Asus GeForce RTX 5090 TUF Gaming OC Review: Power Without Compromise
Asus GeForce RTX 5090 TUF Gaming OC is a top model on the Blackwell architecture with 32 GB of GDDR7. The video card is equipped with an improved cooling system and is designed for high loads in 4K. In the review, we will consider its behavior at different values Power Target.
APPEARANCE AND CHARACTERISTICS |
This section of our review highlights the main aspects of the Asus GeForce RTX 5090 TUF Gaming OC, such as its appearance and hardware configuration.
APPEARANCE |
In front of us is the ASUS TUF Gaming branded box in black and green tones. The front side shows the video card itself, as well as the main characteristics: 32 GB GDDR7, support DLSS 4, Ray Tracing, Reflex and Studio, as well as the proprietary TUF cooling system. In the lower right corner you can see the NVIDIA logo and the model name. The box has a stylish minimalist design, without unnecessary elements, emphasizing the premium status of the video card.
When opening the package, we are greeted by a neatly laid video card in a protective film, securely fixed in a foam niche. In addition to the video card itself, there are also accessories inside: power cables, metal support to reduce stress on the PCIe slot, as well as ASUS documentation. All packaging is of a high standard, providing maximum protection for the device during transportation.
In the kit we see all the necessary accessories: TUF Gaming signature metal bandage, designed to prevent the heavy card from sagging, power cables, documentation and a branded ASUS envelope. I would like to separately note the quality of the materials - all accessories look solid and reliable.
The Asus GeForce RTX 5090 TUF Gaming OC has a massive three-fan cooling system. The fans are black, and the central one is decorated with the ASUS logo, while the two side ones carry the TUF series logo. The metal shroud adds strength to the structure and gives it a stylish brutal look. The design is clearly focused on effective cooling even under maximum loads.
On the back of the card is installed metal backplate, which not only protects the printed circuit board, but also participates in heat dissipation. The central part has a large ventilation window for better cooling. The backplate has the TUF Gaming logo, as well as the GeForce RTX model name. The overall design looks strict, without frills, but very reliable.
You can see it on the side of the map huge radiator, which occupies the entire length of the printed circuit board. The aluminum fins of the cooling system are clearly visible through the ventilation grilles, which indicates high efficiency of heat dissipation. In addition, a 12-pin power connector can be seen here.
The connection panel contains three DisplayPort 2.1 ports and two HDMI 2.1, providing support for modern monitors with resolutions up to 8K. Also visible are vents that allow hot air to escape outside the case, minimizing the heating of other system components.
On one of the edges of the map is located BIOS switch, which allows you to choose between P Mode (Performance Mode) и Q Mode (silent mode)This is convenient for users who want to balance between maximum performance and noise level.
Asus GeForce RTX 5090 TUF Gaming OC takes 3,5 expansion slots, making it one of the most massive cards on the market. Its impressive dimensions require a spacious case and a powerful cooling system in the PC. Visually, the card looks impressive, and its design inspires confidence.
RTX 5090 TUF Gaming OC clearly superior RTX 4090 TUF Gaming OC in terms of dimensions and massiveness of the design. The card has become longer and wider, and its casing has received an updated design with more expressive metal elements. Visually, it looks more monumental, and three large fans occupying the entire front panel emphasize the serious approach to cooling. In comparison with RTX 4090, the new model seems even more massive, which requires a spacious case for comfortable installation.
When comparing the sides, it is noticeable that the cooling system of the RTX 5090 TUF Gaming OC has been significantly enlarged. Its radiator has become thicker, and the ventilation slots have been widened, which contributes to better heat dissipation. Unlike RTX 4090 TUF Gaming OC, the new model has improved air circulation, which should have a positive effect on operating temperatures. This solution is especially important for users working with high levels Power Target, as the cooling system now copes better with heat dissipation.
The video output connectors on both cards are identical - three DisplayPort 2.1 and two HDMI 2.1. However, on the back panel of the RTX 5090 TUF Gaming OC, the ventilation holes have been redesigned. Unlike RTX 4090, the air is now distributed more evenly, which should reduce the temperature inside the case. This improvement will allow the system to work more stably during long gaming sessions and under maximum loads.
