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.

 

[kurosaki]

based chip at all? I thought farther back in the thread that it was thought the 80CU card would use HBM, and the two smaller chips would be GDDR6?

You can pretty much bet NV was saying the 3070 is faster when RTX and DLSS is enabled. I can believe that if that's what they were trying to say.

So, its faster if you play the same game, but on the newer card, you play it on a lower resolution. makes sene.... Hate DLSS. That ugly son of a hog should just be shot on the spot.

 

[Det0x]

NVIDIA GeForce RTX 3090 reviewed, only 10% faster than RTX 3080?

NVIDIA GeForce RTX 3090 gaming performance review leaks out - VideoCardz.com

It seems that TecLab is back with yet another leaked review. NVIDIA GeForce RTX 3090 reviewed, only 10% faster than RTX 3080? The guys who were the first to publish RTX 3080 review are now back with another piece, but this time covering the flagship model — GeForce RTX 3090. The video by the...

videocardz.com

The guys who were the first to publish RTX 3080 review are now back with another piece, but this time covering the flagship model — GeForce RTX 3090. The video by the Chinese TecLab team features synthetic and gaming performance.

The test was performed on the same system as before, featuring an Intel Core i9-10900K processor, and DDR4 memory clocked at 4133 MHz.

In the video the leakers do not refer to the RTX 3090 by its actual name, instead, it is shown as a comparison between a “10 thousand Yuan Flagship” and “5 thousand Yuan Flagship” (which is RTX 3090 1500 USD and RTX 3080 700 USD respectively).

We took screenshots of the video in case it will be taken down (like it was the last time). We attached a table with all the important values for a quick comparison.

Please note that we have not confirmed the values presented by the leakers, but the last video covering RTX 3080 performance was indeed accurate. Also, check our coverage of synthetic performance here and here.



If i were to guess the reason for this small performance difference between the 3080 and 3090 i would say it all comes down to the TDP targets.. 3090's 350w is servery power limited compare to the 3080's 320w.. (350/320= is only ~9% more)
The 3090 have more then twice the memory +20% CUDA cores and bandwidth, but it doesn't have the powertargets to to properly utilize it

This probably means that the 3090 would gain more from overclocking then the 3080, if you are prepared to run the 3090 around and above 400w

 

[Zstream]

NVIDIA GeForce RTX 3090 reviewed, only 10% faster than RTX 3080?

NVIDIA GeForce RTX 3090 gaming performance review leaks out - VideoCardz.com

It seems that TecLab is back with yet another leaked review. NVIDIA GeForce RTX 3090 reviewed, only 10% faster than RTX 3080? The guys who were the first to publish RTX 3080 review are now back with another piece, but this time covering the flagship model — GeForce RTX 3090. The video by the...

videocardz.com

The guys who were the first to publish RTX 3080 review are now back with another piece, but this time covering the flagship model — GeForce RTX 3090. The video by the Chinese TecLab team features synthetic and gaming performance.

The test was performed on the same system as before, featuring an Intel Core i9-10900K processor, and DDR4 memory clocked at 4133 MHz.

In the video the leakers do not refer to the RTX 3090 by its actual name, instead, it is shown as a comparison between a “10 thousand Yuan Flagship” and “5 thousand Yuan Flagship” (which is RTX 3090 1500 USD and RTX 3080 700 USD respectively).

We took screenshots of the video in case it will be taken down (like it was the last time). We attached a table with all the important values for a quick comparison.

Please note that we have not confirmed the values presented by the leakers, but the last video covering RTX 3080 performance was indeed accurate. Also, check our coverage of synthetic performance here and here.

View attachment 30169

If i were to guess the reason for this small performance difference between the 3080 and 3090 i would say it all comes down to the TDP targets.. 3090's 350w is servery power limited compare to the 3080's 320w.. (350/320= is only ~9% more)
The 3090 have more then twice the memory +20% CUDA cores and bandwidth, but it doesn't have the powertargets to to properly utilize it

This probably means that the 3090 would gain more from overclocking then the 3080, if you are prepared to run the 3090 around and above 400w

Well yeah, you can’t have 400W space heater. You will get plenty of RMA’s for lack of cooling in a case or PSU’s blowing up.

 

[Head1985]

Well those game benchmark looks good for 3070 .3090 is only 5-11% faster than 3080 in 4k.I can see now 3080 25-30% faster than 3070 so 3070 could match 2080TI or even be 5% faster.There is something wrong with ampere and this architecture just dont scale well.
Probably number of GPC???

 

[Hitman928]

Well those game benchmark looks good for 3070 .3090 is only 5-11% faster than 3080 in 4k.I can see now 3080 25-30% faster than 3070 so 3070 could match 2080TI or even be 5% fatser.There is something wrong with ampere and this architecture just dont scale well.
Probably number of GPC???

