Nvidia unveils Geforce RTX 4000 and ‘Ada Lovelace’ architecture

The annual event GTC is usually a platform where Nvidia offers demonstrations and news related to one of their main areas, namely artificial intelligence (AI) and the hardware to drive it. When Jensen Huang enters the digital scene, it is clear that this year’s edition is no exception. However, Nvidia isn’t just appeasing professional users, but is topping the bill with the next consumer-oriented Geforce series.

In accordance with previous information, Nvidia is letting large parts of the two-year-old Geforce RTX 3000 series live on, while the RTX 4000 defines a new top tier with the two models RTX 4090 and RTX 4080. Under the skin of the fresh duo is the graphics architecture “Ada Lovelace “, which is manufactured on TSMC’s 4-nanometer technology and puts the features under the RTX umbrella in focus.

Lovelace.jpg
geforce-rtx-4090-100vp-l@2x.jpg

Specification-poor uncovering of worst-case cards

Geforce RTX 4090 a worst-case model with 24 GB of GDDR6 memory and 76 billion transistors, which translates to a whopping 16,384 CUDA cores with a turbo frequency of 2.52 GHz. In terms of performance, Nvidia compares the successor to the current top graphics card RTX 3090 Ti and believes that the raster performance is around twice better with the RTX 4090. When ray tracing is on the menu, the performance step can be turned up to four times higher.

Specifications Geforce RTX 4090, RTX 4080 16 GB and RTX 4080 12 GB

RTX 4090

RTX 3090 Ti

RTX 4080 16 GB

RTX 4080 12 GB

RTX 3080 Ti

Technique

4 nm TSMC

8 nm Samsung

4 nm TSMC

4 nm TSMC

8 nm Samsung

Circuit

AD102?

GA102

AD103?

AD104?

GA102

Circuit area

?

628 mm²

?

?

628 mm²

Transistors

76 billion

28.3 billion

?

?

28.3 billion

Architecture

There’s Lovelace

Ampere

There’s Lovelace

There’s Lovelace

Ampere

CUDA kernels

16 384 st.

10 752 st.

9 728 st.

7 680 st.

10 240 st.

RT kernels

128 st.?

84 st.

76 st. ?

60 st. ?

80 st.

Tensor kernels

512 st.?

336 st.

304 st.?

240 st.?

320 st.

Texture units

512 st.?

336 st.

304 st.?

240 st.?

320 st.

Raster units

?

112 st.

?

?

112 st.

Clock frequency

2 230 MHz

1 560 MHz

2 210 MHz

2 310 MHz

1 365 MHz

GPU Boost

2 520 MHz

1 860 MHz

2 510 MHz

2 610 MHz

1 665 MHz

Computing power

82 575 GFLOPS

39 997 GFLOPS

48 835 GFLOPS

40 090 GFLOPS

34 099 GFLOPS

Amount of memory

24 GB GDDR6X

24 GB GDDR6X

16 GB GDDR6X

12 GB GDDR6X

12 GB GDDR6X

Memory frequency

?

21 000 MHz

?

?

19 000 MHz

Memory bus

384-bit

384-bit

256-bit

192-bit

384-bit

Minnesbandbredd

?

1008 GB/s

?

?

912 GB/s

Power supply

12VHPWR1×12-pin

12-pin

12VHPWR

12VHPWR

12-pin

SLI connection

No

NVLink 3.0 x4

No

No

TBP

450 W

450 W

320 W

285 W

350 W

Launch price

1 599 USD

1 999 USD

1 199 USD

899 USD

1 199 USD

In advance, high power output has always been on the cards, but when the official debut now takes place, the concerned can now breathe a sigh of relief – 600 and 800 watts are missing. It is indeed possible to overclock the Geforce RTX 4090 north of 3 GHz and thus a heavy power draw, but in the standard version it is about 450 watts according to the specification. Assuming Nvidia’s claim about the performance gains is true, so is “Ada Lovelace” an energy efficient story – up to twice as energy efficient.

