• Competitor rules

    Please remember that any mention of competitors, hinting at competitors or offering to provide details of competitors will result in an account suspension. The full rules can be found under the 'Terms and Rules' link in the bottom right corner of your screen. Just don't mention competitors in any way, shape or form and you'll be OK.

NVIDIA ‘Ampere’ 8nm Graphics Cards

Caporegime
OP
Joined
8 Jul 2003
Posts
30,062
Location
In a house
Nvidia is slowly but steadily preparing the launch of their ampere generation GPUs, allegedly starting March 2020, the first models will be out for data-centers. Geforce 3000 for desktop PCs would appear in June. This is reported by the Chinese website HKEPC, citing Chris Caso, an analyst with Raymond James.

Accordingly, Nvidia's in-house GTC (GPU Technology Conference, March 22-26, 2020) will initially feature data center GPUs, such as ampere-based Tesla accelerator cards with a focus on scientific computation and artificial intelligence algorithms.

This means around Computex in June 2020 the first consumer variants with Ampere chips would see the light of day. The launch would involve the high-end cards first and slower models. Pascal was manufactured in the 16nm FinFET Variation, Turing is produced in 12nm FinFET, Ampere is said to be manufactured at Samsung on 7nm - using its Extreme Ultraviolet (EUV) method 7nm+. In comparison to Turing, the clock rates could be increased by 200 to 300 MHz. 7nm makes transistors smaller, and that allows for more transistors to be placed on a chip. It is expected that Ampere will offer significantly more Raytracing cores than Turing did.

https://www.guru3d.com/news_story/geforce_rtx_3000_ampere_data_center_marchconsumers_june_2020.html
 
Soldato
Joined
6 Feb 2019
Posts
17,566
A whole bunch of tech stats dropped today - take with salt, though it's a lot of data

  • Nvidia GA100 (HPC)

  • HPC / professional solution of the Ampere generation
  • 6144 Bit HBM2 interface

  • 8 Graphics Processing Cluster (GPC) with each 16 Streaming Multiprocessors (SM) and each 8 Texture Processing Clusters (TPC)

  • each SM contains 64 CUDA cores (CU) with 4 Texture Mapping Units (TMU), but 16 Tensor cores

  • Total of 128 Streaming Multiprocessors, which contain a total of 8192 CUDA cores, 512 TMUs and 2048 Tensor cores

  • PCI Express 4.0

  • 7nm+ (EUV) from TSMC

  • Launch first half 2020

  • Nvidia Ampere Gaming

  • 5 gaming chips, all at Samsung 7nm (with EUV)
https://www.forum-3dcenter.org/vbulletin/showthread.php?p=12138675#post12138675

While not mentioned, if the memory bit rate and type is true it translates to 48GB of memory
 
Last edited:
Man of Honour
Joined
13 Oct 2006
Posts
91,058
While not mentioned, if the memory bit rate and type is true it translates to 48GB of memory

That would be the HPC boards though - gaming cards would be quite a bit less - nVidia have kind of said they are going to significantly increase VRAM amounts next generation though.
 
Soldato
Joined
19 Jan 2010
Posts
4,806
A whole bunch of tech stats dropped today - take with salt, though it's a lot of data

  • Nvidia Ampere HPC

  • GA100 chip with a 6144 Bit HBM2 interface, doubled Tensor cores per SM
  • TSMC 7nm+ (with EUV)
  • Nvidia Ampere Gaming

  • 5 gaming chips, all at Samsung 7nm (with EUV)
  • Nvidia GA100

  • HPC / professional solution of the Ampere generation

  • 8 Graphics Processing Cluster (GPC) with each 16 Streaming Multiprocessors (SM) and each 8 Texture Processing Clusters (TPC)

  • each SM contains 64 CUDA cores (CU) with 4 Texture Mapping Units (TMU), but 16 Tensor cores

  • Total of 128 Streaming Multiprocessors, which contain a total of 8192 CUDA cores, 512 TMUs and 2048 Tensor cores

  • PCI Express 4.0

  • 7nm+ (EUV) from TSMC

  • Launch first half 2020
https://www.forum-3dcenter.org/vbulletin/showthread.php?p=12138675#post12138675

While not mentioned, if the memory bit rate and type is true it translates to 48GB of memory
Is it really going to use the pci 4.0 bandwidth or are we just looking at "extra special writeup goodness" here?
 

bru

bru

Soldato
Joined
21 Oct 2002
Posts
7,360
Location
kent
A whole bunch of tech stats dropped today - take with salt, though it's a lot of data

