NVIDIA Announces the Turing GPU Architecture

NVIDIA announced the Turing GPU architecture at SIGGRAPH on August 13, 2018, with the GeForce RTX 2080 Ti ($999/$1,199), RTX 2080 ($699/$799), and RTX 2070 ($499) going on sale September 20, 2018. Turing (TU102 die for RTX 2080 Ti, TU104 for RTX 2080) was manufactured on TSMC’s 12nm FinFET process and contained 18.6 billion transistors in the flagship TU102 — 54% more than Pascal’s GP102 (12 billion). The two defining new hardware blocks were RT Cores (first ever in a consumer GPU, one per SM) and 2nd-generation Tensor Cores (inherited from the Volta data-center GPU, brought to consumer hardware for the first time). RT Cores accelerated the most computationally expensive parts of ray tracing: BVH (Bounding Volume Hierarchy) traversal (the tree search to find which scene geometry might intersect a ray) and ray-primitive intersection testing (the geometric calculation to determine exact intersection point and surface normal). Without RT Cores, ray tracing required thousands of CUDA cores to evaluate these tests in software, consuming the majority of the compute budget for even one bounce of reflected light per pixel at 1080p resolution. With RT Cores offloading these two operations to fixed-function hardware, the CUDA cores could focus on shading and color calculation, allowing real-time ray-traced shadows, reflections, and ambient occlusion at playable frame rates for the first time.

The 2nd-generation Tensor Cores (a 4×4×4 matrix multiply–accumulate unit capable of mixed FP16/FP32 precision) powered DLSS (Deep Learning Super Sampling) — a feature where a neural network trained offline on high-resolution ground-truth rendered frames learned to upscale a lower-resolution rendered image to a higher-resolution output. DLSS 1.0 (launched with Control, Battlefield V, and a handful of early RTX titles in late 2018 and early 2019) used game-specific trained models and received mixed reviews: some implementations improved performance significantly (30–40% more frame rate by rendering at 1440p internally and outputting 4K) but with visible softness compared to native resolution. DLSS 2.0 (March 2020, released with Control and Wolfenstein Youngblood patches) replaced the game-specific model with a general temporal upscaling network using motion vectors and per-pixel depth, dramatically improving quality and requiring no per-game training — a version that shipped with RTX 2000 hardware from day one via driver update. DirectX Raytracing (DXR) — the Microsoft API enabling DX12 applications to trace rays against scene geometry — had been announced by Microsoft at GDC in March 2018 and launched alongside Windows 10 October 2018 Update (RS5), providing the API surface that RTX games used.

The RTX 2080 Ti’s $999 starting price was $350 more than the GTX 1080 Ti had been at launch two years earlier, generating criticism that NVIDIA had priced RT hardware beyond the reach of most gamers. Performance on traditional rasterized games (without RT or DLSS) was approximately 25–35% faster than the GTX 1080 Ti — a meaningful improvement but not enough to justify the price premium for gamers who didn’t use the RT/DLSS features, which required game developer implementation to benefit from. By January 2019, fewer than 20 games supported DXR. AMD’s RDNA 1 (RX 5700 series, July 2019) offered no RT hardware; AMD’s first RT hardware (RDNA 2, RX 6000 series) launched in November 2020, two years after Turing established RT as a standard GPU feature. Turing’s Tensor Cores also marked the first time consumer GPU hardware executed neural network inference as a first-class rendering feature, establishing the GPU as the natural hardware for AI-augmented graphics that continued through DLSS 3 (frame generation, Ada Lovelace, 2022) and DLSS 4 (multi-frame generation, Blackwell, 2025).