arestavo
My 3090 FTW3 Ultra at stock (power slider maxed) only gets about 1100 points more for the GPU score, so I don't think you're that far off considering that the 30 series is totally power draw limited - https://www.3dmark.com/spy/14211835

redarman
Here is some more detail on my best 3090 score vs my best 2080ti score so far. Graphics score is 20% improvement from 2080ti to 3090. In bitwit's video that matches his gain to a 3080. Shouldn't I get an additional 5-10% the a 3090 instead of a 3080? I know there are a lot of variables here.

attached screenshot
 

wolfe155
Close X1 when you run time spy. I got an extra 1000 from that. 
 
I got a score of 20112 on my 3090 FTW3 Ultra with +900 on Memory and +90 on Clock. it was pretty stable and i could play RDR2 maxed out at 65 fps
 
also my scores started at in the low 18000 before i started messing with fan speeds and overclocking

MatthewAMEL

vulcan1978
 
Here's what upsetting the discrete GPU market looks like: 
 
Navi 21 is shaping up to be as fast as the 3090 with 16GB of video memory on 7nm TSMC node with a 256 bit bus (128mb cache will compensate for this) @ 300w (100w less than GA-102-300) for $1000 and will run all of the next gen console ports, whose underlying architecture, RDNA 2, is the same and whose engines, including Unreal engine, will be heavily optimized around. Both next-gen consoles will have native path tracing, AI super-sampling and their equivalent of 'RTX IO' (AMD invented this first, NGreedia followed suit). 
 

You were doing fine with your post until this.
 
There are no leaked numbers or expectations that say Big Navi will be on par with the 3090. Everything to date says it slots between 3070 and 3080. Wish it was a better performer, but that lines up with their goal of double a 5700XT performance. What remains to be seen is how much headroom is avail due to TSMC's 7nm process.
 
I assume when you talk about RTX IO you are referring to AMD's HBCC? HBCC was forward thinking of AMD, but still sits unused. RTX IO is Nvidia's implementation of DX12's DirectStorage. Since XBSX is using RDNA2(ish), I would assume AMD has HBCC or a newly re-vamped version of it ready to go.
 
Also, just FYI, the Samsung 8nm isn't EUV. Their 7nm is and their 5nm will be then transitioning to GAA. It's one of the reasons there are such power inefficiencies with Ampere.
https://news.samsung.com/...ion-of-8nm-lpp-process

MatthewAMEL

vulcan1978
 
Here's what upsetting the discrete GPU market looks like: 
 
Navi 21 is shaping up to be as fast as the 3090 with 16GB of video memory on 7nm TSMC node with a 256 bit bus (128mb cache will compensate for this) @ 300w (100w less than GA-102-300) for $1000 and will run all of the next gen console ports, whose underlying architecture, RDNA 2, is the same and whose engines, including Unreal engine, will be heavily optimized around. Both next-gen consoles will have native path tracing, AI super-sampling and their equivalent of 'RTX IO' (AMD invented this first, NGreedia followed suit). 
 

You were doing fine with your post until this.
 
There are no leaked numbers or expectations that say Big Navi will be on par with the 3090. Everything to date says it slots between 3070 and 3080. Wish it was a better performer, but that lines up with their goal of double a 5700XT performance. What remains to be seen is how much headroom is avail due to TSMC's 7nm process.
 
I assume when you talk about RTX IO you are referring to AMD's HBCC? HBCC was forward thinking of AMD, but still sits unused. RTX IO is Nvidia's implementation of DX12's DirectStorage. Since XBSX is using RDNA2(ish), I would assume AMD has HBCC or a newly re-vamped version of it ready to go.
 
