NVIDIA Announces GeForce RTX 40 Series Up to 4X Faster with New Ada Lovelace Architecture.

https://news.mixi.jp/view_news.pl?id=7118511&media_id=38

Sep 21, 2022 02:31 ITmedia PC USER

View Tweets( 7 )Read Diary( 1 )

Check tweet write diary share on facebook

ITmedia PC USER

PhotoCEO Jensen Huang holding the GeForce RTX 4090 Founders Edition, a graphics card of his own design.

CEO Jens Huang holds the GeForce RTX 4090 Founders Edition, a graphics card designed by NVIDIA.

　NVIDIA announced the new GeForce RTX 4090 and GeForce RTX 4080 GPUs on September 21 (US Pacific Daylight Time). The GeForce RTX 4090 and GeForce RTX 4080 are scheduled to be released in the U.S. on October 12 and November, respectively, and the estimated retail prices of graphics cards equipped with each GPU are as follows.

[more images]

GeForce RTX 4090: Starting at $1599 (approx. 229,800 yen)

GeForce RTX 4080 (16GB model): starting at $1199 (approximately 172,300 yen)

GeForce RTX 4080 (12GB model): starting at $899 (approx. ¥129,200)

 

Ada Lovelace, a new GPU core

　The GeForce RTX 40 series is the first GPU to adopt the new GPU architecture "Ada Lovelace.

　The manufacturing process is Taiwan TSMC's 4nm, with up to 76 billion transistors integrated. The graphics memory is integrated with Micron's GDDR6X standard.

　Up to 90 shaders (SM: Streaming Multiprocessor) and up to 90 trillion FLOPS of computing power will be used. Sharder Execution Reordering (SER), a proprietary technology that dynamically organizes data before it is input to the shaders, improves shader performance (based on power consumption) by up to 2x and game frame rates by up to 25% compared to the previous generation Ampere architecture. and game frame rates by up to 25% compared to the previous generation Ampere architecture.

　The RT Core, the "heart" of the GeForce RTX series, is now in its third generation. The number of cores is up to 200, the computing power is up to 200 trillion FLOPS (up to 2.8 times more than Ampere), and the throughput of Ray-Triangle Intersection (judging the intersection of rays and triangles) is up to twice that of Ampere. This enables faster real-time ray tracing and more immersive 3D graphics.

　The 3rd generation RT core also includes the new "Opacity Micro-Map Engine" and "Micro-Mesh Engine," which assist in the tasks required for RT processing and are also very useful in realizing the "DLSS 3 The "Micro-Mesh Engine" and the "Micro-Mesh Engine" are very useful in realizing the "DLSS 3" described below.

　The "Tensor core" used for machine learning operations is now in its fourth generation. The number of cores is up to 1,400, and by including an FP8 (8-bit floating-point arithmetic) engine, the effective computing speed has been increased up to five times faster than that of Ampere.

Super-resolution technology "NVIDIA DLSS" is now in its third generation.

　With the Ada Lovelace architecture GPU cores, NVIDIA's super-resolution technology, NVIDIA DLSS (Deep Learning Super Sampling), has evolved to the third generation (DLSS 3).

　GPUs with the Ada Lovelace architecture are equipped with an "Optical Flow Accelerator," which can be leveraged to increase frame rates at high quality and high speed. The effect may vary depending on the game, but it seems to be more effective than expected. 

Overview of GeForce RTX 4090/4080

　The GeForce RTX 40 series will consist of two types of chips: the top-of-the-line GeForce RTX 4090 and GeForce RTX 4080, the latter of which will have two different chips with different specifications.

　Both models are equipped with two 8th generation NVENCs (NVIDIA Encoders) capable of hardware-based encoding of AV1 codec videos, which can be used for live game play using OBS Studio and Discord (*1).

(*1) OBS Studio will be supported in October, and Discord in the second half of 2022.

　The main specifications of each GPU are as follows. The operating clock and power consumption are NVIDIA's "rated values.

GeForce RTX 4090

CUDA cores: 16,384

Operating clock: 2.23GHz to 2.52GHz

Graphics memory: 24GB (GDDR6X/384bit)

External video output: HDMI x 1 + DisplayPort x 3

Power consumption: 450W (GPU auxiliary power: 8-pin x 3 equivalent)

GeForce RTX 4080 (16GB memory model)

CUDA cores: 9728

Operating clock: 2.21GHz to 2.51GHz

Graphics memory: 16GB (GDDR6X/256bit)

External video output: HDMI x 1 + DisplayPort x 3

Power consumption: 320W (GPU auxiliary power: 8-pin × 3 equivalent)

GeForce RTX 4080 (12GB memory model)

CUDA cores: 7680

Operating clock: 2.31GHz to 2.6s1GHz

Graphics memory: 12GB (GDDR6X/192bit)

External video output: HDMI x 1 + DisplayPort x 3

Power consumption: 285W (GPU auxiliary power: 8-pin x 2 equivalent)

　The GeForce RTX 4090 and GeForce RTX 4080 (16GB model) will also have a "Founders Edition" graphics card designed by NVIDIA itself, but its availability in Japan has not yet been determined.