Question 'Ampere'/Next-gen gaming uarch speculation thread

[JujuFish]

1050 Ti had 4 GB.

The 700 series had this all over the place. Nearly every model had 2 memory configurations, with the higher one being more memory than the next tier up GPU. Heck, the 750Ti had a 4GB model, more than the 3GB the 780Ti had.

 

[ozzy702]

1050 Ti had 4 GB.

I thought about that too but the 1060 6GB was released first and the 3GB version was also cut down so different than the current situation.

 

[guidryp]

I thought about that too but the 1060 6GB was released first and the 3GB version was also cut down so different than the current situation.

There is no current situation either.

 

[Capt Caveman]

Any thoughts on how long a 3090 will remain as the top Nvidia gaming card? Will we see a refresh next year? Or will it be two years? Can't remember what the life cycle history is for each generation or Nvidia's road map.

Have an EVGA 3090 FTW3 Ultra coming in two weeks and still trying to justify it when for $1k less, 3080 is only 10% slower.

 

[jpiniero]

Any thoughts on how long a 3090 will remain as the top Nvidia gaming card? Will we see a refresh next year? Or will it be two years? Can't remember what the life cycle history is for each generation or Nvidia's road map.

Have an EVGA 3090 FTW3 Ultra coming in two weeks and still trying to justify it when for $1k less, 3080 is only 10% slower.

A completely full GA102 gaming card is possible sometime in 2021 but that would be a minor update. The bigger deal would be the 3080 Ti which is basically a 3090 with "only" 20 GB for $999.

 

[Bouowmx]

Kopite talking about next-gen already:

https://twitter.com/i/web/status/1343462867811495937

12 GPC, 6 TPC per GPC, assuming the rest is the same as in Ampere, 2 SM per TPC, 128 cores per SM = 18432 cores. At 1.8 GHz, 66 TFLOPS.

I'm surprised there will be a 1.7x jump, when the density jump from Samsung 10/8 nm to 7/5 nm is less than 2x.

For such number of TFLOPS, HBM2 may be suitable. 4096-bit, 3.2 GT/s, 1.64 GB/s maintains about the same bandwidth-to-FLOPS ratio as in GA102.

 

[fleshconsumed]

Re: 3060 non-ti


Do I hear "mining card"? I think that I do!

Edit: With the 12GB VRAM buffer, it should be able to hash the Grin/CuckooWHATEVER algos that require the "larger" VRAM frame-buffers.

Memory bandwidth is still only 384gb/s compared to 448gb/s on 3060ti (which can also be overclocked higher). It'll be a decent mining card especially if it comes in at $300 level and is widely available, but nothing too special.

 

[CakeMonster]

Have an EVGA 3090 FTW3 Ultra coming in two weeks and still trying to justify it when for $1k less, 3080 is only 10% slower.

The 3090 doesn't show its theoretical potential unless you push it, meaning new games at 4K, or 1440p with absolutely all bells and whistles. At the former it does actually give you that ~20%.

 

[GodisanAtheist]

Kopite talking about next-gen already:

https://twitter.com/i/web/status/1343462867811495937

12 GPC, 6 TPC per GPC, assuming the rest is the same as in Ampere, 2 SM per TPC, 128 cores per SM = 18432 cores. At 1.8 GHz, 66 TFLOPS.

I'm surprised there will be a 1.7x jump, when the density jump from Samsung 10/8 nm to 7/5 nm is less than 2x.

For such number of TFLOPS, HBM2 may be suitable. 4096-bit, 3.2 GT/s, 1.64 GB/s maintains about the same bandwidth-to-FLOPS ratio as in GA102.

- JFC this is ridiculous. Can barely buy the current gen and only 2 months from launch and the rumor mill is starting up on... not the refresh... but the whole next gen?!

*Let us pray* Lord, please let there be a glut of Ampere and RDNA2 inventory that has to be heavily discounted due to a quick generational cycle that allows us peasants in the $200-$350 bracket to pick up huge upgrade increases. Lord, please keep the devil of perpetual scarcity and artifically limited supply shortages at bay lest it become the norm. Lord, please keep all natural disasters away from Taiwan or any part of the PC DIY supply chain. Amen.