It's hard to scale when your next card up only has a 9.4% increase in TDP.

 

[Heartbreaker]

Just the day before we were discussing this one with ~19%:

NVIDIA GeForce RTX 3090 3DMark Time Spy scores leaked - VideoCardz.com

We have the first benchmarks results for the upcoming flagship TITAN-class graphics card, the GeForce RTX 3090. NVIDIA GeForce RTX 3090 is ~19% faster than RTX 3080 The GeForce RTX 3090 scores 20387 ‘graphics’ points in Time Spy 1440p preset. The same card scores 10328 points in Time Spy Extreme...

videocardz.com

One may be with power limits on, and the other with power limits off.

20% more cores on 3090 could deliver up to 20% more performance, but it never made sense that it didn't have at least 20% larger TDP to go with it.

Really, no one should be buying this for gaming. The small increase in performance and massive increase in price makes no sense over the 3080 for gaming, and even no sense for professional workloads that can fit in 10GB.

3090 makes more sense for people with professional workloads that need big memory. For them it could be a bargain that quickly pays for itself.

Mind you that won't stop gamers with money to burn and scalper from making them sell out in seconds. There will be a much more limited supply of 3090s given they that it's a much more fully enabled die.

 

[raghu78]

Well those game benchmark looks good for 3070 .3090 is only 5-11% faster than 3080 in 4k.I can see now 3080 25-30% faster than 3070 so 3070 could match 2080TI or even be 5% faster.There is something wrong with ampere and this architecture just dont scale well.
Probably number of GPC???

The number of GPC increasing by only 1 for the RTX 3090 and remaining same for RTX 3080 compared to RTX 2080 Ti, RTX 2080 Super, RTX 2080 and RTX 3070 is limiting the performance gain at 1440p where its likely to be not purely shader performance bound but geometry bound. At 4K the shader bound scenario is more common so we see the RTX 3090 and RTX 3080 pull a little bit more from RTX 2080 Ti, RTX 2080 Super and RTX 2080. We also need to see if Vermeer is better than CML for gaming by a significant enough margin to allow CPU performance scaling at 1440p for RTX 3080 / RTX 3090 vs 10900k

 

[Head1985]

The number of GPC increasing by only 1 for the RTX 3090 and remaining same for RTX 3080 compared to RTX 2080 Ti, RTX 2080 Super, RTX 2080 and RTX 3070 is limiting the performance gain at 1440p where its likely to be not purely shader performance bound but geometry bound. At 4K the shader bound scenario is more common so we see the RTX 3090 and RTX 3080 pull a little bit more from RTX 2080 Ti, RTX 2080 Super and RTX 2080. We also need to see if Vermeer is better than CML for gaming by a significant enough margin to allow CPU performance scaling at 1440p for RTX 3080 / RTX 3090 vs 10900k

And 3070 this time have 96rops+6GPC like 3080.

 

[Heartbreaker]

And 3070 this time have 96rops+6GPC like 3080.

3080 has ~50% more CUDA/RT/Tensor cores, and 70% more memory BW than the 3070. ROPs aren't going to make up for all those deficits on the 3070.

IMO, Anyone expecting 3070 to really beat a 2080 Ti is in for a disappointment when the card with 50-70% more resources, only exceeds 2080 Ti by up to 30%.

 

[Grooveriding]

It better be more than 10% for twice the price :O Otherwise 3080 is going to be even more impossible to get and price gauged to the moon and back.

 

[DJinPrime]

3080 has ~50% more CUDA/RT/Tensor cores, and 70% more memory BW than the 3070. ROPs aren't going to make up for all those deficits on the 3070.

IMO, Anyone expecting 3070 to really beat a 2080 Ti is in for a disappointment when the card with 50-70% more resources, only exceeds 2080 Ti by up to 30%.

The 3070 might be more balance for gaming. It has the same # of GPC as the 3080 and around 1000 more cores than the 2080ti. Memory BW is overhyped, it's important for compute but I don't think it's as important for gaming. Otherwise, the Radeon HBM cards would have destroyed the NV cards of the same generation. Also with lesser cores, the 3070 should boost better than the 3080. On paper, it should beat the 2080ti (1% counts), but will see.

It better be more than 10% for twice the price :O Otherwise 3080 is going to be even more impossible to get and price gauged to the moon and back.

Specs wise, you shouldn't expect that big of a jump, 10-15% is probably about it. Looks like it's just an extra GPC with a few extra cores per and slightly more bandwidth. The extra memory is completely wasted for gaming. The extra cores and unnecessary memory will just eat up power, so it will boost even less than 3080. Don't buy 3090 for gaming, unless money is not an issue. Should have kept the Titan name, would be less confusing.
Quoting the white paper: "The GeForce RTX 3090 is a beast. Like the previous generation TITAN RTX, ", so why not call it Titan RTX2?