Screenshot (85).png
4090.jpg

Geforce RTX 4080 and Geforce RTX 4080 will be significantly different radar pairs

For a moment, Nvidia also shows what awaits a step down in the “Ada Lovelace” signed top segment, namely the Geforce RTX 4080 with 12 or 16 GB of GDDR6X memory. The graphics cards are compared against the RTX 3080 Ti and here too a performance step of two to four times is nailed. Taking a closer look at the spec sheet, it’s clear that this comparison likely only applies to the more well-equipped card.

4080.jpg

Nvidia is doing its best to make it difficult for customers. The Geforce RTX 4080 model with 16 GB of graphics memory has 27 percent more CUDA cores – 9,728 compared to the 12 GB model’s 7,680 CUDA cores. However, the turbo frequency is slightly higher for the latter, specifically 2.61 GHz instead of 2.51 GHz. Thus, the power budget for the models is also different: 320 and 285 watts respectively. The connector used is 12VHPWR, or supplied adapter.

Unchanged radiators and high prices

Judging from the presentation, Nvidia is breaking the two-generation-old tradition of changing the cooling solution for the reference graphics cards. The Founder’s Edition variants shown on the broadcast appear to have the same design as the ‘Ampere’ siblings, but whether they will be available in Europe this generation remains to be seen. It can be added that Nvidia does not offer a Founder’s Edition variant of the simpler Geforce RTX 4080 version.

The top model Geforce RTX 4090 will be the first out of the two graphics cards and the day to mark in the calendar is October 12. The target price is 1,599 USD, but with the current exchange rate, the Swedish price indicated via Nvidia’s website is not as pleasing, with a starting price of 21,590 kroner. The RTX 4080 duo will debut sometime in November, with target prices from SEK 12,200 and SEK 16,199.

Screenshot (86).png
Screenshot (98).png

For those who usually follow hardware launches, it is clear that Nvidia is not opening all the floodgates when it comes to specifications. More details about the “Ada Lovelace” architecture and the specifications of the unveiled graphics cards will appear in the near future, once the commotion surrounding the GTC presentation has subsided.

Smarter ray tracing and extra frames keywords for “Ada Lovelace”

RTX is a label that covers Nvidia’s capabilities for both machine learning and ray tracing. On ray tracingpage, the company highlights that the RT cores are getting new specific hardware to handle and accelerate specific functions. To further speed up the workflow, Nvidia equips the SM clusters of “Ada Lovelace” with something they call Shader-Execution Reordering (TO BE).

Graphics cards typically work with tasks that are easy to parallelize, but when ray tracing mixed into the game, this disappears. The light rays bounce against different materials and different calculation time and memory access are given parts of such loads. According to Nvidia, SER is to be equated with that of the processor world Out-of-order Execution. Graphics card’s different ray tracing-operations can thus be sorted in a way that makes better use of the hardware, which is claimed to give a performance boost of two to three times.

Furthermore, Nvidia emphasizes that the machine learning-oriented Tensor cores are also getting a boost, but the company is mainly focusing on DLSS 3. The acronym stands for Deep-Learning Super Sampling and has so far used machine learning to scale up low-resolution images to a target resolution in real time. With DLSS 3, upscaling can be set aside to interpolate intermediate frames to reach a higher frame rate, something that becomes exclusively available with the updated Tensor cores of the “Ada Lovelace” family.

The technology takes advantage of machine learning to be able to predict what the intermediate images will look like, where information about how and in which direction pixels move from image to image. Nvidia describes the fact that DLSS is not part of the typical rendering flow, because then neither the graphics card nor the processor becomes a limiting factor. The company gives a taste of how the function works in Microsoft Flight Simulator 2020, where the frame rate with ray tracing rising from barely 50 to over 100 FPS.


Source: SweClockers by www.sweclockers.com.

*The article has been translated based on the content of SweClockers by www.sweclockers.com. If there is any problem regarding the content, copyright, please leave a report below the article. We will try to process as quickly as possible to protect the rights of the author. Thank you very much!

*We just want readers to access information more quickly and easily with other multilingual content, instead of information only available in a certain language.

*We always respect the copyright of the content of the author and always include the original link of the source article.If the author disagrees, just leave the report below the article, the article will be edited or deleted at the request of the author. Thanks very much! Best regards!