launching with 3080ti or Titan is more like a naming/marketing decision than a pricing difference. Though calling a 2080ti instead of Titan did make a lot of people completely lose it.

To me the naming is irrelevant. They will launch with their Biggest die Ampere gaming part (GA 102?), and it will be expensive (> $1000) regardless of whether they call lit Titan or 3080Ti.

True, and agree, but there is still a huge price difference between Titan and "Ti" - for example Titan RTX is $2,500 MSRP while RTX 2080 Ti MSRP was $1,200 - so Titan is 2x the cost for negligible gain (in gaming). To me this is still a significant difference.

 

[exlink]

I fully anticipate the RTX 3000 series to follow a similar launch and pricing structure as the RTX 2000 series. We'll see the RTX 3070, 3080, and 3080 Ti this fall and then probably the 3060 in the winter or really early spring 2021. But this also depends on just how competitive AMD is this time around. May push Nvidia to release the 3060 cards earlier if necessary.

 

[n370zed]

Here’s hoping my $600 max I’m willing to spend on Ampere gets me an RTX **70 part at launch. Oh gawd I hope so. I can’t imagine it being more.....

 

[Dayaks]

A very aggressive Nvidia is good for us. AMD can compete with price vs performance. Lisa Sue previous goal was to make an overall 45% profit margin which they were able to do.

I really like what Nvidia is doing with HPC, I don't think anyone will be close to them. Nvidia could be the next Apple. With there incredible AI performance, I can see some major potential shifts coming especially if Mvidia creates an AI driven OS that works with their hardware.

Ampere is looking to be the break away product, looking forward to their gaming cards and surprises coming with more AI driven software.

I am pretty sure you meant 45% gross margin, which is healthy. nVidia is at 60% gross margin / 20% profit iirc.

Low overhead businesses (not tech) can run around 30-40% gross and be healthy.

 

[5150Joker]

I am pretty sure you meant 45% gross margin, which is healthy. nVidia is at 60% gross margin / 20% profit iirc.

Low overhead businesses (not tech) can run around 30-40% gross and be healthy.

It's the same thing. They calculate gross margins as a way to track profit % made and that helps with efficiency for a company like nVidia or AMD. It doesn't count rent, administrative fees etc but strictly speaking if their gross margin is 60% then that's the profit % nVidia makes which is really good.

 

[Dayaks]

It's the same thing. They calculate gross margins as a way to track profit % made and that helps with efficiency for a company like nVidia or AMD. It doesn't count rent, administrative fees etc but strictly speaking if their gross margin is 60% then that's the profit % nVidia makes which is really good.

He said 45% which refers to gross margin (less used gross profit). The profit margin is what actually shows the company is profitable.... you can be at 40% gross margin (gross profit) and still lose money (negative profit margin) as a company.

I only point it out because people get riled up over nVidia but they are 20% profit margin (which is still high but not like old Pharma high).

 

[5150Joker]

He said 45% which refers to gross margin (less used gross profit). The profit margin is what actually shows the company is profitable.... you can be at 40% gross margin (gross profit) and still lose money (negative profit margin) as a company.

I only point it out because people get riled up over nVidia but they are 20% profit margin (which is still high but not like old Pharma high).

Right but if you're only looking at gpus their gross profit margin is essentially the profit % they are making, they can be used interchangeably. As a company their overall profit is lower from other overhead. I think a lot of people fail to realize how expensive it is to produce cutting edge gpus because they only account for raw production cost and neglect the years of R&D cost behind it in addition to operating expenses all those years, marketing, driver development etc. I'm surprised we get chips with billions of transistors for as "cheap" as we do.

 

[strufa]

Do you all think the 3080ti will be vastly superior to amd's big navi card? Since its rumored to be at 862 mm^2 vs 502 mm^2 i think nvidia has a big advantage here.

 

[Snowdog]

Do you all think the 3080ti will be vastly superior to amd's big navi card? Since its rumored to be at 862 mm^2 vs 502 mm^2 i think nvidia has a big advantage here.

The giant Ampere is the Data Center part, not the gaming part. We really have no idea how big the Gaming Ampere part will be.

 

[noko]

Do you all think the 3080ti will be vastly superior to amd's big navi card? Since its rumored to be at 862 mm^2 vs 502 mm^2 i think nvidia has a big advantage here.

rumoured, raining salt, tasting like salt. Most of us do not know one way or another. The other factor is price per performance for what one will use it for. Just have to wait for the dust to clear to decide which is better for one own purposes. Both could be superior in some aspect or feature or one could be so worth the cost that even if the other is somewhat better it is not worth it. If Nvidia can capitalize more on their AI lead, useful functions that many would want, like DLSS, be implemented easily and used, then the added extra cost may will be worth it. I believe Nvidia is at a breakaway point, a clear break away where every competitor may not even be in the rearview mirror so to speak. Won't know until the stuff hits the ground.

 

[strufa]

Do you mind elaborating how nvidia would be at a breakaway point from your point of view? Yes they have an AI lead but does nvidia have other up and coming technologies that would push their lead performance wise in the gaming market?

 

[Dayaks]

Do you mind elaborating how nvidia would be at a breakaway point from your point of view? Yes they have an AI lead but does nvidia have other up and coming technologies that would push their lead performance wise in the gaming market?

NVENC for real time encoding, better performance per transistor, and if they can get better adoption - Ray Tracing and DLSS.

And usually better software and drivers.

 

[XoR_]

The giant Ampere is the Data Center part, not the gaming part. We really have no idea how big the Gaming Ampere part will be.

They might release it as normal PCI-E card which you could then use for gaming.
Price would be at least $3000 so not really cheap but it can and imho should happen.

 

[Snowdog]

They might release it as normal PCI-E card which you could then use for gaming.
Price would be at least $3000 so not really cheap but it can and imho should happen.

It doesn't even contain RTX cores, so no real chance of that.

 

[pippenainteasy]

It doesn't even contain RTX cores, so no real chance of that.

It would be like buying the Titan V CEO edition for gaming even if you could. No point dropping Gs just to brag you have the most capable DX11 card ever.

 

[XoR_]

It doesn't even contain RTX cores, so no real chance of that.

Too bad. It now seems a lot less probable they will do it.

 

[madpistol]

Just want HDMI 2.1. Seriously. I'll be happy for 30% performance increase, but I REALLY want HDMI 2.1.

 

[melk]

Man dropping back down to my old 1060 after using my 2060 Super was a rude awakening. Crazy how big a jump that was. Cant wait for Hot Chips conference in August!