Introduction

Palit GeForce RTX 5090 GameRock OC is the company's premium custom-design take on NVIDIA's flagship GeForce Blackwell generation graphics card. It's designed to bring you never-before-seen levels of performance, letting you max out any of today's games at 4K Ultra HD. The Palit GameRock line of graphics cards combine the company's highest factory overclock for a given GPU, as well as its most artisanal product design. This card is best meant to be installed on a vertical slot, given that most of its lighting and design bits are arranged around the fan intakes. Even if you install it conventionally, there's plenty going on with the design, plus you can take advantage of its reinforced outer frame that counteracts bending of the card over time.



The GeForce RTX 5090 and the Blackwell graphics architecture it's based on, write a new chapter in real time 3D graphics with the introduction of neural rendering. You're likely already aware of the awesome potential of generative AI in creating photorealistic images and video. NVIDIA figured out a way to get generative AI models render specific objects in a 3D scene with an AI model running in tandem with the main raster 3D workload. NVIDIA worked with Microsoft to standardize this technology at an API level, letting applications directly address the Tensor cores. There's also a dedicated component on the GPU called the AI management processor (AMP), which controls the various AI acceleration resources on the silicon; Blackwell has a head start on this technology, with all the hardware-level groundwork done.

The Blackwell graphics architecture is built around the same TSMC 4N process as the RTX 40-series Ada, so any efficiency gains you see in this review are purely on the backs of the graphics architecture, and the new power management advancements introduced with it. The new Blackwell streaming multiprocessor introduces several new bits. All 128 CUDA cores can perform concurrent FP32 and INT32 math, and shader execution reordering is now aware of neural shaders. The new 5th Gen Tensor cores support FP4 data formats for up to 32 times the throughput of the original Tensor core. The new 4th Gen RT core introduces Triangle Cluster Intersection engine, and Triangle Cluster Decompression engine, making it possible for the RT core to ray test objects with significantly higher polygon counts.

At the heart of the GeForce RTX 5090 is the huge GB202 chip, a 750 mm² slab of silicon with over 92 billion transistors, and 192 SM. The GPU introduces several firsts-to-market features, including the PCI-Express 5.0 x16 host interface, ATX 3.1 + PCIe Gen 5 power architecture, DisplayPort 2.1b UHBR20, and the new GDDR7 memory standard. The GPU features a broad 512-bit GDDR7 memory interface driving 32 GB of 28 Gbps GDDR7 memory for a jaw-dropping 1,792 GB/s memory bandwidth. The GPU needs this for neural rendering, DLSS 4 multi frame generation, and other new technologies to work. The RTX 5090 enables 170 out 192 SM, and 96 MB out of the 128 MB L2 cache physically present on the silicon. It also gets 3 out of 4 NVENC accelerators, and 2 out of 4 NVDEC. This works out to 21,760 CUDA cores, 680 Tensor cores, and 170 RT cores, across 11 GPCs.

The Palit RTX 5090 GameRock introduces the latest version of the company's GameRock cooling solution. Remember what we said about its topside being the best one, it features what Palit calls the Chameleon Panel, an acrylic ARGB LED diffuser with wavy light guides inside, which wrap around the three fan intakes. You'll see in the photography pages just how eye catching this is. The cooling solution features a vapor chamber plate to pull heat from the GPU and memory; and no less than eight heatpipes spreading heat across the large aluminium fin-stack heatsink that's ventilated by a trio of Palit's latest TurboFan 4.0 units, which feature serrated fan blades for increased turbulence, and dual ball bearings. The card puts out an ARGB header, letting you sync the rest of your lighting setup with the Chameleon Panel. Palit's RTX 5090 GameRock runs at default clocks, but there's also an OC model with 2527 MHz (compared to 2410 MHz reference). The company hasn't revealed its price, but we are assuming a price of $2,200 or a 10% premium over NVIDIA's baseline.