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.

 

[sontin]

I wont gloat. But this "leakbench" guy has no history of leaks towards Ampere. Unlike these other two people. And both confirmed the same Timespy Extreme score three weeks ago:

https://twitter.com/x/status/1289148415293628417

 

[pj-]

The thing that baffles me is...

No way Jenhsen will let it be this slow. Not compared to Navi 2. Nvidia Engineers know, (well they know 90-95% at this point...) how Navi 2 will behave. They can clock it to hell.

Why only 35%?

I don't buy it. It will be faster. Around 40-45% in 1080p, and faster in 4K.

I'm baffled by you linking to a tweet showing 35% uplift, then linking to a related tweet to "disprove" a claim that maybe it meant 2080ti -> 3080, then posting that you don't believe the first tweet. It seems like you're picking and choosing which leaks to believe, even from the same source. I don't see how that's any different than pure speculation. Are you trying to post every possibility at least once so that no matter what happens you can say you weren't wrong?

 

[Glo.]

I'm baffled by you linking to a tweet showing 35% uplift, then linking to a related tweet to "disprove" a claim that maybe it meant 2080ti -> 3080, then posting that you don't believe the first tweet. It seems like you're picking and choosing which leaks to believe, even from the same source. I don't see how that's any different than pure speculation. Are you trying to post every possibility at least once so that no matter what happens you can say you weren't wrong?

I simply do not believe it will be only 35% faster.

Am I allowed NOT to believe a "leaker"/"rumorer"?

 

[blckgrffn]

I simply do not believe it will be only 35% faster.

Am I allowed NOT to believe a "leaker"/"rumorer"?

Only the accurate ones.

 

[CastleBravo]

Only the accurate ones.

How about this?

https://twitter.com/x/status/1298039731595415552

 

[Krteq]

MLID again

 

[Maverick177]

I wont gloat. But this "leakbench" guy has no history of leaks towards Ampere. Unlike these other two people. And both confirmed the same Timespy Extreme score three weeks ago:

https://twitter.com/x/status/1289148415293628417

Yeah, 1 random tweet and suddenly everyone swears by it. 35% faster with almost twice the graphics score ? Yeah something does not add up here.

 

[Maverick177]

How about this?

https://twitter.com/x/status/1298039731595415552

Just no. The guy is not trustworthy.

 

[AtenRa]

Thanks! Looking over that documentation; I think it’s pretty clear that for GA102, the total cuda core count will go down as will FP64 FMAs and number of tensor cores. I don’t think consumer TPUs will need fp64 either (unless A102 is also for engineering workstations as well). So I think some reasonable cuts can be made, especially if NV is aiming for higher clocks. Hopefully we’ll see actual products soon and know for sure.

As with Turing, NVIDIA will use the same dies for both consumer and Quadro products with Ampere. So hardware availability will differ in GPU products but hardware will exist in all dies.
For example RTX2080Ti (TU102) had 1:64 FP64 but RTX Quadro 8000 (TU102) had 1:32 FP64.

And since they want to sell more ML products, I dont expect less Tensor Cores per SM in GA102 vs A100

 

[Glo.]

How about this?

https://twitter.com/x/status/1298039731595415552

Depends on the context.

RTX 3070 CAN be as fast as RTX 2080 Ti, if DLSS is what MLID is implying it is.

 

[AtenRa]

Just no. The guy is not trustworthy.

You expecting the RTX3070 to be more than the RTX2080ti ?? and if so by how much ??

Because RTX2070 Super to RTX2080Ti is roughly at +32% at 1440p (according to latest reviews in Techpowerup)

 

[Maverick177]

You expecting the RTX3070 to be more than the RTX2080ti ?? and if so by how much ??

Because RTX2070 Super to RTX2080Ti is roughly at +32% at 1440p (according to latest reviews in Techpowerup)

I expect 3070 to be faster than 2080ti by at least 10% with 3 times less deficit with RT enabled. AMD still gonna lose this round by a large margin unfortunately.

 

[CastleBravo]

I expect 3070 to be faster than 2080ti by at least 10% with 3 times less deficit with RT enabled. AMD still gonna lose this round by a large margin unfortunately.

I would say +10% is close enough to falling into the "margin of error" based on what games/benchmarks you select for comparison that your prediction could line up with 3070 ~= 2080ti.

 

[Maverick177]

I would say +10% is close enough to falling into the "margin of error" based on what games/benchmarks you select for comparison that your prediction could line up with 3070 ~= 2080ti.

