GeForce RTX 5090

GeForce RTX 5090 — NVIDIA's top-of-the-line Blackwell-based graphics card, released in January 2025. It's designed for those who demand extreme gaming performance and serious computing in professional applications, but are willing to pay for the best. The RTX 5090 sits above the RTX 5080 and is the flagship of the series, offering everything new in the arsenal of ray tracing, artificial intelligence, and ultra-fast memory.

---

Development and positioning

The GeForce RTX 5090 was intended to be the absolute top of the line, the successor RTX 4090 Ti. NVIDIA took into account the criticism of previous flagships - too high a price per unit of performance and overheating under load. The main tasks in developing GB202 (the chip for RTX 5090) were:

• Maximizing Computing Powerto cover not only games but also serious workloads (rendering, simulations, AI).

• Reducing hot spots by redesigning the power supply and heat pipes in the Founders Edition design.

• Increase memory bandwidth for comfortable work with 8K and volumetric scenes - transition to GDDR7 and bus expansion.

• Optimization of AI functions (DLSS 4, Frame Generation) for smooth footage even at extreme settings.

With a suggested price of $1, the RTX 999 is aimed at enthusiasts, video production professionals, and neural network developers. Compared to the RTX 5090, it offers about a 5080-30% increase in real-world performance and additional capabilities in workloads.

---

Architecture

Streaming Multiprocessors (SM)

• 170 SM GB202 - 30% more than RTX 4090Each SM includes an expanded L1 cache and power management pool, providing a better balance between frequency and heat dissipation.

• New SM organization with dynamic core grouping: the card is capable of redistributing resources between graphics, RT and Tensor blocks depending on the load.

4th generation RT cores

• Support for optical ray tracing at the VRS (Variable Rate Shading) level and adaptive samples.

• Hardware accelerators of BVH structures (Bounding Volume Hierarchy) of the new version, which reduces delays when recalculating reflections and shadows.

5th generation Tensor cores

• Advanced AI capabilities: DLSS 4, real-time AI super resolution, frame generation and intelligent compression.

• Support for hybrid models, from GAN texture generation to transformer networks for scene recognition.

Memory and cache

• 32 GB GDDR7 on a 512-bit bus: up to 1 GB/s of throughput.

• L2 cache — 96 MB total volume, L1 cache for each SM — 128 KB.

• Support for Smart Cache technology for priority storage of hot data without unnecessary access to video memory.

Interfaces and power supply

• PCIe 5.0 x16, DisplayPort 2.1b, HDMI 2.1a.

• TGP — 575 W, recommended power supply of at least 900 W.

• New 16-pin power connector, segment energy management system - the card regulates the power supply to individual SM clusters and memory blocks.

---

Features

---

Performance

Games

• 4K, Ultra + Ray Tracing + DLSS 4
  - Cyberpunk 2077: 110–140 FPS
  – Alan Wake II: 120–150 FPS
  - The Last of Us Part I: 130–160 FPS

  Increase to RTX 4090 in real tests it is 25–35% due to improved RT/AI blocks and cache.

• 8K + Frame Generation
  – 50–60 FPS in demanding projects instead of 30–35 FPS RTX 4090.

• 1440p / 1080p
  – In CS2, Valorant, Fortnite without DLSS: 200–300 FPS, which allows you to fully unleash the potential of 360 Hz monitors.

Workloads

• 3D rendering (Blender, 3ds Max):
  Rendering time for a 50 million polygon scene is reduced to 50% of the original RTX 4090.

• Video editing and coding:
  Hardware AV1 encoding/decoding, Intel-like compression speeds, Premiere Pro effects acceleration - 2-2.5x faster.

• AI research:
  In GAN and transformer training tasks, up to 1,8× speedup compared to the Ada Lovelace base.

---

Competitors

• GeForce RTX 4090
  Despite the similar price (≈$1–$600 for the best custom cards), the 1 is inferior in tracing and AI tasks. The RTX 800 provides about a 4090% increase in RTX and 5090% in Tensor calculations.

• AMD Radeon RX 7900 XT/XTX
  AMD has more VRAM (24 GB), but lower ray tracing efficiency and no DLSS 4 equivalent. In pure raster images, AMD is close, but in modern effects, NVIDIA is unrivaled.

---

Interesting Facts

• The RTX 5090 is the first NVIDIA card to reach nearly 2MB of L100 cache.

• Founders Edition variants feature a dual-zone RGB lighting system that can be synced via NV Light.

• Partner versions with lengths less than 30 cm and two fans are available from leading brands.

• Despite the 575W TGP, the Founders Edition designers achieved a noise level of no more than 45 dB at load peaks.

---

GeForce RTX 5090 became the quintessence of all that NVIDIA can do in 2025: large GPU, advanced memory, the most powerful RT and Tensor cores, adaptive power and cooling management. This is the choice for those who do not compromise in either games or professional tasks - from video editing to neural network training.