ARCHITECTURE |
The Blackwell architecture, introduced by Nvidia in 2024, was a significant step forward in graphics technology, replacing the previous Hopper and Ada Lovelace architectures. Named after renowned mathematician and statistician David Blackwell, the architecture was officially announced at Nvidia's GTC 2024 keynote on March 18, 2024.
Manufacturing process and technical characteristics
Blackwell is manufactured using TSMC's advanced 4NP process technology, an improvement on the previous 4N node used for the Hopper and Ada Lovelace architectures. This process enables increased transistor density, which results in higher performance and energy efficiency. For example, the GB100 chip contains 104 billion transistors, which is 30% more than the 80 billion transistors in the previous-generation GH100 chip.
Multi-chip layout
One of the key features of the Blackwell architecture is the use of multi-chip layout. Specifically, Nvidia's B100 accelerator uses two GB100 chips in a single package, connected by a 10 TB/s NV-High Bandwidth Interface (NV-HBI) based on the NVLink 5.0 protocol. The total volume of transistors in such a solution reaches 208 billion. The two GB100 chips are housed on a silicon wafer manufactured using TSMC's CoWoS-L 2.5D packaging technology.
Improved computing units
The Blackwell architecture represents the fifth generation of Tensor Cores for AI computing and floating point operations. For data centers, Blackwell adds support for FP4 and FP6 data types with eight-precision number handling. The previous Hopper architecture introduced the Transformer Engine to split FP32 data into FP8, increasing peak compute performance. In Blackwell, the second-generation Transformer Engine enables even further FP32 partitioning, doubling FP8 compute performance. Using 4-bit data improves efficiency and throughput when training AI models. Nvidia claims to have achieved 20 petaflops of FP4 compute with the B100 accelerator based on two GB100 chips.
NVLink 5.0 Interface and Switching
Blackwell features fifth generation NVLink, which provides connectivity between GPU with a throughput of up to 1,8 TB/s, which is twice as fast as the previous generation. Until 576 GPU can be combined using this technology. The new NV-Link Switch delivers 14,4 TB/s switching throughput, significantly improving performance for parallel computing and big data processing.
Security hardware
The Blackwell architecture incorporates NVIDIA Confidential Computing technology, which provides secure communications based on hardware encryption. This allows you to protect data and computing processes from unauthorized access, which is especially important in today's conditions of increased requirements for information security.
Decompression engine
The new decompression engine in Blackwell is capable of decompressing data at speeds of up to 800 GB/s. Combined with eight stacks of high-speed HBM3e memory and a high-performance bus, this significantly accelerates database queries and analytics workloads, improving overall system efficiency.
Improvements for working with large language models
The second generation of the Transformer Engine in Blackwell delivers improved performance for training and inference of large language models (LLM) and Mixture-of-Experts models with up to 10 trillion parameters. By comparison, GPT-4 has 1,8 trillion parameters. New floating point compute units also support 6-bit and 4-bit data types, making training and inference faster and more efficient.
Reliability, Availability and Serviceability (RAS)
Blackwell has a dedicated RAS (Reliability, Availability, and Serviceability) engine powered by thousands of sensors. This allows for early detection and localization of errors, minimizing downtime and increasing overall system reliability.
Application in data centers
The GB200 NVL72 cluster, which combines 36 GB200 chips, is designed for use in data centers. It supports a memory capacity of up to 13,5 TB and provides data transfer rates of up to 576 TB/s. An improved version called Blackwell is planned for release in 2025. Ultra, which will be used
NVIDIA GeForce RTX 5090 Architecture
GeForce RTX 5090 is built on the latest architecture Blackwell, which replaced Ada Lovelace. The card is equipped with a powerful graphics processor GB202, which is made according to 5nm process technology. This made it possible to place 92 billion transistors, which significantly exceeds the performance of the previous generation. The new technological process provides not only increased energy efficiencybut higher clock speeds without a significant increase in energy consumption.
RTX 5090 uses 176 ROPs and 680 TMUs, making it one of the most powerful cards for handling complex scenes. The GPU is equipped with 21 CUDA cores, which are responsible for general computing and rendering. In combination with support DLSS 4 and updated Ray Tracing new generation, the video card provides the highest quality graphics with an increased frame rate even in the most resource-intensive projects.