 

[Shivansps]

It's strange for sure. NVIDIA got caught with their pants down, that's all there is to it.

More like: its 192-bit and 6GB is just not enoght these days.

 

[jpiniero]

Alleged NVIDIA GeForce RTX 3080, RTX 3070 and RTX 3060 Mobile GPU specifications emerge - VideoCardz.com

NVIDIA GeForce RTX 30 Mobile Yesterday the first notebook maker listed its gaming devices featuring GeForce RTX 30 Mobile series up for preorder. Since then the company has removed the listing, but our article is still up in case you are wondering what was revealed. NVIDIA will introduce three...

videocardz.com

Videocardz has what is possibly the final mobile specs. The 3080 is a full GA104 while the 3070 mobile is further cut down than the desktop version.

 

[Mopetar]

Alleged NVIDIA GeForce RTX 3080, RTX 3070 and RTX 3060 Mobile GPU specifications emerge - VideoCardz.com

NVIDIA GeForce RTX 30 Mobile Yesterday the first notebook maker listed its gaming devices featuring GeForce RTX 30 Mobile series up for preorder. Since then the company has removed the listing, but our article is still up in case you are wondering what was revealed. NVIDIA will introduce three...

videocardz.com

Videocardz has what is possibly the final mobile specs. The 3080 is a full GA104 while the 3070 mobile is further cut down than the desktop version.

We know that Ampere is actually pretty efficient when the clocks aren't pushed to the limits, so these should make for some great mobile chips. The funny part is that the mobile 3080 is listed as potentially having 16 GB VRAM. Even though it's a smaller bus and slower non-X GDDR6 memory, it's still humorous for the mobile part to have over 50% more VRAM than the desktop part.

 

[JoeRambo]

Ampere is fine i think, what is not fine is ridiculous amount of overvolt applied and how hungry GDDR6X turned out to be. Take a look at the following photo from Port Royal run.

24GB of RAM is 100W of power stock clocks. SRAM takes another 50-60w and that leaves peanuts for actual processing arrays. One can easily see that 20GB of this hot mess wont be far behind.

 

[uzzi38]

Let me leave this here:

NVIDIA GeForce RTX 3080 Mobile Specs

NVIDIA GA104, 1545 MHz, 6144 Cores, 192 TMUs, 96 ROPs, 8192 MB GDDR6, 1500 MHz, 256 bit

www.techpowerup.com

 

[jpiniero]

https://twitter.com/i/web/status/1344619563770527747

Possible two new SKUs coming, which they are calling the 3080 Super and 3070 Super. The 3080 Super may just be a 3080 with 12 GB and 384 bit.

 

[insertcarehere]

Ampere is fine i think, what is not fine is ridiculous amount of overvolt applied and how hungry GDDR6X turned out to be. Take a look at the following photo from Port Royal run.

24GB of RAM is 100W of power stock clocks. SRAM takes another 50-60w and that leaves peanuts for actual processing arrays. One can easily see that 20GB of this hot mess wont be far behind.

I do wonder how much GA102 may have been bottlenecked with 16gbps GDDR6, because that sure seems like a solution with less power consumption and allows for more VRAM, which would've made the whole "lower SKU has more VRAM than higher SKU" problem moot.

 

[tviceman]

https://twitter.com/i/web/status/1344619563770527747

Possible two new SKUs coming, which they are calling the 3080 Super and 3070 Super. The 3080 Super may just be a 3080 with 12 GB and 384 bit.

This guy has literally made nearly every plausible prediction possible since Ampere has come out. There are higher vram cards, no there isn't. There is a GA103, no there isn't. There are TI models, no there aren't. There are super models, no there aren't.

When a person makes this many contradicting opinions, it stands to reason he is now currently begging for attention OR he's being fed bad information.

 

[tviceman]

Everyone here trying to say Ampere's vram is a huge source of its power problems - that doesn't explain how much power is saved by undervolting the GPU and also the fact that the 3070 is only 5-6% more efficient than the 3080.

But also, has anyone ran benchmarks with the 3080 at 16gb/s vram? If GDDR 6X is a power hog (and costs more), it seems like GDDR6 running at 16gb/s with the full 384-bit bus would have been the proper trade off. Less power, 12gb of vram, 99.2% of the performance.

