• Light/Dark
  • Ever wanted an RSS feed of all your favorite gaming news sites? Go check out our new Gaming Headlines feed! Read more about it here.

If the 3070 is on par with the 2080Ti why the big Teraflop difference?

Flandy

Flandy

Community Resettler
Member
Oct 25, 2017
3,445
2080Ti is 14 TFLOPs give or take
3070 is 20 TFLOPs
Nvidia said the 3070 is only slightly more powerful than the 2080Ti. Why the huge increase in TFLOPs? It's making me take the 30 and 36 numbers for the 3080/3090 with a much bigger grain of salt
NVIDIA_Gamescom_2018_GeForce_RTX_20_Series_Launch_7.png

Eg1-sIpVgAEEZnX.jpg
 
Linus815

Linus815

Member
Oct 29, 2017
19,720
because you cant cmpare teraflops between architectures, and there is several other factors outside of it as well.
 
Komo

Komo

Info Analyst
Verified
Jan 3, 2019
7,110
Cause the architecture has changed and teraflops don't mean everything lol
 

kabel

Member
Feb 3, 2020
55
FLOPS = Floating Point Operations Per Second

Ampere has doubled the FP32 throughput. But games also need integer operations.
 
djplaeskool

djplaeskool

Member
Oct 26, 2017
19,736
There are so many differing TFLOP calculations and no definitive standard, so it's kind of a BS metric for comparing unlike platforms.
 
Stickman

Stickman

Member
Oct 27, 2017
381
Teraflops should only be compared within the same GPU-architecture. They are not what horsepower is for cars.
 
LordRuyn

LordRuyn

Member
Oct 29, 2017
3,909
Flandy said:
This was my takeaway when comparing them. Keep seeing people mentioning 30TF for the 3080 though like it means something so I'm left confused ¯\_(ツ)_/¯
Click to expand...
Click to shrink...
TFLOPs are not the be-all end-all in terms of performance measurements, especially with two different GPU architectures.
 
R.T Straker

R.T Straker

Chicken Chaser
Member
Oct 25, 2017
4,715
I find it funny that the console manufacturers pretty much made some to think that tflops is all that matters in a GPU with their marketing.
 
theAllseeingEYE

theAllseeingEYE

Member
Apr 4, 2018
4,509
Vancouver, BC
Hold you horses OP, we haven't see actual performance data from the 3070 yet. Jensen said the 3070 was more powerful, not slightly more powerful, and said the same thing for the 3080. In reality, based on the Digital Foundry performance numbers, the 3080 is actually significantly more powerful than the 2080ti.

For example, a chart I've been working on:


Borderlands 3 (4k/Bad Ass/DX12/4K/TAA)
RTX 2080 - 34.33fps
RTX 2080 ti - 48FPS
RTX 3080 - 61.79fps (180%)


Shadow of the Tomb Raider (4k/Highest/TAA)
RTX 2080 - 47fps
RTX 2080 ti - 61FPS
RTX 3080 - 78.96fps (168%)


Battlefield V (4k/Ultra/TAA/DXR OFf)
RTX 2080 - 67fps
RTX 2080 ti - 84FPS
RTX 3080 - 112fps (168%)


Control (4k/High/TAA/DLSS OFf)
RTX 2080 - 31.8fps
RTX 2080 ti - 41.7FPS
RTX 3080 - 56.47fps (177.6%)


Doom Eternal (4k/Ultra Nightmare)
RTX 2080 - 82.3fps
RTX 2080 ti - 105.8FPS
RTX 3080 - 156fps (189.6%)


The 3070 could come in somewhere between the 2080ti and 3080, but we don't know. Also, comparing TFLOPs across generations is not indicative of much. Compare TFLOPS within the same generation of cards to get a better indication of performance.
 
OP
OP
Flandy

Flandy

Community Resettler
Member
Oct 25, 2017
3,445
theAllseeingEYE said:
Hold you horses OP, we haven't see actual performance data from the 3070 yet. Jensen said the 3070 was more powerful, not slightly more powerful, and said the same thing for the 3080. In reality, based on the Digital Foundry performance numbers, the 3080 is actually significantly more powerful than the 2080ti.

For example, a chart I've been working on:


Borderlands 3 (4k/Bad Ass/DX12/4K/TAA)
RTX 2080 - 34.33fps
RTX 2080 ti - 48FPS
RTX 3080 - 61.79fps (180%)


Shadow of the Tomb Raider (4k/Highest/TAA)
RTX 2080 - 47fps
RTX 2080 ti - 61FPS
RTX 3080 - 78.96fps (168%)


Battlefield V (4k/Ultra/TAA/DXR OFf)
RTX 2080 - 67fps
RTX 2080 ti - 84FPS
RTX 3080 - 112fps (168%)


Control (4k/High/TAA/DLSS OFf)
RTX 2080 - 31.8fps
RTX 2080 ti - 41.7FPS
RTX 3080 - 56.47fps (177.6%)


Doom Eternal (4k/Ultra Nightmare)
RTX 2080 - 82.3fps
RTX 2080 ti - 105.8FPS
RTX 3080 - 156fps (189.6%)


The 3070 could come in somewhere between the 2080ti and 3080, but we don't know. Also, comparing TFLOPs across generations is not indicative of much. Compare TFLOPS within the same generation of cards to get a better indication of performance.
Click to expand...
Click to shrink...
They're basically on the same level on this chart though
Nvidia-RTX-3000-performance-graph-1212x681.jpg
 
Tagyhag

Tagyhag

Member
Oct 27, 2017
12,483
As others have said, TFLOPS should only ever be compared between the same architecture, and even then, you might as well not even bother until we get actual benchmarks.

