NVIDIA GeForce RTX 5090 Founders Edition Review: Still King Of PC Gaming


NVIDIA GeForce RTX 5090: MSRP $1,999
NVIDIA's powerful but pricey GeForce RTX 5090 is the company’s latest flagship GPU and our tests show it crushing everything in AI, gaming, and content creator workloads.


hot flat
  • Fastest GPU Ever
  • Bleeding Edge Features
  • DLSS 4
  • RTX Nerual Rendering
  • Great For AI, Creators, And Gaming
  • Smaller Form Factor Than Previous-Gen
  • 32B Of VRAM
not flat
  • Steep Price Premium
  • High Power Consumption
  • Not A Massive Leap Over 4090
hothardware approved small


Since it was announced at CES a few weeks back, the GeForce RTX 50 series, the Blackwell GPU architecture, and the flagship GeForce RTX 5090 in particular have been the topics of much discussion. Over two years out from the launch of the previous-gen GeForce RTX 4090, enthusiasts, gamers and creators were itching to see what NVIDIA had in store. As such, company CEO Jensen Huang revealed entire desktop and mobile GPU line-ups, including the GeForce RTX 5070 all the way on up the top-end GeForce RTX 5090 we’ll be showing you here.

Before we continue though, we’ve provided a ton of Blackwell and GeForce RTX 50 series coverage, that are must-reads to better understand what the GeForce RTX 5090 is all about. At the very least, please check out our Blackwell GPU architecture piece, which details the many advancements coming with the architecture, including RTX Neural Rendering, RTX Mega Geometry, DLSS 4 with Multi-Frame Gen, the new AI Management Processor,  and the RTX 50 series’ updated media engine. Details on the rest of the line-up are also available here. We not going to re-hash many of the deeper details available in those previous articles here, so please give them a read if you want the full scoop before proceeding...

NVIDIA GeForce RTX 5090 Specifications

geforce rtx 5090 specs comparison

The GeForce RTX 5090 is built around the GB202 GPU, which is manufactured on the same TSMC 4nm process as the previous gen. The GB202 features 92 billion transistors, and as configured on the RTX 5090, it has 21,760 CUDA Cores, 170 RT Cores, and 680 Tensor cores. The GB202 is based on the NVIDIA Blackwell architecture and is outfitted with new 5th gen Tensor Cores, 4th gen Ray Tracing Cores, and updated CUDA cores.

Connected to the GPU via a wide 512-bit interface is 32GB of GDDR7 memory offering up to 1.8 TB/s of total memory bandwidth, which is a 78% increase over the previous gen. The GB202 also features a native PCIe Gen5 interface, and its updated display engine supports DisplayPort 2.1b with UHBR20 and HDMI 2.1b.

blackwell gpu

In total, a full GB202 GPU includes 12 Graphics Processing Clusters (GPCs), 96 Texture Processing Clusters (TPCs), 192 Streaming Multiprocessors (SMs) each with 128 CUDA Cores, and a 512-bit memory interface with sixteen 32-bit memory controllers. As such, a full GB202 GPU includes 24,576 CUDA Cores, 192 RT Cores, 768 Tensor Cores, and 768 Texture Units. A full GB202 GPU also features 128 MB of total L2 cache, along with a 256 KB Register File, and 128 KB of L1/Shared cache.

The GB202 is not fully enabled on the GeForce RTX 5090, however. While the full GB202 GPU has a total of 12 GPCs, 96 TPCs, and 192 SMs, “only” 11 GPCs, 85 TPCs, and 170 SMs are enabled on the RTX 5090. As such, although the GB202 is outfitted with 128 MB of total L2 cache, the RTX 5090 has 96 MB of L2 enabled.

The GB202 may not be fully enabled on the GeForce RTX 5090, but it’s still a step forward over the AD102 powering the RTX 4090. The number of texture units has increased from 512 on the RTX 4090 to 680 on the RTX 5090, which results in higher bilinear-filtered texel fill rates. The RTX 5090 delivers 1636.76 Gigatexels/sec, compared to 1290.2 Gigatexels per second on RTX 4090, though the RTX 5090’s pixel fill rate is actually a bit lower than the 4090’s.

