NVIDIA details GeForce RTX 30 Ampere architecture

Published: 4th Sep 2020, 21:46 GMT   Comments

A more detailed look at the Ampere GeForce architecture has been provided by NVIDIA. The company released block diagrams, technology descriptions, and detailed specifications of the next-generation GeForce graphics cards. 

NVIDIA details Ampere architecture

NVIDIA has confirmed the expected performance figures for its new RTX Ampere models. These are official figures based on NVIDIA in-house testing, likely to be taken with a grain of salt. However, the slides do show what the company expects from the upcoming GeForce 30 stack and where should it settle in the comparison charts.

Official NVIDIA RTX 30 Performance (according to NVIDIA)

NVIDIA also revealed detailed specs of the Ampere GPUs. The GPUs build using the Samsung 8nm node have a transistor destiny of 44.6MT/mm2 (GA102) compared to 24.7 MT/mm2 for Turing (TU102). The biggest Ampere RTX GPU – GA102 has 28 billion transistors and a die size of 628.4 mm2.  This GPU will be used for GeForce RTX 3090 and RTX 3080. The flagship model features 82 Streaming Multiprocessors with 328 Tensor cores and 82 RT cores (ray tracing cores). It also has a 384-bit memory bus and 6MB of L2 cache.

A smaller Ampere GPU, GA104, features 17.4 billion transistors and a die size of 392 mm2. The RTX 3070 graphics card has 46 SMs enabled and thus the number of Tensors is 184 and RT cores is 46. This GPU has a memory bus of 256-bit and 4MB of L2 cache.

NVIDIA RTX 3080 (GA102) Block Diagram

NVIDIA Ampere Streaming Multiprocessor

NVIDIA GeForce RTX 30 Series Specifications
VideoCardz.comRTX 3090RTX 3080RTX 3070RTX 2080 Ti
Picture
BoardPG132 SKU 30PG132 SKU 10PG142 SKU 10PG150 SKU 32
GPU8nm GA102-3008nm GA102-2008nm GA104-30012nm TU102-300
Die Size
 
628 mm2
 
628 mm2
 
392 mm2
 
754 mm2
Transistors
 
28 B
 
28 B
 
17.4 B
 
18.6 B
CUDA Cores
 
10496
 
8704
 
5888
 
4352
Tensor Cores
 
328 (4 per SM)
 
272 (4 per SM)
 
184 (4 per SM)
 
544 (8 per SM)
RT Cores
 
82
 
68
 
46
 
68
Base Clock
 
1395 MHz
 
1440 MHz
 
1500 MHz
 
1350 MHz
Boost Clock
 
1695 MHz
 
1710 MHz
 
1725 MHz
 
1545 MHz
Shader Perf.
 
35.6 TFLOPS
 
29.8 TFLOPS
 
20.3 TFLOPS
 
13.4 TFLOPS
Tensor Perf.
 
285 TFLOPS
 
238 TFLOPS
 
163 TFLOPS
 
110 TFLOPS
Memory
 
24GB G6X
 
10GB G6X
 
8GB G6
 
11GB G6
Memory Clock
 
19.5 Gbps
 
19 Gbps
 
14 Gbps
 
14 Gbps
Memory Bus
 
384-bit
 
320-bit
 
256-bit
 
352-bit
Bandwidth
 
936 GB/s
 
760 GB/s
 
448 GB/s
 
616 GB/s
TDP
 
350W
 
320W
 
220W
 
250W
MSRP
 
$1499
 
$699
 
$499
 
$999

RT Cores

NVIDIA GeForce RTX 30 series introduces Second Generation RT cores, further improving ray tracing acceleration. NVIDIA RT cores have pure hardware-based bounding volume hierarchy (BVH) which greatly improves performance over traditional and minimalistic approach with SIMD stream processors.

The second-generation Ampere RT core adds a triangle interpolation component along a time-scale, in coordination with the triangle intersection unit to the RT core architecture. According to NVIDIA, this should be useful during motion blur effects in real-time ray tracing.

Tensor Cores

A third-generation Tensor core largely copies the design from (formerly known as Tesla) A100 Accelerator. This core will boost the performance during DLSS AI-super resolution upscaling in gaming.

Ampere’s tensors are designed to leverage sparsity in deep learning neural nets. This is a process of reducing the matrixes without affecting its accuracy. This process can improve AI performance by an order of magnitude.

NVIDIA GA102 PCB

NVIDIA released a first official picture of the GA102-based Founders Edition graphics card featuring a PCI-Express 4.0 connector. This particular board design is for the GeForce RTX 3080. Unlike the RTX 3090, this board design has two memory modules missing and it lacks the NVLink connector (which is exclusive to RTX 3090 models). A total of 12 GDDR6X modules can be mounted on this board, but the RTX 3080 only has 10.

The PCB features 20-phase VRM which are put on both sides of the GPU. The DrMOS and tantalum capacitors were mostly put on the rear of the board.

The RTX 30 series biggest innovation is the new Molex Microfit 12-pin power connector, capable of delivering up to 300W of power. NVIDIA will not restrict Intel or AMD from using this connector, so there is a change that we see an adoption of the new standard pretty soon.

NVIDIA GeForce RTX 3080 PCB

HDMI 2.1 and 8k GAMING

The RTX 3090 is the first graphics card advertised by NVIDIA for 8K gaming. This is possible thanks to the NVIDIA’s DLSS2 technology, which artificially upscales the resolution to 8k using 3rd Gen Tensor cores.

The 8K resolution has 16x more pixels than FullHD. This resolution has not really seen its prime popularity yet, as the availability of proper displays is still very limited. However, the addition of the HDMI 2.1 connector to the RTX 30 series (and also upcoming Xbox Series X and Playstation 5) are a good starting point for technology adoption in the near future.

Source: TechPowerUP, HotHardware




Comment Policy
  • Comments must be written in English.
  • Comments deemed to be spam or solely promotional in nature will be deleted. Including a link to relevant content is permitted, but comments should be relevant to the post topic.
  • Comments containing language or concepts that could be deemed offensive will be deleted. Note this may include abusive, threatening, pornographic, offensive, misleading or libelous language.
  • A failure to comply with these rules will result in a warning and, in extreme cases, a ban.
  • Please note that comments that attack or harass an individual directly will be deleted and such comments will result in a ban.
  • VideoCardz Moderating Team reserves the right to edit or delete any comments submitted to the site without notice.
  • If you have any questions about the commenting policy, please let us know through the Contact Page.
Hide Comment Policy
Comments

This website relies on third-party cookies for advertisement, comments and social media integration. Check our Privacy and Cookie Policy for details.