News Nvidia RTX 3080 and Ampere: Everything We Know

Zizo007

Respectable
Feb 23, 2019
1,655
229
2,290
55
How is it possible that the 3080Ti has double the GFLOPS of 3080? The 2080Ti is not even close to double the speed of 2080, its only 30% faster in games and has 10.07GFLOPS.

Also the 2080Ti is 14.2GFLOPS, I don't believe that the 3080Ti is more than double at 28.7GFLOPS at 102%.

Even if game performance is lower than GFLOPS this is still unbelievable.
 
Last edited:

JarredWaltonGPU

Senior GPU Editor
Editor
Feb 21, 2020
228
186
260
0
How is it possible that the 3080Ti has double the GFLOPS of 3080? The 2080Ti is not even close to double the speed of 2080, its only 30% faster in games and has 10.07GFLOPS.

Also the 2080Ti is 14.2GFLOPS, I don't believe that the 3080Ti is more than double at 28.7GFLOPS at 102%.

Even if game performance is lower than GFLOPS this is still unbelievable.
It all depends on the configuration of the chips -- and note the use of "up to" and less than signs. There are at least plausible leaks of a 124 SM Ampere chip. As I note in the text, that's not a "good" number for the maximum configuration, because GPCs usually all have the same number of SMs. So 16 GPCs with 8 SMs each would work, or 12 GPCs with 10 SMs each (which is fine for the 118 SM leak but doesn't cover the 124 SM leak).

What will the chips actually have enabled, though? My guess is around 112 SMs on an RTX 3080 Ti would make sense. If we use that figure, and the same clockspeed, TFLOPS drops to 25.1. So, that's part one of the story.

The second part is the rumored GA103. I put down 64 SMs as the top configuration, but 72 or even 80 SMs is entirely possible, as noted in the text. So I probably undershot on the GA103 estimate and overshot on the GA100 estimate -- or RTX 3080 and RTX 3080 Ti, if you prefer. I also gave GA103 a higher potential clockspeed advantage. If RTX 3080 ends up with somewhere between 72-80 SMs instead of 64 SMs, performance ends up at 18.4-20.5 TFLOPS.

Realistically, then, RTX 3080 Ti will probably be in the 20-25 TFLOPS range (for FP32). RTX 3080 will end up closer to 14-16 TFLOPS. Or Nvidia does something completely unexpected and the theoretical TFLOPS values end up much lower but real-world performance (ie, the utilization of the available compute) goes up.

Actually, having thought on it a bit more, I think GA103 will probably have more than 64 SMs available. RTX 3080 could still go with 64 SMs enabled, but if Nvidia can do a 128 SM chip, the second tier will probably be much higher than 64. I've adjusted the tables and text slightly to reflect this.
 

Zizo007

Respectable
Feb 23, 2019
1,655
229
2,290
55
It all depends on the configuration of the chips -- and note the use of "up to" and less than signs. There are at least plausible leaks of a 124 SM Ampere chip. As I note in the text, that's not a "good" number for the maximum configuration, because GPCs usually all have the same number of SMs. So 16 GPCs with 8 SMs each would work, or 12 GPCs with 10 SMs each (which is fine for the 118 SM leak but doesn't cover the 124 SM leak).

What will the chips actually have enabled, though? My guess is around 112 SMs on an RTX 3080 Ti would make sense. If we use that figure, and the same clockspeed, TFLOPS drops to 25.1. So, that's part one of the story.

The second part is the rumored GA103. I put down 64 SMs as the top configuration, but 72 or even 80 SMs is entirely possible, as noted in the text. So I probably undershot on the GA103 estimate and overshot on the GA100 estimate -- or RTX 3080 and RTX 3080 Ti, if you prefer. I also gave GA103 a higher potential clockspeed advantage. If RTX 3080 ends up with somewhere between 72-80 SMs instead of 64 SMs, performance ends up at 18.4-20.5 TFLOPS.

Realistically, then, RTX 3080 Ti will probably be in the 20-25 TFLOPS range (for FP32). RTX 3080 will end up closer to 14-16 TFLOPS. Or Nvidia does something completely unexpected and the theoretical TFLOPS values end up much lower but real-world performance (ie, the utilization of the available compute) goes up.

Actually, having thought on it a bit more, I think GA103 will probably have more than 64 SMs available. RTX 3080 could still go with 64 SMs enabled, but if Nvidia can do a 128 SM chip, the second tier will probably be much higher than 64. I've adjusted the tables and text slightly to reflect this.
Thanks for the elaborate answer, that makes more sense now.
 
Reactions: JarredWaltonGPU
"As with Big Navi, the best advice right now is to wait and see what actually materializes."
Took a while to get there but I agree. Way to much rumor mill stuff going on about graphics cards and very little seems likely to be factual. I'll just wait until AMD and Nvidia both release their next cards.
 
Reactions: bit_user

2Be_or_Not2Be

Honorable
Aug 23, 2013
843
12
11,365
115
"As with Big Navi, the best advice right now is to wait and see what actually materializes."
Took a while to get there but I agree. Way to much rumor mill stuff going on about graphics cards and very little seems likely to be factual. I'll just wait until AMD and Nvidia both release their next cards.
What else are people doing now that most are staying at home? You need articles to fill the "new" queue for the greater idle time people have now. Speculative articles are easy choices to keep something "new" in the queue.
 