Also, just FYI, the Samsung 8nm isn't EUV. Their 7nm is and their 5nm will be then transitioning to GAA. It's one of the reasons there are such power inefficiencies with Ampere.
https://news.samsung.com/...ion-of-8nm-lpp-process

terry2776

MatthewAMEL

vulcan1978
 
Here's what upsetting the discrete GPU market looks like: 
 
Navi 21 is shaping up to be as fast as the 3090 with 16GB of video memory on 7nm TSMC node with a 256 bit bus (128mb cache will compensate for this) @ 300w (100w less than GA-102-300) for $1000 and will run all of the next gen console ports, whose underlying architecture, RDNA 2, is the same and whose engines, including Unreal engine, will be heavily optimized around. Both next-gen consoles will have native path tracing, AI super-sampling and their equivalent of 'RTX IO' (AMD invented this first, NGreedia followed suit). 
 

You were doing fine with your post until this.
 
There are no leaked numbers or expectations that say Big Navi will be on par with the 3090. Everything to date says it slots between 3070 and 3080. Wish it was a better performer, but that lines up with their goal of double a 5700XT performance. What remains to be seen is how much headroom is avail due to TSMC's 7nm process.
 
I assume when you talk about RTX IO you are referring to AMD's HBCC? HBCC was forward thinking of AMD, but still sits unused. RTX IO is Nvidia's implementation of DX12's DirectStorage. Since XBSX is using RDNA2(ish), I would assume AMD has HBCC or a newly re-vamped version of it ready to go.
 
Also, just FYI, the Samsung 8nm isn't EUV. Their 7nm is and their 5nm will be then transitioning to GAA. It's one of the reasons there are such power inefficiencies with Ampere.
https://news.samsung.com/...ion-of-8nm-lpp-process

Holy crap dude, sounds like you need a hug and a big navi card, but then you would probably need to code your own drivers to have something stable but at least you would have real problems
 
Edit: That was supposed to quote Vulcan1978 not Matt

vulcan1978

arestavo
My 3090 FTW3 Ultra at stock (power slider maxed) only gets about 1100 points more for the GPU score, so I don't think you're that far off considering that the 30 series is totally power draw limited - https://www.3dmark.com/spy/14211835

It's not "power draw limited", 3090 Strix only gains another 3% at 480W vs 390w and we see that shunted 3080 ( 'vmanuelgm') is only 7-10% faster @ 550w vs 390w!
 
The problem is that NGreedia had to pre-overclock Ampere right at the point in the performance-efficiency curve where any more freq requires a non-linear, inordinate increase in voltage and wattage. 
 
This is the case with any silicon. My 2080 Ti for example, it will do 2055 MHz core with an undervolt (1.025v on volt/freq curve). It can do 2130 MHz but this requires nearly another 50mv (1.069v) and another 50w+ on average. 
 
2055 MHz is done with ~300-320w and then the last 80 MHz requires another 50w up and over that. 
 
This is where Ampere is clocked from the factory. 
 
It's why people are having black-screens (aside from the POSCAP, MLCC issue). 
 
Had NGreedia gone with 7nm TSMC the 3080 would have it's present performance with a 425mm2 die @ 225w and the 3090 would easily be 40% faster with a 628mm2 die at only 320w! 
 
And it would overclock way better, 25% overclock, like what is capable with Turing, would be possible. 
 
They couldn't do that though, they needed to save $26 per yield. They are really smart, they know that with slick marketing and by gimping the outgoing architecture via driver updates for newer games, i.e. Cyberpunk 2077 where the 3080, although only 10-15% faster than 2080 Ti at the same power draw, will probably "magically" be something ridiculous like 30-40% faster in Cyberpunk 2077 as NGreedia "focuses on the new architecture", a euphemism they use to mean that they will make the outgoing architecture appear slow in comparison via driver shenanigans. 
 
That's great! So instead of actually releasing a faster, quality product they can simply buy out tech-tubers (Digital Foundry) and gimp the outgoing architecture via driver updates with newer games that have yet to be released on the older architecture so that you don't even know it's happening and have no reference point. 
 
RTX 3090, 25% faster than 2080 Ti (same power draw): $1500-1800.
GTX 2080 Ti, 50% faster than 1080 Ti at the same power draw: $1200.
GTX 1080 Ti, 50% faster than 980 Ti at the same power draw: $700
GTX 980 Ti, 50% faster than 780 Ti at the same power draw: $700
GTX 680, 50% faster than 580 at the same power draw
GTX 580, 50% faster than 480 at the same power draw
 
All of these cards I've personally owned. 
 