 

[DooKey]

Everyone here trying to say Ampere's vram is a huge source of its power problems - that doesn't explain how much power is saved by undervolting the GPU and also the fact that the 3070 is only 5-6% more efficient than the 3080.

But also, has anyone ran benchmarks with the 3080 at 16gb/s vram? If GDDR 6X is a power hog (and costs more), it seems like GDDR6 running at 16gb/s with the full 384-bit bus would have been the proper trade off. Less power, 12gb of vram, 99.2% of the performance.

NV likes to jump on this new stuff. Maybe this is just a test bed for the next Gen card?

 

[jpiniero]

But also, has anyone ran benchmarks with the 3080 at 16gb/s vram? If GDDR 6X is a power hog (and costs more), it seems like GDDR6 running at 16gb/s with the full 384-bit bus would have been the proper trade off. Less power, 12gb of vram, 99.2% of the performance.

I'm not sure that 12 GB GDDR6 would actually be cheaper or use less power than 10 GB GDDR6X. Esp when you factor in the consoles eating up a ton of the GDDR6 supply.

 

[Ajay]

This guy has literally made nearly every plausible prediction possible since Ampere has come out. There are higher vram cards, no there isn't. There is a GA103, no there isn't. There are TI models, no there aren't. There are super models, no there aren't.

When a person makes this many contradicting opinions, it stands to reason he is now currently begging for attention OR he's being fed bad information.

Yeah, he's been better in the past. Like you point out, could just be NV or AIBs are tossing around names left and right.

 

[JoeRambo]

But also, has anyone ran benchmarks with the 3080 at 16gb/s vram? If GDDR 6X is a power hog (and costs more), it seems like GDDR6 running at 16gb/s with the full 384-bit bus would have been the proper trade off. Less power, 12gb of vram, 99.2% of the performance.

Impossible to test, as Afterburner only allows a minor underclock on memory.

I did the following test of memory scaling in ANNO 1800 DX11 3440x1440 benchmark on fixed 3090 1815mhz clock.

18500 effective mem clock : 106.51 FPS
19500 effective mem clock : 108.14 FPS (stock)
20500 effective mem clock : 109.35 FPS

 

[exquisitechar]

This guy has literally made nearly every plausible prediction possible since Ampere has come out. There are higher vram cards, no there isn't. There is a GA103, no there isn't. There are TI models, no there aren't. There are super models, no there aren't.

When a person makes this many contradicting opinions, it stands to reason he is now currently begging for attention OR he's being fed bad information.

I don't think so. What he's posted is probably real and is/was in consideration for Nvidia, but not all of it will end up being released. He acknowledges that it's a mess himself, but I don't think it's because he's being fed false information.

https://twitter.com/i/web/status/1346356287903326209

I have no reason to doubt him given his perfect track record concerning Ampere leaks, long before launch.

 

[menhera]

Those "leakers" are the same. They change their statements every time. Kopite, he doesn't seem to be legit. He said 3080 Ti would be 384-bit 12GB and that it would be 320-bit 20GB later and then it could be a new GA103-based chip. Why would NVIDIA design another silicon with 6 GPCs when they already have GA102 with 7 GPCs?

He's now saying 3080 Ti could be GA102 384-bit 12GB, GA102 320-bit 20GB or GA103 320-bit 10GB. He literally couldn't be wrong since he's made every prediction. I don't trust those guys.

 

[jpiniero]

Those "leakers" are the same. They change their statements every time. Kopite, he doesn't seem to be legit. He said 3080 Ti would be 384-bit 12GB and that it would be 320-bit 20GB later and then it could be a new GA103-based chip. Why would NVIDIA design another silicon with 6 GPCs when they already have GA102 with 7 GPCs?

He's now saying 3080 Ti could be GA102 384-bit 12GB, GA102 320-bit 20GB or GA103 320-bit 10GB. He literally couldn't be wrong since he's made every prediction. I don't trust those guys.

nVidia pretty clearly changed the 3080 specs when they found out about what AMD was doing. That AMD didn't end up not really making many doesn't change that.