JarredWaltonGPU

Senior GPU Editor
Editor
Feb 21, 2020
228
186
260
0
What else are people doing now that most are staying at home? You need articles to fill the "new" queue for the greater idle time people have now. Speculative articles are easy choices to keep something "new" in the queue.
Plus we didn't actually have a central hub for all the Ampere (and Big Navi) rumors, which is a good thing to put together for any website that covers GPUs. There's not a lot of concrete information right now, but we'll keep this updated as new stuff comes out, and eventually it will morph into our Everything We Know article with final architectural details when the parts get announced.
 
Nov 27, 2019
39
6
35
0
An Arm and a leg wont cut it. With a 2080 Ti easily costing 60% to 80% AUD more than the previous generations, I expect this will be at least 100% - 120% more than the GTX 1080 Ti. About 2 months rent ought to cover it. Actually it could be more as the AUD is plummeting to 50c USD, so that might be A150% or more. I don't think many Australians will get to run one of these new cards.
 

Zizo007

Respectable
Feb 23, 2019
1,655
229
2,290
55
An Arm and a leg wont cut it. With a 2080 Ti easily costing 60% to 80% AUD more than the previous generations, I expect this will be at least 100% - 120% more than the GTX 1080 Ti. About 2 months rent ought to cover it. Actually it could be more as the AUD is plummeting to 50c USD, so that might be A150% or more. I don't think many Australians will get to run one of these new cards.
Same here with Canadian dollar, my Asus 2080 Ti cost 2200$ with taxes now but I got it used 6 months ago for 1500$CA from ebay.

I wouldn"t be surprised if the 3080Ti cost 3000+$CA here with the corona and lack of supply going on.
 
Last edited:

JarredWaltonGPU

Senior GPU Editor
Editor
Feb 21, 2020
228
186
260
0
Ouch, the prices for other countries are sounding super painful these days. I really do hope AMD knocks one out of the park with Big Navi so that things are more competitive. That will help bring prices down as well. But 7nm is costly, so I'm not counting on prices dropping next gen. :-(
 

wr3zzz

Distinguished
Dec 31, 2007
2
1
18,515
0
Shouldn't 8-16GB be the new base line for VRAM? It's not just a technical need but also for marketing reason. Nvidia used 4-8GB memory for the last console gen and bump the baseline to 6GB when XboxPro came out with 12GB. given the XboxSeX and PS5 specs I would think the 3060-3070-3080 spec would be 8GB-12GB-16GB. It doesn't look good for the top 3080 to have less memory than a console.

Nvidia probably will use the memory hike to justify a big price hike. It used to be that we pay the same price every gen for more performance but it's crazy that for two gens now we effectively have to pay linearly for more.
 
Last edited:
Reactions: 2Be_or_Not2Be

spongiemaster

Proper
Dec 12, 2019
199
74
160
0
Shouldn't 8-16GB be the new base line for VRAM? It's not just a technical need but also for marketing reason. Nvidia used 4-8GB memory for the last console gen and bump the baseline to 6GB when XboxPro came out with 12GB. given the XboxSeX and PS5 specs I would think the 3060-3070-3080 spec would be 8GB-12GB-16GB. It doesn't look good for the top 3080 to have less memory than a console.
Consoles use one pool of RAM for system and video memory, so the GPU will never be able to use all of the RAM on the spec sheet.
 

JarredWaltonGPU

Senior GPU Editor
Editor
Feb 21, 2020
228
186
260
0
Depends whether or not 3xxx series comes with HDMI 2.1 - If not that will be a dealbreaker for a lot of people
Considering Turing at least has partial HDMI 2.1 support (it can do VRR with HDMI 2.0), I strongly anticipate HDMI 2.1 in Ampere. Nothing is certain of course, but I'd be very surprised if HDMI 2.1 and DP 2.0 aren't included.
 

cerealkeller

Distinguished
Jul 30, 2009
22
3
18,520
1
I think Ampere will probably be HDMI 2.1 and PCI-e 4.0. My LG C9 OLED is HDMI 2.1. So there's already hardware out there waiting for compatibility. Nvidia is keeping up with the latest standards, my TV supports G Sync with an RTX series graphics card via HDMI 2.0, it's pretty awesome. HDMI 2.1 is going to be even better for me, that will unlock 120 hz at 4K on my TV over HDMI.
I would like to see an bump up in VRAM. I'm sure some people think it's excessive, but with SLI 2080 Ti's running at very high resolution I have run out of VRAM on a number of occasions, and that's with 11 GB, I really ran into a wall at 7K and 8K. Not that I plan to buy two 3080 Ti's if they cost $1500. At least selling my 2080 Ti's should pay for one, maybe I will buy two :ROFLMAO: I just hope there isn't a repeat of people's cards dying like there was with the RTX launch. I preordered mine, an Nvidia FE card. But it has been going strong for about a year and a half at this point, with no trouble whatsoever. Of course my other card, which I bought a few months later, over clocks much better.
Another issue for me is that I will have to buy a water block for each, adding to the already absurd expense. Selling my 2080 Ti's, I dread the thought of selling the blocks separately. Reassembling an FE, with its 50+ screws, is a Goddamn nightmare.
 

ASK THE COMMUNITY

TRENDING THREADS