TU-102 is capable of a 30% overclock. My personal sample is doing nearly 25% overclock (13,600 Timespy GPU @ 260w to 16,900 @ 373w @ 1.063v)
 
GA-102-200 is capable of a 7% overclock (~17,500 to 18,750-19k @ 370w) 
 
GA-102-300 is capable of a 5% overclock (20,500 to 21,500 @ 393w)
 
Ampere is clocked so aggressively because it needed to impress the masses on the inferior node and the only way to do that is to pre-overclock GA-102 20% into it's 30% overclock. 
 
Hence the black screens (although the mismatch between POSCAP and MLCC caps is implicated AIB's have been doing this for years, only now is this a problem because Ampere is clocked so aggressively. The solution short of RMA? Underclock by 100 MHz)
 
If anyone is buying this mediocrity before waiting to see how Big Navi shapes up is voting for more mediocre gains and price gouging with whatever NGreedia dishes up next. 
 
Don't buy this garbage. 
 
Nvidia's decision to got with 8nm EUV was their dumbest decision, now AMD is going to clean their clock, will do to them what they did to Intel with 14+++++++++. NGreedia has become like Intel. Their sitting on their laurels because they falsely believe they have no competitor. 
 
Ampere will be viewed in retrospect as NGreedia's Comet Lake moment when AMD upset the discrete GPU market. 
 
Here's what upsetting the discrete GPU market looks like: 
 
Navi 21 is shaping up to be as fast as the 3090 with 16GB of video memory on 7nm TSMC node with a 256 bit bus (128mb cache will compensate for this) @ 300w (100w less than GA-102-300) for $1000 and will run all of the next gen console ports, whose underlying architecture, RDNA 2, is the same and whose engines, including Unreal engine, will be heavily optimized around. Both next-gen consoles will have native path tracing, AI super-sampling and their equivalent of 'RTX IO' (AMD invented this first, NGreedia followed suit). 
 
Here's an idea as to what's coming with Big Navi. 
 
The APU in the XBOX Series X is the CPU and GPU compute equivalent of an AMD 3700x and an RTX 2080 @ 170w. 
 
Navi 21 will be that all GPU compute @ 300w. 
 
NGreedia rushed Ampere out of the door to try to beat AMD to the punch. 
 
NGreedia is done, anyone still supporting them is voting against their own interests as a consumer. 
 
They killed off 3D Vision to save pennies. I literally have no reason other than the G-Sync module in my AW3418DW to remain trapped in their ecosystem any longer. 
 
How much work and money did adding the 3D Vision driver bits forward to the next display driver update entail? 5 minutes of work, copying and pasting a few lines of code? 
 
There are thousands of us still enjoying 3D Vision (PG278Q here). We have simply modified newer display drivers, taking the bits from 425.31 and adding them to newer drivers via 3D Fix Manager (credit to Paul Dusler). 
 
Now it appears that we can no longer do that. 
 
I have zero reason to remain with this shady ass corporation. 
 
The decision to go with 8nm EUV was an economic one (to save $26 per yield, and no, the 7nm queue was not forecasted to be full as it was known crApple was moving to 5nm. 7nm TSMC is currently only at 30% capacity). 
 
They knew that they had a captive market that would buy anything with sophisticated marketing. $26 per yield is $26 per yield! Granted, the consumer would evaporate $26 in the form of higher electricity bills in 3-4 months vs the same performance from a 225w 425mm2 7nm TSMC die, and guess what? None of that $26 savings was passed on to the consumer. Youre still buying ridiculously overpriced garbage, ESPECIALLY the 3090. And the environment (what's that? Unprecedented wild-fires in the Arctic, West Coast, Australia, the Atlantic ocean hurricane season starting 1 month early because the surface temp is 7 degrees above average?)? Imagine everyone who is buying a 3080 @ 370w when that same level of compute is possible with a 7nm TSMC die @ 425mm2 @ 225w. This is criminal. 
 