The NVIDIA GeForce RTX 5090 Founders Edition

If you watched our unboxing video above, you’ve already seen the new, more environmentally friendly packaging NVIDIA is using with the GeForce RTX 50 series. The new packaging is not only much smaller than the previous-gen, but the inner box is 100% plastic free, inkless and made of paper fibers. The carboard outer box also has a few cool design elements that hint at what’s inside. Watch the vid if you haven’t already.

geforce rtx 5090 front
geforce rtx 5090 back
What’s inside that box is a totally redesigned GPU, which features a number of innovative firsts. Although it is much more powerful than its predecessor, the GeForce RTX 5090 is much thinner. The RTX 5090 measures 304mm in length, 137mm in height, and it is truly two-slots (40mm) wide – it’s not just a 2-slot wide case bracket. It's a dense package for sure, and feels premium through-and-through.

NVIDIA was able to shrink the GeForce RTX 5090 by developing what it calls a “Double Flow Through” cooler design, which allows both cooling fans to blow air straight through the heatsink, for optimal performance.

rtx 5090 double flow though

To achieve this, the GeForce RTX 5090 has a multi-part PCB setup. The central PCB houses the GPU, memory, and power circuitry. It rests in the center of the card, with only small portions protruding underneath each fan. A separate daughterboard attaches to the central PCB for the PCIe x16 connector, and a third, flexible board runs perpendicular along the bottom with high-speed signaling connections from the central PCB to the display outputs.

The amount of engineering that went into the GeForce RTX 5090’s design is nothing short of impressive. Not only is the cooler far more capable, but NVIDIA had to solve a number of problems to ensure proper signaling over the adjacent PCBs and their connectors. NVIDIA discusses much of the design process in a video available here if you’re interested in some of the nitty-gritty details. 

At the top of the card, you'll find the familiar 12VHPWR 16-pin connector, also used on the RTX 40 series. On the RTX 5090 though, the connector is angled off the back on the PCB, and recessed slightly in the shroud. This configuration should allow for easier cable management and minimize the need to bend the power feed in shallower PC cases. The included power adapter requires four PCIe 8-pin feeds, like the RTX 4090 did, but the cabling is longer and far more flexible. The connector on the adapter is also beefier and emits a sold 'click' when pushed fully into place. Hopefully, these tweaks and updates to the power connector and adapter help prevent some of the issues that affected the RTX 4090.

geforce rtx 5090 bottom
geforce rtx 5090 top

Sitting atop the GPU, memory and power circuitry is a new, unified 3D Vapor Chamber that allows vapor to travel directly to heat pipes to increasing thermal performance, which is connected to a combination of straight and curved fins in the heatsink assembly. designed to reduce back pressure and optimizes airflow through the fins. Directional outlets on the side direct heat up-and-out to combat recirculation of warm air as well.

To ensure optimal heat transfer from the GPU to the vapor chamber, NVIDIA is also using a very high performance liquid metal thermal interface material, similar to what's in the PS5. Liquid metal TIM has much lower thermal resistance than even today's best thermal pastes, which allows more heat to transfer more quickly into the vapor chamber, ultimately lowering GPU temperatures, but there is a catch. Liquid metal is electrically conductive, and if not secured in place, it can run like any liquid and potentially cause shorts. NVIDIA sealed the GPU to sure the liquid metal stays in place though, with the card installed in any orientation. Just be aware that "re-pasting" the RTX 5090 down the road will require extra care to prevent potential problems.

geforce rtx 5090 ports
Outputs in the GeForce RTX 5090 include a trio of DisplayPorts (2.1b) and a single HDMI port (2.1b). Their orientations have been reversed versus previous-gen cards though, and the new case bracket -- which doesn't require venting due to the dual flow-though cooler design -- also features an anti-fingerprint coating.

And with all of that out of the way, let's get to some benchmarks...

Related content