Rumor: NVIDIA's Next Generation GeForce RTX 3080 and RTX 3070 "Ampere" Graphics Cards Detailed

[AleksandarK]

NVIDIA's next-generation of graphics cards codenamed Ampere is set to arrive sometime this year, presumably around GTC 2020 which takes place on March 22nd. Before the CEO of NVIDIA, Jensen Huang officially reveals the specifications of these new GPUs, we have the latest round of rumors coming our way. According to VideoCardz, which cites multiple sources, the die configurations of the upcoming GeForce RTX 3070 and RTX 3080 have been detailed. Using the latest 7 nm manufacturing process from Samsung, this generation of NVIDIA GPU offers a big improvement from the previous generation.

For starters the two dies which have appeared have codenames like GA103 and GA104, standing for RTX 3080 and RTX 3070 respectively. Perhaps the biggest surprise is the Streaming Multiprocessor (SM) count. The smaller GA104 die has as much as 48 SMs, resulting in 3072 CUDA cores, while the bigger, oddly named, GA103 die has as much as 60 SMs that result in 3840 CUDA cores in total. These improvements in SM count should result in a notable performance increase across the board. Alongside the increase in SM count, there is also a new memory bus width. The smaller GA104 die that should end up in RTX 3070 uses a 256-bit memory bus allowing for 8/16 GB of GDDR6 memory, while its bigger brother, the GA103, has a 320-bit wide bus that allows the card to be configured with either 10 or 20 GB of GDDR6 memory. In the images below you can check out the alleged diagrams for yourself and see if this looks fake or not, however, it is recommended to take this rumor with a grain of salt.


View at TechPowerUp Main Site

 

[londiste]

Tl;dr
GA103 - 60 SM (3480 SP) - 320-bit VRAM (10/20 GB) - assumed 3080
GA104 - 48 SM (3072 SP) - 256-bit VRAM (8/16 GB) - assumed 3070

7nm has not been favourable to higher clocks so far but has been considerably better in power consumption and efficiency. With frequency of around 2GHz (and some architectural improvements) GA104 is faster than 2080 Super and GA103 faster than 2080Ti.

Assuming Nvidia's usual chip numeration there should be another higher-end GA102 in the works...

 

[kapone32]

Depending on pricing I may make the jump to Nvidia if this holds true. I cringe at the potential price point for these particular cards.

 

[Deleted member 177333]

Depending on pricing I may make the jump to Nvidia if this holds true. I cringe at the potential price point for these particular cards.

Ya, I will wait and see what the 3080 Ti is priced at when released. It needs to have at bare minimum, 16GB of RAM or more. If they can get it priced appropriately, I'll buy a pair of them. If not, I will continue to be patient and wait for a good deal on ebay down the road.

 

[xkm1948]

I do wonder how much different the overall design would be from Turing uArc. Also I hope they don't cut down the Tensorflow units. It has been really nice to use consumer level GPU for dl/ml acceleration.

 

[CrAsHnBuRnXp]

Ya, I will wait and see what the 3080 Ti is priced at when released. It needs to have at bare minimum, 16GB of RAM or more. If they can get it priced appropriately, I'll buy a pair of them. If not, I will continue to be patient and wait for a good deal on ebay down the road.

Why? Dual card systems are basically dead these days. Yes, there is still support for SLI but it's at the point where its unneeded and troublesome.

 

[Otonel88]

Not really an expert on the depth details, but currently the 2080 TI has 4352 Cuda cores, and b, and the rumored GA103 die that could have potentially 3840 Cuda cores and as much as 60 SM's.
So on paper this is not faster than the 2080TI. Right?
Obviously the smaller die would result in other benefit, such as lower TDP which I am really interested, and other stuff.
Do you guys think the GA103 card would be under 200W TDP?
Always looking at lower TDP cards, which are more silent as there is less heath to deal with by the cooler which translates into quieter card operation.

 

[xkm1948]

Not really an expert on the depth details, but currently the 2080 TI has 4352 Cuda cores, and b, and the rumored GA103 die that could have potentially 3840 Cuda cores and as much as 60 SM's.
So on paper this is not faster than the 2080TI. Right?
Obviously the smaller die would result in other benefit, such as lower TDP which I am really interested, and other stuff.
Do you guys think the GA103 card would be under 200W TDP?
Always looking at lower TDP cards, which are more silent as there is less heath to deal with by the cooler which translates into quieter card operation.

Different generation of stream processors cannot be compared just by numbers.

 

[Otonel88]

Different generation of stream processors cannot be compared just by numbers.

Good point

 

[kapone32]

Why? Dual card systems are basically dead these days. Yes, there is still support for SLI but it's at the point where its unneeded and troublesome.

Multi GPU is for anyone that wants it. I have not replaced my Vega 64s with 5700Xts for exactly that reason. If Multi GPU was really dead MBs would not be touting it and giving you 2, 3 and 4 SLI bridges. As much as people complain about it being troublesome it is not as bad as people make it out to be. There are plenty of Games that support Multi GPU anyway. As an example I get 107 FPS average at 4K playing Jedi Fallen Order @ Ultra settings.

 

[CrAsHnBuRnXp]

Multi GPU is for anyone that wants it. I have not replaced my Vega 64s with 5700Xts for exactly that reason. If Multi GPU was really dead MBs would not be touting it and giving you 2, 3 and 4 SLI bridges. As much as people complain about it being troublesome it is not as bad as people make it out to be. There are plenty of Games that support Multi GPU anyway. As an example I get 107 FPS average at 4K playing Jedi Fallen Order @ Ultra settings.