Their first stupid decision was the decision to sell discrete GPU's hand over fist to mining farms in Asia and the U.S. in 2017 and telling investors that these were purchases from people building gaming PC's for the first time. 
 
Crypto-bubble crashed, and they had all of this built up investor activity, so they had to keep profits high without moving inventory. Best way to accomplish that? Price the new GPU's higher! How can they do that? Introduce some new technology! Introducing the 20 series GPU's where every GPU was renamed one GPU higher on the product stack in a vain attempt to conceal a literal doubling in historical pricing! 
 
RTX 2070? TU-106 SKU. 106 designates 60 card SKU. No SLI, heretofore every 70 card has had an SLI tab. Doesn't share PCB with the 80 card, heretofore every 70 card shares PCB with 80 card. Only as fast as outgoing 80 card, not 80 Ti card like every new 70 card before it, i.e. GTX 1070, as fast as 980 Ti and GTX 970, as fast as 780 Ti. Making for a $600 60 card in actuality that can't even do ray tracing at 1080p 
 
Same story for the 2080, which is basically a $900 70 card. 
 
And the 2080 Ti, a glorified $1200 80 card. (I bought my Samsung B-Die equipped XC2 used, local, no sales tax with 2 years remaining on the transferable warranty for $900 around 6 months ago and I already had the Phanteks water block on hand. I need the upgrade because 1080 Ti wasn't enough for 3440x1440, 3D Vision, and VR and the upgrade has yielded 50% gains in all of my demanding titles, as represented by the Timespy GPU improvement. I would NOT have paid $1200 before taxes for this and another $250 for water-block and back-plate before taxes, even at 50% performance gain and Ampere only represents a 25% gain from here for $2300 estimated after taxes for quality AIB variant and water-block. Twice the price for half the gain. Masterful marketing Jacket Man!) 
 
And now they are at it again with their shenanigans! They learned real quick this time that the best way to overstate the performance jump is to simply buy out a tech-tuber and allow them to show specially curated "benchmarks" showing how fast the new architecture is two weeks before anyone else! According to Digital Foundry the 3080 is 90% faster than the 2080 and 65% faster than 2080 Ti! 
 
In reality the 3080 is 10-15% faster than 2080 Ti when running both cards as the same power draw. 
 
And the brain-dead zombified masses eat it up. 
 
Then, NGreedia intentionally botch the launch (they knew scalpers with bots would occur, this is by design: 1. it lets them off the hook from warranty responsibility. 2. It creates a situation of artificial scarcity, which creates hysteria and and amplifies demand. Then they invariably "run out" of the FE variant, which as MLID has point out, is intentional, as they intend to make the bulk of their profit through AIB sales, replete with inferior components and which cannot run as fast "just underclock your AIB 3080 by 100 MHz to resolve the black-screen issues". 
 
And now? 
 
Now they will simply gimp the outgoing architecture driver-side with newer games. They will call it "focusing on the new architecture". Here's how it works. They set out to UNOPTIMIZE the outgoing architecture with a newer game where the consuming public has no reference point. (they can't do it with existing games, well they can, but this is already becoming TLDR). 
 
So then the benchmarks will show: 
 
RTX 3080 = 50% faster than 2080 Ti in Cyberpunk 2077!
 
When in reality, the 3080 isn't actually faster, they simply GIMPED the 2080 Ti with this game to make it look faster. 
 
Stop buying this mediocrity. Do not support this shady corporation. It's time to support the competition. We are in this mess with a $1500 80 Ti card that is only 25% faster than the outgoing 80 Ti card at the same power draw because mindless morons support it with their wallet. 
 
Wait for Big Navi, due out in one month. 
 
 

slayer6288
 
You are an idiot the 2080ti was not 50% faster than a 1080ti unless you had an extreme ocer. Dont spread fud like the clown you are

arestavo
My 3090 FTW3 Ultra at stock (power slider maxed) only gets about 1100 points more for the GPU score, so I don't think you're that far off considering that the 30 series is totally power draw limited -