RTX 4090D — and why it is not an RTX 4090

What the D means

In December 2023 the United States Department of Commerce set an export threshold that the RTX 4090 exceeded. NVIDIA cut a version down to clear it and sold that version in China: the RTX 4090 D. It launched on 28 December 2023 and it has 14,592 CUDA cores against the full card’s 16,384 — about eleven per cent fewer shaders.

Everything else is the same card. Twenty-four gigabytes of GDDR6X on a 384-bit bus, the same 1,008 GB/s of memory bandwidth, the same 2.52 GHz boost clock, the same third-generation ray-tracing cores and fourth-generation tensor cores, the same two AV1 encoders. Nothing that depends on memory capacity or memory speed changes at all, which covers most of what a rented card is asked to do: the models that fit in 24 GB still fit, and the frames that stream still stream.

Where it does cost you is raw arithmetic throughput on work that saturates the shaders — a long render, a batch training run. Budget roughly a tenth less than a full 4090 and you will not be surprised. We say so here rather than printing “RTX 4090” on the page and letting you find out, which is what most listings do.

Across the 56 cards in the catalogue: RTX 4090D holds 24 GB, 13 cards we fit hold more, and 11 cost more per month. Read the whole ladder on the whole catalogue, priced.

The specification, as published

Ada Lovelace. Figures from NVIDIA’s own RTX 40 series documentation, cited in the source of this page.

What it replaced, and what replaces it

From our own shelves rather than from a vendor roadmap. All three of these are orderable today, in the same chassis, so the question is not which is newest but which is worth the difference.

The RTX 3090 has the same 24 GB and is the older, slower, cheaper way to get it — if capacity is what you are buying and speed is not, read the RTX 30 series page before you spend the difference. The RTX 5090 steps up to 32 GB, and is a PCIe Gen 5 card on platforms that are Gen 3 and Gen 4, so read the section below before assuming newer is faster here.

Which machine it goes in, and what link it gets there

The 4090 D is a PCIe Gen 4 card. Two of our three platforms that take it are Gen 3, and one is Gen 4.

This is the fact behind support ticket #3347, where a customer moved to a newer card on an older platform and rendered slower than before. Once the scene or the weights are resident on the card the slot is idle and its generation stops mattering. It matters when the work crosses the slot on every step.

What actually fits in 24 GB

The other ceiling is power: this card draws up to 425 W on its own, which is what makes more than one of them in a chassis a build we quote rather than a box you tick.

What it costs, with the machine it goes in

The card is one line on the invoice and the machine is another. Both halves below come from one reading of the order catalogue, so the price and the specification beside it always belong to each other.

Bandwidth is unmetered in both directions with no egress charge, one IPv4 address is included, there is no contract and one month is the default term. For every other card we fit, priced the same way, read the whole catalogue, priced. For machines racked and ready today, /instant.