The new architecture also includes 512-bit memory bus, which was a significant improvement over RTX 4090. This allows you to achieve throughput 1792 GB / s, which is 40% more than its predecessor. It is used 32 GB GDDR7 video memory, operating at a frequency 1750 MHz. The new GDDR7 standard provides increased storage density and lower latency, which is extremely important for high resolutions and ray tracing.
The RTX 5090's clock speeds have also been significantly increased. The base clock GPU is 2017 MHz, with the possibility of automatic boost up to 2550 MHz. Unlike previous generations, NVIDIA has optimized the work resizable BAR, which allows the processor to work more efficiently with video memory and increases the speed of texture loading.
The innovations also affected the connection interface. The video card uses a new PCIe 5.0, which provides double the bandwidth compared to PCIe 4.0. This is especially important for future games and applications where data processing requires maximum transfer speed.
In the end, the RTX 5090 is flagship graphics card of the new generation, in which NVIDIA significantly increased the power GPU, increased memory bandwidth and optimized power consumption. Blackwell is not just another update, but a real technological leap, providing unprecedented graphics quality in games and professional tasks.
OPERATING MODES GPU Tweak III |
Using proprietary software GPU Tweak III available four modes of operation, which differ in the level of energy consumption, performance and noise of the cooling system. We will consider each of them and analyze how they affect the operation of the video card.
Default Mode
In this mode, the card operates on standard factory frequencies, providing optimal balance between speed, noise and power consumption.
- Power Target: 100%
- GPU Boost: up to 2550 MHz
- Video memory frequency: 28002 MHz
It is a the most balanced profile, in which the card automatically regulates the frequency and speed of fans, adjusting to the load. This mode is ideal for most usersAs does not require intervention and works stably.
Overclocking mode (OC Mode)
This profile should provide additional performance, but in practice the difference with Default Mode is minimal.
- Power Target: 100%
- GPU Boost: 2580 MHz (all +30 MHz to standard mode)
- Video memory frequency: 28002 MHz
There is no real FPS increaseAs 30 MHz is too little change, to make at least some difference. However, the map it heats up a little more and the fans work more activelywhich can lead to slightly higher noise level.
In essence, this mode no sense, as it does not provide any real overclocking or performance gains. It is more of a marketing feature than a useful setting.
Silent Mode
This mode reduces power consumption for quieter operation, without significant loss of performance.
- Power Target: 90%
- GPU Boost: 2520 MHz
- Video memory frequency: 28002 MHz
The difference with Default Mode is total 30 MHz minus that does not affect FPS in any way in games. However the fans are running slower, what makes a map much quieter.
This profile ideal for all-day work and even games, if you don't strive to squeeze out the maximum. Reducing temperature and noise makes Silent Mode is a great choice for those who value comfort.
Manual mode (User Mode – Power Target 67%)
This mode we set it up manuallyTo Minimize energy consumption and heating of the video card.
- Power Target: 67%
- GPU Boost: dynamic frequency reduction depending on load
- Video memory frequency: 28002 MHz
Unlike Silent Mode, here power consumption limitation is more stringent, that leads to core frequency drop under high loads.
- В simple scenes frequencies remain high enough, and there is almost no difference in FPS.
- В heavy games Power Target reduction may lead to a decrease in frequency GPU, because of what FPS will be lower than in other modes.
This mode Suitable for heavy gamesBut with limited productivity. For example, if the card in Default Mode gives 120 FPS, in User Mode (67%) it may be 90-100 FPS.
The mode is perfect for those who want:
✔ Minimize energy consumption and heating.
✔ Make the video card quieter under load.
✔ Use the card in compact cases with limited ventilation.
But it is worth considering that in particularly demanding games, frequencies may drop, reducing FPS more than desired.
HEATING AND POWER CONSUMPTION |
TEST PART |
GAMES |
Black Myth Wukong |
Video cards were tested at resolution 3840x2160 with maximum graphics quality settings.

Dragon Age The Veilguard |

Ghost of Tsushima DIRECTOR'S CUT |

God of War Ragnarök |

Horizon Forbidden West |

Indiana Jones and the Great Circle |

STALKER 2 Heart of Chornobyl |

Senua's Saga - Hellblade II |

Cyberpunk 2077 2.2 |

Kingdom Come Deliverance II |


CONCLUSION |
![]() |