Introduction

Galax GeForce RTX 5070 Ti 1-Click OC White is a simple yet elegant custom design of NVIDIA’s latest premium performance segment GPU, which is coming in at its launch MSRP of $750, which is why we are able to post its review today (February 19). We have many other RTX 5070 Ti, but those are overclocked SKUs that are priced well above the MSRP, and we’ll be able to talk more about them tomorrow, so do check back. The new GeForce RTX 5070 Ti occupies a gray area between the performance and enthusiast segments. As our testing will show you, this card is capable of 4K Ultra HD gaming with the latest titles, although that isn’t its recommended use-case, and it’s still being pushed for the 1440p-class that has GPUs across a vast price range. The RTX 5070 Ti is a successor to the RTX 4070 Ti, and not its refresh, the RTX 4070 Ti Super.

The GeForce RTX 5070 Ti has quite a bit in common with the enthusiast segment RTX 5080 we tested last month, besides the swanky new Blackwell graphics architecture driving the two. They share the GB203 silicon, which the RTX 5080 maxes out, enabling all 84 SM, all 112 ROPs, and all 64 MB of L2 cache; but the RTX 5070 Ti is cut down, with 70 SM, 96 ROPs, and 48 MB of cache. The memory sub-system is largely similar between the two except for the on-die cache. You get 16 GB of GDDR7 memory across a 256-bit wide memory bus, which is a big upgrade over the 12 GB of 192-bit GDDR6X of the RTX 4070 Ti. This memory ticks at 28 Gbps, yielding 896 GB/s of memory bandwidth, a huge 77% increase in bandwidth over last gen.

The SM count of 70 remains higher than even the 66 of the RTX 4070 Ti Super. It yields 8,960 CUDA cores, 280 Tensor cores, 70 RT cores, 280 TMUs, and 96 ROPs. The GB203 silicon is built on the same 5 nm-class silicon foundry node that NVIDIA built its RTX 40-series Ada generation on. This node is called the NVIDIA 4N, and has been co-developed by the company with TSMC. All energy efficiency improvements of this generation are purely a function of the architecture and the new power management technology it brings.

Blackwell introduces Neural Rendering, a revolutionary new technology in consumer 3D graphics, which brings objects created by a generative AI model to conventional raster 3D scenes much in the same way as RTX brings ray traced objects to it. You need little introduction to the awesome capabilities of generative AI models to create photorealistic images and video, and can imagine its impact on gaming. AI hence plays a more directly participative role in rendering, and isn’t just relegated to the DLSS upscaler. This is made possible due to a new hardware-based scheduler component called the AI Management Processor (AMP), which lets the GPU accelerate AI models and render graphics in tandem.

The new Blackwell SM sees all 128 CUDA cores being capable of concurrent FP32 and INT32 math; only half the cores in an older Ada generation SM were capable of INT32. The shader execution reordering component of Blackwell comes with the ability to reorder neural shaders. The 5th Gen Tensor core leverages FP4 data formats to increase throughput in lieu of precision. The 4th Gen RT comes with even more fixed function hardware, this time to enable Mega Geometry—a concept similar to Mega Textures, which allows ray traced objects to have exponentially higher triangle counts by leveraging hierarchies.

The new DLSS 4 technology introduces a Transformer-based AI model replacing the convoluted neural networks powering DLSS till now. These models replace the CNN-based ones for upscaling (super resolution) and ray-reconstruction. These are also made available to older RTX 40-series and RTX 30-series. What’s exclusive to the RTX 50-series is Multi Frame Generation (MFG). NVIDIA introduced Frame Generation with the RTX 40-series, where in an AI model is used to draw an entire frame as an intermediate to two conventionally generated ones. With MFG, the model is able to draw up to three such frames between every two conventionally-rendered ones, effectively quadrupling frame rates. This technology relies on an innovation in the display engine of Blackwell called flip metering, which is why it’s exclusive to this generation.

The Galax RTX 5070 Ti 1-Click OC White physically resembles the RTX 5080 1-Click OC in size and shape, but comes in a white design scheme. Both cards meet NVIDIA’s SFF-Ready specs, which is an effort by the company to ensure its latest graphics cards aren’t oversized for some of the smaller mid-tower cases or SFF cases with no more than 3 expansion slots. The card comes with a minor factory overclock of 2467 MHz compared to 2452 MHz reference, the company’s Xtreme Tuner app engages a slightly higher software-defined OC, hence the name. There’s a surprisingly good RGB LED setup for a baseline product, with an RGB LED diffuser along the top, and illumination for each of the three fans.