  • Nvidia GA100 (HPC)

  • HPC / professional solution of the Ampere generation
  • 6144 Bit HBM2 interface

  • 8 Graphics Processing Cluster (GPC) with each 16 Streaming Multiprocessors (SM) and each 8 Texture Processing Clusters (TPC)

  • each SM contains 64 CUDA cores (CU) with 4 Texture Mapping Units (TMU), but 16 Tensor cores

  • Total of 128 Streaming Multiprocessors, which contain a total of 8192 CUDA cores, 512 TMUs and 2048 Tensor cores

  • PCI Express 4.0

  • 7nm+ (EUV) from TSMC

  • Launch first half 2020
  • Nvidia Ampere Gaming

  • 5 gaming chips, all at Samsung 7nm (with EUV)
https://www.forum-3dcenter.org/vbulletin/showthread.php?p=12138675#post12138675

While not mentioned, if the memory bit rate and type is true it translates to 48GB of memory


Well that shows us what sort of size and core count the mainstream cards could have. potentially 8192 cuda cores on the big boy Titan card would be very powerful

In theory a 3070 could be looking at about 4500 cuda cores which would be between a Titan RTX and 2080ti in core count.

This could be monstrous in performance, without taking into account any improvements in the architecture itself.

Ok now I'm definitely looking forward to it.:)
 
Soldato
Joined
6 Feb 2019
Posts
17,566
Well that shows us what sort of size and core count the mainstream cards could have. potentially 8192 cuda cores on the big boy Titan card would be very powerful

In theory a 3070 could be looking at about 4500 cuda cores which would be between a Titan RTX and 2080ti in core count.

This could be monstrous in performance, without taking into account any improvements in the architecture itself.

Ok now I'm definitely looking forward to it.:)

haha definitely. I didnt post it, but there was also some guestimation for the gaming cards based on the HPC specs and you're in the right range with the 3080 estimated to be around 6000-6500 cuda cores.
Due to the bump in tensor and assuming RT cores gets the same type of bump - it was also estimated ray tracing could be up to 3 times faster than the RTX2000 cards when comparing something like a theoretical Ampere Titan to the 2080ti.

As long as prices don't go up for the 3070, then getting 2080ti performance raster performance and increased ray tracing performance for 2070 prices ($500usd) would be very good.
 
Permabanned
Joined
2 Sep 2017
Posts
10,490
A whole bunch of tech stats dropped today - take with salt, though it's a lot of data

  • Nvidia GA100 (HPC)

  • HPC / professional solution of the Ampere generation
  • 6144 Bit HBM2 interface

  • 8 Graphics Processing Cluster (GPC) with each 16 Streaming Multiprocessors (SM) and each 8 Texture Processing Clusters (TPC)

  • each SM contains 64 CUDA cores (CU) with 4 Texture Mapping Units (TMU), but 16 Tensor cores

  • Total of 128 Streaming Multiprocessors, which contain a total of 8192 CUDA cores, 512 TMUs and 2048 Tensor cores

  • PCI Express 4.0

  • 7nm+ (EUV) from TSMC

  • Launch first half 2020
  • Nvidia Ampere Gaming

  • 5 gaming chips, all at Samsung 7nm (with EUV)
https://www.forum-3dcenter.org/vbulletin/showthread.php?p=12138675#post12138675

While not mentioned, if the memory bit rate and type is true it translates to 48GB of memory

What is really written there is:

"I'm trying to recapitulate his deleted posts to Ampere:
Since he had posted that HPC Ampere consists of 2 chips:

GA100 with 6 HBM stacks, 6072Bit, double the number of tensor cores per SM.
GA101 with 3 HBM stacks, half GA100, was canceled according to him.

The further info he had marked as not safe, but I'll post it anyway for discussion:
GA100 8GPCx8TPC = 8192 shaders
GA101 and GA101 at TSMC with EUV

Gaming were five chips completely at Samsung with EUV according to him, but also uncertain.

Now, Samsung really does not know what their EUV process really is. But after David Schor posted the other day that 1 main customer is now turning his back on Samsung (Can only be Nvidia or QC, otherwise there are not any big ones anymore, tap on QC) it does not sound good. Ne 2080 TI Super in Q1 would definitely indicate a shift in amps and I guess that could be due to Samsung's process.

Oh, and finally, the ampere successor is called "Hopper" according to him and is no longer monolithic in the HPC area."

https://www.forum-3dcenter.org/vbulletin/showthread.php?p=12138675#post12138675
 
Back
Top Bottom