Companies use it purely for marketing.
 
BloodHound

BloodHound

Member
Oct 27, 2017
8,999
Best way to measure performance is to measure performance.
TFLOPS like a lot of numbers used in computing technology means jack shit in 2020.
 
LorentzFactor

LorentzFactor

Member
Oct 25, 2017
696
I feel like this is a problem with taking one metric that's potentially useful for computer/computational scientists and treating it as an objective context free number. Maybe looking a HPC (not gaming mind you so it's missing RT-Cores and other important considerations) whitepaper would put things into perspective of just how many things are to be compared (https://www.nvidia.com/content/dam/...ter/nvidia-ampere-architecture-whitepaper.pdf) -- and to be clear, each application has specific needs and is limited in particular ways which is why knowing the games being benchmarked will be important. I'm a computational scientist, so this is what I thought to grab. Maybe someone knows of a similar whitepaper but for consumer gpus?
 

shark97

Banned
Nov 7, 2017
5,327
Mullet2000 said:
Because as people have been aggressively stressing all year TFLOPS is a bad way to measure performance.
Click to expand...
Click to shrink...


Between nearly identical architectures though it's pretty good.

People like to downplay Tflops now but there's a reason it became the one catch all stat.

Yes you could have other factors of the chip be the bottleneck (a higher TF chip bottlenecked by bandwidth, for example), but in most cases everything scales around TFlops. To do anything else would be bad design, and GPU companies aren't in the business of bad design.

Myself I immediately spotted the extremely high TF in Nvidia's new cards relative to past. Just waking up groggily, my initial impulse is that they are changing the way they count flops or pumping the numbers? However I'll have to do more reading to find out the real story.
 
hanshen

hanshen

Member
Jun 24, 2018
3,855
Chicago, IL
because no all flops flops equally. some flops flops more than other flops. 2080ti flops more flops per flops than 3070 flops flops.
 
Sean Mirrsen

Sean Mirrsen

Banned
May 9, 2018
1,159
Here's what I think.

The info card for the 2080Ti says "14 TFLOPS + 14 TIPS", "tips" being a bad acronym for Tera-IOPS. IOPS are operating with, roughly, four times less precision than FP32 FLOPS - at least, if I remember correctly. Like FP16 FLOPS can be "twice as fast" as FP32 FLOPS, IOPS are roughly four times as "fast", albeit requiring specialized hardware and middleware to do so efficiently.

14 divided by 4 is 3.5. So on a sheer performance level, if 3070's Ampere architecture delivers a combined 20 TFLOPS of processing power, Turing's 14+14 adds up to around 17.5. There's still a gap between 17.5 and 20, but it's small enough where I'd say just "more powerful than" is more or less accurate.

Might be completely off base, but makes sense to me at least. :P
 
brawndolicious

brawndolicious

Member
Oct 27, 2017
4,919
Looking at the DF performance comparison, it looks like there is actually a pretty good correlation between the TFlop difference and the in-game fps difference. They compared the 3080 to the 2080 (30TF vs 14TF) and there was about an 80% performance boost across several games.

I may be completely wrong, but I'm thinking maybe Nvidia wants to continue selling 2000 series cards as they might have a better margin on those than the 3000 series. Could explain why all their new cards are $500+, it may be to push the mid-market to get a 2000 series card.
 
Henrar

Henrar

Member
Nov 27, 2017
1,905
digitalrelic said:
Only if comparing two different architectures. If comparing two GPUs from the same architecture, it's very relevant.
Click to expand...
Click to shrink...
It's not quite relevant in that case since GPUs still have some fixed pipeline hardware related to rendering. TFLOPs is pure compute performance.
Within the same architecture GPUs often have more of that hardware, but it still doesn't mean more TFLOPS = more graphics performance.
 
AegonSnake

AegonSnake

Banned
Oct 25, 2017
9,566
Flandy said:
2080Ti is 14 TFLOPs give or take
3070 is 20 TFLOPs
Nvidia said the 3070 is only slightly more powerful than the 2080Ti. Why the huge increase in TFLOPs? It's making me take the 30 and 36 numbers for the 3080/3090 with a much bigger grain of salt
NVIDIA_Gamescom_2018_GeForce_RTX_20_Series_Launch_7.png

Eg1-sIpVgAEEZnX.jpg
Click to expand...
Click to shrink...
rtx 2080 ti isnt 14 tflops. its 17 tflops. dont go about advertised nvidia clocks. look at the gpu clocks during benchmarks. its always around 2ghz.
 
Deimos

Deimos

Member
Oct 25, 2017
5,767
TFLOPS mean nothing. You have to look at TFLIPS for a meaningful measure of performance.
 
You must log in or register to reply here.