It's not my prediction. Turing has multiple problems with tensor cores and RT function, they will be fixed and upgraded with Ampere. Ampere is still ahead of RDNA 2 in terms of everything, the gap will not be as closed as you might think. AMD had a bit of trouble with their power delivery for the big die in their labs last 2 months, might explain why Lisa kept saying RDNA 2 will come out this year. They have been pulling resources since the start of this year for the software alone. The 1x Xbox One power consumption figure you saw on Hotchip slide is actually more like 1.3 ~ 1.4. Big Navi die is power hungry as in tradition. Think Fiji vs GM102 although the die difference will even make it worse.

 

[Glo.]

It's not my prediction. Turing has multiple problems with tensor cores and RT function, they will be fixed and upgraded with Ampere. Ampere is still ahead of RDNA 2 in terms of everything, the gap will not be as closed as you might think. AMD had a bit of trouble with their power delivery for the big die in their labs last 2 months, might explain why Lisa kept saying RDNA 2 will come out this year. They have been pulling resources since the start of this year for the software alone. The 1x Xbox One power consumption figure you saw on Hotchip slide is actually more like 1.3 ~ 1.4. Big Navi die is power hungry as in tradition. Think Fiji vs GM102 although the die difference will even make it worse.

This will not age well.

Especially since we know full well the specs of Xbox Series X...

 

[DiogoDX]

How about this?

https://twitter.com/x/status/1298039731595415552

Well, 2070 super plus 35% is about a stock 2080Ti.

 

[Maverick177]

This will not age well.

Especially since we know full well the specs of Xbox Series X...

Forgot to include 50% is actually meant for the smaller BIG NAVI die at lower clock speed - 5600xt successor. Big Navi is actually only about 30% with very minor IPC uplift. Coretek got it right, partially. 78CU only put it ahead of 2080Ti about 15% in selected titles with tweaked settings. Big Navi is competitor for the most cut down GA102 die. Naming scheme is really weird atm.

 

[Glo.]

How about +35% above RTX 2070?

 

[Maverick177]

How about +35% above RTX 2070?

Ampere will run circles around Turing, Nvidia has plenty of resources to think far ahead. Turing features will eventually be phased out and renamed with something else, all Ampere related. Like I said, naming scheme is really weird.

 

[Glo.]

Forgot to include 50% is actually meant for the smaller BIG NAVI die at lower clock speed - 5600xt successor. Big Navi is actually only about 30% with very minor IPC uplift. Coretek got it right, partially. 78CU only put it ahead of 2080Ti about 15% in selected titles with tweaked settings. Big Navi is competitor for the most cut down GA102 die. Naming scheme is really weird atm.

Its funny that Nvidia had to put you out of coma right now, 3 years since your last activity on this forum in order to make damage control, and post FUD .

 

[AtenRa]

It's not my prediction. Turing has multiple problems with tensor cores and RT function, they will be fixed and upgraded with Ampere.

We are talking about Raster performance not ML/RT.

 

[Maverick177]

Its funny that Nvidia had to put you out of coma right now, 3 years since your last activity on this forum in order to make damage control, and post FUD .

If you think so then I'm about as credible as Moore's Law is dead guy.

 

[Glo.]

If you think so then I'm about as credible as Moore's Law is dead guy.

Well, the fact that you came out of silence 3 years since your last post, right now, and all you do is FUD in Nvidia thread, talking about AMD tech, tells a lot.

Yeah, AMD might have for once kicked Nvidia's ass.

 

[Stuka87]

It's not my prediction. Turing has multiple problems with tensor cores and RT function, they will be fixed and upgraded with Ampere. Ampere is still ahead of RDNA 2 in terms of everything, the gap will not be as closed as you might think. AMD had a bit of trouble with their power delivery for the big die in their labs last 2 months, might explain why Lisa kept saying RDNA 2 will come out this year. They have been pulling resources since the start of this year for the software alone. The 1x Xbox One power consumption figure you saw on Hotchip slide is actually more like 1.3 ~ 1.4. Big Navi die is power hungry as in tradition. Think Fiji vs GM102 although the die difference will even make it worse.

Huh??

What does "in terms of everything" even mean? And whatever it means, what are you basing it off?

Where are you getting this bit on AMD having issues on power delivery? Power delivery is a trivial thing, why on earth would this be an issue?

Are you saying that Microsoft is lying about their power consumption?

And how is there already a tradition of being power hungry for a chip that has never been released? Just because Fiji was a bit more power hungry than the 980Ti doesn't mean RDNA2 will be also. There is literally thing compatible between Fiji and Navi. They aren't even the same ISA, much less a similar chip design.

 

[Maverick177]

We are talking about Raster performance not ML/RT.

Rasterisation will be get a huge jump to no one's surprise.