 

[MrTeal]

The 3070 might be more balance for gaming. It has the same # of GPC as the 3080 and around 1000 more cores than the 2080ti. Memory BW is overhyped, it's important for compute but I don't think it's as important for gaming. Otherwise, the Radeon HBM cards would have destroyed the NV cards of the same generation. Also with lesser cores, the 3070 should boost better than the 3080. On paper, it should beat the 2080ti (1% counts), but will see.

It has a good bit more FP32 cores than the 2080 Ti (5888 vs 4352), but keep in mind the 2080Ti effectively also has 4352 Int cores as well while the Ampere has to run Int operations on the same cores as FP. If Nvidia's ratio of 36 Int to 100 FP is correct, some napkin math would show the 2080 Ti using 4352 FP + 1567 Int cores (5918) while the 3070 uses its 5888 cores. Should be interesting to see if that translates to a rough performance tie or not.

 

[Heartbreaker]

The 3070 might be more balance for gaming. It has the same # of GPC as the 3080 and around 1000 more cores than the 2080ti. Memory BW is overhyped, it's important for compute but I don't think it's as important for gaming. Otherwise, the Radeon HBM cards would have destroyed the NV cards of the same generation. Also with lesser cores, the 3070 should boost better than the 3080. On paper, it should beat the 2080ti (1% counts), but will see.

Excess BW is unimportant. So HBM won't make a slower card, faster, but I doubt many cards come with excess excess BW outside of the HBM cards. Starving it for BW does make fast card, slower. If the 3070 only needs 448 GB/s, then the 3080 wouldn't need expensive fancy GDDR6X memory.

BW: 2080 Ti 616 GB/s 3070 448 GB/s. ~38% more BW on 2080 Ti Massive deficit for 3070.

Texel Fill Rate: 2080 Ti 420 Gt/s 3070 317 Gt/s ~32% higher T-Fill Rate on 2080 Ti

Tensor Cores: 2080 Ti 544, 3070 184 ~3X as many on 2080 Ti. New Tensor marketing claim 2x-4x as good. So if marketing numbers true. Trade blows. Really only affects DLSS, probably a wash.

RT cores: 2080 Ti 68, 3070 46 ~48% more cores on 2080 Ti. 70% more perf on 2nd gen marketing claim. In practice looks more like 50% - so Trade blows, so again probably a wash.

From where I sit, on Tensor/RT cores, it looks like a wash. While on older basics like Memory BW and Texture Fill rate 3070 gets decimated.

Only on CUDA FP32 does 3070 pull ahead, and combined with the weak BW and Texture fill rate, it's hard to see how the 3070 leverages this one advantage.

I am betting 3070 being a let down, where the 3070 loses more than it wins.

 

[alcoholbob]

So with the 3090 TDP bound, it seems like for most sane people the best bet is the wait for the 20GB 3080, especially if you are running a SFF PC build.

 

[sze5003]

Nvidia's upcoming RTX 3060 Super 8 GB, RTX 3070 Super 16 GB and RTX 3080 Super 20 GB models confirmed by Gigabyte listing

Gigabyte's Watch Dogs Legion redemption site includes a list with all eligible cards, and, among these, we can spot the unannounced RTX 3060 with 8 GB of VRAM, along with two RTX 3070 GPUs packing 16 GB VRAM and three SKUs for the RTX 3080 featuring 20 GB VRAM. Presumably, all these will bear...

www.notebookcheck.net

Makes me feel better if I cant get a 3090. I'll be more than happy with a 20gb 3080 for the next few years.

 

[IntelUser2000]

Excess BW is unimportant. So HBM won't make a slower card, faster, but I doubt many cards come with excess excess BW outside of the HBM cards. Starving it for BW does make fast card, slower. If the 3070 only needs 448 GB/s, then the 3080 wouldn't need expensive fancy GDDR6X memory.

I am betting 3070 being a let down, where the 3070 loses more than it wins.

Hmm, I'm agreeing with this. Even in RT it doesn't look that much better.

For the 3090, its still too early to make a judgment. Based on all the crap rumors about Ampere, nothing less than 3rd party benchmarks can be trusted.

Remember Tensor Memory Compression? What about 4x faster Ray Tracing? Even a dedicated chip for it?

 

[A///]

16 GB 3070 Super sounds nice.

Not sure how it'll perform. In any case, it may be better than a 2080? There's a few games I want to play over the next 2 years, and I like to play at 1440p 95% of the time, maybe 1080p if it looks better somehow. 4K gaming is... I don't see the point of it, but I'm not a hardcore gamer like most of you.

 

[DJinPrime]

Excess BW is unimportant. So HBM won't make a slower card, faster, but I doubt many cards come with excess excess BW outside of the HBM cards. Starving it for BW does make fast card, slower. If the 3070 only needs 448 GB/s, then the 3080 wouldn't need expensive fancy GDDR6X memory.

