GeForce RTX Turns Eight: How Turing Changed PC Graphics

NVIDIA announced the first GeForce RTX cards at Gamescom on August 20, 2018 and set September 20 as the availability date for the GeForce RTX 2080 and RTX 2080 Ti. Eight years later, the useful anniversary is not the model names themselves but the architecture underneath them: Turing changed how NVIDIA divided real-time graphics work across a consumer GPU.

The launch combined conventional rasterization hardware with dedicated RT Cores for ray-tracing work and Tensor Cores for neural-network operations. NVIDIA described the result as hybrid rendering: rasterization remained central, while selected ray-traced effects and AI-assisted processing could be inserted into the frame. That distinction matters because the first RTX generation was not a proposal to replace rasterization wholesale.

Real-time ray tracing existed before Turing, but tracing useful numbers of rays through complex scenes is expensive when all of the acceleration-structure traversal and intersection work competes with ordinary shader execution. Turing added dedicated RT Cores to accelerate bounding-volume-hierarchy traversal and ray/triangle intersection testing.

That did not make ray tracing free. Games still had to build and update acceleration structures, issue rays, shade the resulting hits, denoise noisy results and fit the work into a frame budget. The architectural change was narrower and more durable: consumer GeForce hardware now contained a fixed-function path specifically intended to accelerate a major part of the ray-tracing workload.

Turing also brought Tensor Cores to GeForce. NVIDIA had introduced Tensor Cores with the Volta architecture for deep-learning workloads; Turing integrated a newer implementation into an architecture sold directly to PC gamers. NVIDIA’s RTX software stack used that capability for neural graphics rather than treating it only as general AI compute.

The most visible launch-era example was Deep Learning Super Sampling. The first DLSS generation and later DLSS versions should not be treated as one unchanged algorithm, but the hardware/software split established in 2018 is important: render part of the graphics workload conventionally, then use dedicated matrix hardware for learned image processing. Modern neural rendering has expanded well beyond that original implementation.

RTX branding can obscure the less dramatic changes inside the Turing Streaming Multiprocessor. NVIDIA’s architecture whitepaper describes an independent integer datapath that can execute integer instructions concurrently with floating-point math. On earlier Pascal hardware, those integer instructions could occupy execution opportunities that otherwise served floating-point work.

Turing also redesigned the SM memory path, unifying shared memory, texture caching and memory-load caching into a configurable architecture. NVIDIA claimed substantially more L1 bandwidth and capacity for common workloads. Those changes benefited conventional rendering and compute code independently of whether a game enabled ray tracing or DLSS.

NVIDIA’s original launch material made aggressive performance claims, including statements about gains versus Pascal and the potential of RTX rendering. Those were vendor claims tied to NVIDIA’s chosen workloads and its new RTX-OPS metric, not evidence that every 2018 game became several times faster on an RTX 2080.

It is more useful to separate architectural facts from launch positioning. Turing did ship dedicated RT hardware, Tensor Cores, a redesigned SM and the software foundations for RTX and DLSS. Whether a particular game benefited depended on engine support, effect selection, image-quality targets, driver maturity and the specific GPU. This retrospective therefore does not infer launch reception, adoption rates or universal performance gains from NVIDIA’s announcement numbers.

Eight years of later hardware and software have changed the implementation details substantially. Current ray-traced games can use more complex acceleration structures and lighting workloads, while neural rendering now covers a broader set of reconstruction and frame-generation techniques than the original DLSS launch pitch. Those later developments should not be back-projected onto first-generation Turing hardware.

The durable idea is the division of graphics work among specialized engines. Rasterization and programmable shaders remain important, ray-tracing hardware accelerates traversal and intersection work, and matrix hardware can run neural processing alongside the rest of the pipeline. Turing did not invent every technique involved, and it did not single-handedly determine the direction of the industry. It did, however, put that combination into the first GeForce RTX generation sold to PC gamers on September 20, 2018.