Triple and quad SLI is officially dead as nvidia no longer supports it.

 

[Otonel88]

Any opinions or guesses on the TDP of this new cards using the 7nm?

 

[Kaotik]

This "new leak" is actually part of rumor from last may

https://twitter.com/i/web/status/1131031428878094337

 

[kapone32]

Triple and quad SLI is officially dead as nvidia no longer supports it.

Yeah I know, I was just referencing that we still get those. Personally I would never go for more than 2 GPUs in a setup. It would be nearly impossible (unless you install single slot water blocks) to fit more than in some cases, 2 GPUS on modern Motherboards.

 

[Adam Krazispeed]

IDK, my thoughts..

AM102 - ?? sm (????? SP) -RT Cores 108 - 384bit VRAM - assumed RTX Titan XXX
AM104 - 68 SM (4352 SP) - RT Cores 88 - 320-bit VRAM - assumed 3080ti -12gb-16gb-32gb
AM106 - 46 SM (2944 SP) - RT Cores 68 - 256-bit VRAM - assumed 3070ti - 12gb-16gb

7nm has not had much higher clocks so ?? but has been much better in power consumption/Eff.. With frequency of 1.7 - 2 GHz GA104 Could be faster than a 2080 Super and GA102 will be faster than 2080Ti. 20% >>??? 30% MAX , but RT-RT up to 50% MAX

Assuming Nvidia's usual chip numeration there should be another higher-end GA102 in the works...TITAN model!!

IDK, my thoughts..

AM102 - ?? sm (????? SP) -RT Cores 108 - 384bit VRAM - assumed RTX Titan XXX
AM104 - 68 SM (4352 SP) - RT Cores 88 - 320-bit VRAM - assumed 3080ti -12gb-16gb-32gb
AM106 - 46 SM (2944 SP) - RT Cores 68 - 256-bit VRAM - assumed 3070ti - 12gb-16gb

7nm has not had much higher clocks so ?? but has been much better in power consumption/Eff.. With frequency of 1.7 - 2 GHz GA104 Could be faster than a 2080 Super and GA102 will be faster than 2080Ti. 20% >>??? 30% MAX , but RT-RT up to 50% MAX

Assuming Nvidia's usual chip numeration there should be another higher-end GA102 in the works...TITAN model!!

depens on also who nvidia using?? TSMC. or SAMSUNG... rumored that TSMC 7nm is a bit better than samsungs 7nm at the moment?? (whether EUV or DUV)

 

[medi01]

https://twitter.com/i/web/status/1219171309298561024

 

[VrOtk]

That 320-bit bus looks weird: allowing your xx80 grade card to have such high memory bandwidth, you start to cripple your xx80Ti's performance at higher resolutions (unless it uses 512-bit bus or HBM2).
Though I'd be happy to be wrong, as better grade products for the same or cheaper price is always a welcome.

 

[Space Lynx]

https://twitter.com/i/web/status/1219171309298561024

going to be a great year for high end gaming.

I'm leaning towards big navi and ryzen 4700x, but if intel/nvidia high end ends up being 30+% fast for a 200-300 bucks more... than yeah I prob will go green.

 

[efikkan]

GTC is a compute/professional graphics conference. I can't remember the last time Nvidia launched consumer products there, so this sounds unlikely but not impossible. Keep in mind that we've had similar rumors about GTC previous years too.
But I would expect them to show data center products and perhaps give some hints about the upcoming consumer products.

 

[Gungar]

That 320-bit bus looks weird: allowing your xx80 grade card to have such high memory bandwidth, you start to cripple your xx80Ti's performance at higher resolutions (unless it uses 512-bit bus or HBM2).
Though I'd be happy to be wrong, as better grade products for the same or cheaper price is always a welcome.

Yeah i was thinking the same but it really wouldn't surprise me that the new 3080Ti comes with HBM.

 

[EarthDog]

Multi GPU is for anyone that wants it.

True... for anyone that is a masochist . I mean if people want to put themselves through spotty/inconsistent performance uptick while always using 2x power, 2x heat mitigation... go for it. mGPU just needs to die already or go all in and get it working. Far too long has this technology just been 'MEH'.

it really wouldn't surprise me that the new 3080Ti comes with HBM.

I don't see this happening...

Does anyone else see HBM making any headway in consumer graphics? When first released, AMD touted it as the second coming... but it turns out, it was fairly useless comparatively (I mean it did shrink gaps between NV cards at high resolutions, but never had the core horsepower in the first place for anything past 1440p).

 

[_Flare]

They could go for 4SMs per TPC leading to doubled RT-Perf.
or they could place 8 or even 12 TPC per GPC, also leading to 2x Perf.
People shouldn´t forget, nvidia gets about 2x the area 12nm vs 7nm.
I bet nvidia has plenty of ideas to get performance from that extra-area and doubled RT Perf is scaled 1:1 with SMs.

 

[DeathtoGnomes]

Wonder what AMD will counter this with.

 

[Gungar]

Does anyone else see HBM making any headway in consumer graphics? When first released, AMD touted it as the second coming... but it turns out, it was fairly useless comparatively (I mean it did shrink gaps between NV cards at high resolutions, but never had the core horsepower in the first place for anything past 1440p).

HBM is by far the most powerful and efficient graphic memory out there right now. AMD is just incapable of producing any good gpu, for sometime now.

 

[EarthDog]

Wonder what AMD will counter this with.

Big Navi is due out at the same time...

... wondering if it will still be 'big' compared to these (or w.e NV's high-end will be) or what...