NVIDIA GeForce RTX 4090 Founders Edition

NVIDIA GeForce RTX 4090 Founders Edition — the reference flagship of the Ada Lovelace generation: 16384 CUDA cores, 24 GB of GDDR6X on a 384-bit bus, a 2520 MHz boost clock and a 450 W board power budget. New and used units are supplied from stock with a 1-year warranty.

The reference Ada Lovelace flagship, built around AD102

The Founders Edition is NVIDIA’s own card, produced to stricter tolerances than most third-party designs: a die-cast aluminium unibody frame, a dual flow-through cooler with two axial fans, and a vapor chamber cold plate over the full AD102 die. It runs the complete 16384-CUDA configuration of the RTX 4090 — 128 third-generation RT cores, 512 fourth-generation Tensor cores and 24 GB of 21 Gbps GDDR6X delivering 1008 GB/s of memory bandwidth.

For AI and compute workloads the card exposes full FP32 throughput with FP8 Tensor support, and its 24 GB frame buffer is the practical ceiling of the consumer line — enough for large diffusion models, local LLM inference up to roughly 30B-parameter quantised models, and 8K timeline work. Display output is 3 × DisplayPort 1.4a plus 1 × HDMI 2.1.

Technical specifications

Software support covers DLSS 3.5 (super resolution, DLAA, ray reconstruction and frame generation), NVIDIA Studio and Game Ready drivers, CUDA 12.x, and two 8th-generation NVENC engines with 5th-generation NVDEC — AV1 encode and decode both included. Compute capability is 8.9, DirectX 12 Ultimate and Vulkan RT are supported, and NVLink is not present on this generation. The card is fully driver-compatible with Windows and Linux distributions shipping CUDA 11.8 or later.

Case and power requirements

The FE is 304 mm long and 61 mm thick — NVIDIA rates it 3-slot, but plan on 3.5 slots of physical clearance — so it needs a chassis with at least 320 mm of GPU clearance and a motherboard area free of the two slots below the PCIe x16 connector. NVIDIA recommends an 850 W PSU; because the card draws power from a single 12VHPWR connector (600 W rated, ATX 3.0 spec as 12V-2x6), either use a native ATX 3.0 cable or the bundled 3 × 8-pin adapter, and make sure the connector is seated fully to the shoulder — partial insertion is the leading cause of melted connectors on this generation.

How the Founders Edition compares with popular AIB 4090s

Every RTX 4090 carries the same 16384-CUDA die and 24 GB frame buffer — AIB models differ mainly in cooler mass, factory clock bins and size. Choose the FE when you want the most compact 3-slot layout and the neutral reference aesthetic; choose the AIB cards when case space allows a larger cooler and you want a few percent of extra factory clock.

Condition, grading and testing

Stock is offered in New and Used condition, priced separately — tell us which you need when you enquire. New units arrive factory-sealed; used units are graded on arrival. Every used card goes through the same intake routine:

1. Visual inspection of the shroud, fans, PCIe fingers and 12VHPWR receptacle — any sign of connector overheating or disassembly rejects the card outright.
2. Full-memory stress testing (VRAM train plus extended 3DMark runs) with temperatures and hotspot deltas recorded.
3. Fan, display-output and power-draw verification at the 450 W envelope.
4. A final driver-clean build and benchmark pass, so the card arrives ready to work.

Warranty and lead time

Every card, new or used, is covered by a 1-year warranty against defects in workmanship and function; a verified faulty unit is replaced or refunded and we cover return shipping on the fault. Stock is held in our own warehouse and normally dispatches within three to seven working days, tracked from dispatch.

Purchasing and volume enquiries

The minimum order quantity is 2 units. Pricing is quotation-based rather than listed, because 4090 pricing — especially on the secondary market — moves with each incoming batch. Send the quantity, condition and destination and we will reply with a firm quote and current lead time; larger quantities can be matched for serial range and driver version on request.