That's the thing, no one really knows how much bandwidth is needed exactly. Didn't someone posted that overclocking the memory barely increased 3080 performance. Might have seen it in some video or read it somewhere else. So, it doesn't seem to be memory bound. Well, GDDR6x might be a gimmick for gaming, would be good for the professional side.

 

[antihelten]

Sorry, but a loaf of bread isn't 10 cents anymore either. Prices have gone up permanently, wafers are much more expensive with each fab steps, with the biggest jumps on the most recent ones.

Obviously Inflation exists, but it isn't really the main determinant in how Nvidia sets their prices, unless you think there was inflation of over 70% between 2017 when Nvidia released the 1080 Ti and 2018 when they released the 2080 Ti.

With regards to wafer prices you are absolutely right, but this clearly goes against your theory. If wafer prices was a major factor in how Nvidia markets their GPUs, then we would expect them to give us smaller and smaller dies over time for the same price, but with the 3080 the exact opposite has happened, they have given us a bigger die (102 instead of 104).

What matters, and you really can't argue against, is performance.

Every x80 Ti cards perform like Titans and always has as long as there were Titan cards.

The 3080 clearly does not perform like this generations Titan.

In a few months when the 3080 Ti is released you will know you were wrong, though I doubt you will admit it even then.

Except it actually looks like the gap between the 3080 and the 3090 is exactly what you would expect between the x80 Ti and the titan card (ie. 5-10%):

leaked numbers for the 3090.

With that small of a gap there is clearly no room for a 3080 Ti in between the two.

So I'm curious. If these numbers hold, will you admit to being wrong, or are you going to do what you're already prematurely accusing me of?

 

[badb0y]

Omg. The 3080 is going to be impossible to get now if those leaks are true.

 

[VirtualLarry]

Well yeah, you can’t have 400W space heater. You will get plenty of RMA’s for lack of cooling in a case or PSU’s blowing up.

400W??? Is that what we've come to?

Think about this factor... CPUs have been increasing (slowly, more slowly than GPUs) in performance, in roughly the same TDP window. The same cannot be said for GPUs, which, argueably, are much still in the same phase as early CPU development, most of the new-found performance, is gained, from increasing the TDP budget over previous cards. Granted, they do have more transistors as well. So, perhaps, the power usage scales up with transistor count and Mhz, and this increase in TDP is overall to be expected. Although, a bit disappointing, for a company that started down the performance/watt path, with the Maxwell V1 GTX 750ti, which had a premium price over the equivalent-performance AMD card, almost solely based on the lower TDP.

So viewing through THAT lens, the GTX 3000 series is probably an abject failure, especially if AMD's "Big Navi" slots in at around the same performance, with better performance/watt.

 

[DJinPrime]

That's the thing, no one really knows how much bandwidth is needed exactly. Didn't someone posted that overclocking the memory barely increased 3080 performance. Might have seen it in some video or read it somewhere else. So, it doesn't seem to be memory bound. Well, GDDR6x might be a gimmick for gaming, would be good for the professional side.

Found it, in Kitguru's review https://www.kitguru.net/components/...s/nvidia-rtx-3080-founders-edition-review/30/
they bumped the core by 60 and then bumped the memory until performance started dropping due to built in protection. His results show very little gain with the memory oc, maybe 1 or 2 fps.

 

[Justinus]

While the 3090 seems to be TDP bound, a report came out where someone has shunt modded a 3090, overclocking it to the max.

It drew 550w and only improved performance 6% from the stock 3090.

The 3090 is a dumpster fire.

Nvidia really screwed up by attempting to teach TSMC a lesson on pricing and using Samsung.

 

[.vodka]

It drew 550w and only improved performance 6% from the stock 3090.

Nvidia really screwed up by attempting to teach TSMC a lesson on pricing and using Samsung.

LOL. This makes Fermi look like a success...

At least Samsung can manufacture the chips, that's something.

 

[Heartbreaker]

Except it actually looks like the gap between the 3080 and the 3090 is exactly what you would expect between the x80 Ti and the titan card (ie. 5-10%):

leaked numbers for the 3090.

With that small of a gap there is clearly no room for a 3080 Ti in between the two.

So I'm curious. If these numbers hold, will you admit to being wrong, or are you going to do what you're already prematurely accusing me of?

It might be power limited for FE, but it won't for AIB cards that aren't as power limited. With 20% more functional units, ~20% gains will happen.

Even if they didn't, 3080 Ti, does NOT have to be slower than 3090.

Go back and check Kepler 780, Titan and 780 Ti. 780 Ti outperformed the Titan.

Also note that was the last time that x80, Titan and x80 Ti all shared the same die.