NVIDIA Announces the RTX 30 Series

NVIDIA CEO Jensen Huang announced the GeForce RTX 30 series on September 1, 2020 in an online event, with the RTX 3080 ($699), RTX 3090 ($1,499), and later RTX 3070 ($499, October 29) going on sale September 17, October 24, and October 29 respectively. The Ampere architecture (GA102/GA104 dies manufactured on Samsung 8nm, not TSMC as NVIDIA had used previously) reorganized the streaming multiprocessor design so that each SM contained twice the FP32 CUDA core count of Turing by adding a second FP32 data path that could also execute INT32 simultaneously — effectively doubling FP32 throughput without doubling the SM count, at the cost of the INT32-FP32 overlap being an architecture-specific optimization rather than general-purpose parallelism. The RTX 3080 shipped with 8,704 CUDA cores and 10 GB GDDR6X at 320 GB/s bandwidth; the RTX 3090 shipped with 10,496 CUDA cores and 24 GB GDDR6X at 936 GB/s. The GDDR6X memory (developed by Micron) used PAM4 (Pulse-Amplitude Modulation 4) signaling at 19.5 Gbps per pin versus GDDR6’s 16 Gbps, adding approximately 21% bandwidth at equivalent pin count and frequency. Compared to the RTX 2080 Ti ($999, Turing’s flagship, 4,352 CUDA cores, 11 GB GDDR6, 14.2 TFLOPS FP32), the RTX 3080 offered approximately 29.8 TFLOPS FP32 — more than 2× the shader throughput — at $699.

The second-generation RT Cores (hardware BVH traversal and ray-box/ray-triangle intersection, same function as Turing’s first generation but higher throughput) and third-generation Tensor Cores (supporting structured sparsity — 2:4 sparse weight matrices providing 2× effective throughput for AI inference) powered two key features: real-time ray tracing in games (available since Turing but now with enough performance headroom for ray-traced reflections and shadows at 1080p/1440p without crippling frame rates) and DLSS 2.0. DLSS 2.0 (launched earlier in 2020 with Control and Minecraft RTX) used a temporal upscaling model trained offline by NVIDIA on 16K ground-truth frames, producing quality comparable to native 4K resolution from internal rendering at 1080p or lower — a 2–4× performance multiplier that made high-quality ray tracing practical on the RTX 3080 at 4K. DLSS 2.0 was a substantial improvement over the game-specific, often-poor-quality DLSS 1.0 that had shipped with early Turing games.

Cards sold out on launch day within seconds despite NVIDIA implementing a “retailer queue” system, and remained unavailable at retail prices through 2020 and most of 2021. Three concurrent factors constrained supply: the COVID-19 pandemic’s effect on semiconductor manufacturing capacity and logistics (TSMC and Samsung fab allocation was shared across all chip categories simultaneously facing demand spikes), Ethereum cryptocurrency mining (Ethereum’s proof-of-work algorithm was heavily GPU-optimized, and the RTX 3080’s GDDR6X made it approximately 8% faster than alternatives; Ethereum’s $4,000+ price in early 2021 made buying GPUs at double retail profitable for miners), and general semiconductor shortages affecting GDDR6 and GDDR6X DRAM supply chains. NVIDIA responded by launching “LHR” (Lite Hash Rate) variants of the RTX 30 series in May 2021 with firmware-limited Ethereum mining throughput (~50% of full rate), reducing the mining profitability of new inventory. The RTX 30 series remained the de facto standard for PC gaming and local machine learning development through 2022, when the RTX 40 series (Ada Lovelace, TSMC N4, September 2022) launched with another substantial performance jump and improved availability.