RTX PRO 4000 SFF, RTX PRO 2000 or L4: choosing a 70 W GPU for compact workstations and edge servers

70 W GPUs that run without a power cable

A graphics card that stays at or below about 70 W can run from the PCIe slot alone. It therefore fits machines that cannot take a larger card: small-form-factor workstations, edge systems in a factory or a shop, and rack servers with low-profile slots and no GPU power cables. NVIDIA’s quick start guide states that the “RTX PRO 4000 SFF and RTX PRO 2000 do not use a power cable”, and the L4 lists no power connector either.

The slot itself needs checking. NVIDIA’s power guidelines say: “Typically, the PCIe slot is rated for 75 W”. Earlier revisions of the PCIe card specification allowed low-profile cards far less power than full-height ones, so a low-profile slot in a compact PC is not automatically a 75 W slot. Before ordering, check that the maker lists the card, at 70 W and 2.7 inches high, for the exact chassis.

RTX PRO 4000 SFF, RTX PRO 2000 and L4 specifications

NVIDIA datasheets and product pages: RTX PRO 4000 SFF (June 2026), RTX PRO 2000 (June 2026), L4 (2023). NVIDIA publishes the Blackwell cards’ tensor rate in FP4 and the L4’s in FP8, so the tensor row does not compare like with like.

At the same precision, FP8 with sparsity, the two Blackwell cards would come to about half their FP4 figures, roughly 385 and 272 TFLOPS by our arithmetic against the L4’s 485, so on paper the L4 has the highest FP8 and FP32 rates of the three. The RTX PRO 4000 SFF has the most memory bandwidth, and bandwidth decides how fast a language model generates tokens.

Cooling: blower cards for desktops, the passive L4 for servers

The L4 is a data-centre card. NVIDIA describes it as “passively cooled” and “requiring system airflow to operate”, with a heatsink that accepts air in either direction, rated for 0 to 50 °C. In a server with front-to-back airflow that is an advantage: no fan to fail, and the server’s fans do the work. In a desktop PC without directed airflow over the slot, a passive card is the wrong choice.

The two RTX PRO cards are workstation cards with their own blower fans and display outputs. NVIDIA calls the SFF card low profile; for the RTX PRO 2000 it gives a height of 2.7 inches, the height of a low-profile card, without naming the brackets supplied. They work in a compact desktop, drive four monitors and need nothing from the chassis except a slot rated for 70 W, room for a dual-slot card 2.7 inches high and some fresh air. In a server they can work where the server maker supports them, but they bring display hardware the server does not need and are not on NVIDIA’s vGPU list. We supply all three, as single cards or in configured workstations and servers. Send us the chassis model and the slot, and we confirm that the card fits and can be cooled there.

Language models in 16 or 24 GB

Memory sets the model size. By our users-per-card rule (90 per cent of the memory the driver reports, which we scale from the 95.6 GiB of a 96 GB card, less 3 GiB, for weights and KV cache), a 24 GB card has about 18.5 GiB and a 16 GB card about 11.3 GiB; on the L4, whose GDDR6 gives up 6.25 per cent of its capacity with ECC on, about 17.2 GiB. A 24 GB card takes gpt-oss-20b (12.8 GiB) with a long context, an 8B model in BF16 or FP8 and a 14B model in FP8; a 32B model in 4-bit, about 18 to 21 GB, fits only in its smallest checkpoints, with cache for about one conversation of 8,192 tokens on the RTX PRO 4000 SFF and less than one on the L4 with ECC on. A 16 GB card takes 8B models at 8-bit and 14B models at 4-bit. OpenAI says gpt-oss-20b runs within 16 GB of memory, but by our rule its 12.8 GiB of weights already exceed the 11.3 GiB a 16 GB card has for weights and cache. Bandwidth then sets the speed of token generation: 432 GB/s on the RTX PRO 4000 SFF against 300 on the L4 and 288 on the RTX PRO 2000.

Formats differ too. The Blackwell cards run NVFP4, the 4-bit format NVIDIA’s TensorRT-LLM supports on Blackwell; the L4 runs FP8, and 4-bit weights as INT4 through AWQ or GPTQ. Our VRAM guide has the full sizing arithmetic.

Video: the L4’s speciality

For video, the L4 has four decoders and four JPEG decoders, more than either RTX PRO card, and two encoders, as many as the RTX PRO 4000 SFF; all three cards encode AV1. NVIDIA quotes up to 1,040 concurrent AV1 streams at 720p30 for a server with eight L4 cards, about 130 per card, and 120 times the AI video performance of a dual-socket CPU server in an end-to-end analytics pipeline, both for eight cards. For camera analytics, transcoding and video search, the L4 is the card of the three designed for this work.

The Blackwell cards add two things the L4 does not have: 4:2:2 encoding and decoding for H.264 and HEVC, which broadcast and professional video formats use, and display outputs for a video workstation. The RTX PRO 4000 SFF has two encoders and two decoders, the RTX PRO 2000 one of each.

Sharing: vGPU only on the L4

Of the three, only the L4 is on NVIDIA’s vGPU list, supported since vGPU 15.2 and still listed as fully supported on NVIDIA’s vGPU lifecycle page (last updated 2 October 2026). It splits into profiles from 1 to 24 GB: up to 24 virtual desktops per card with 1 GB profiles, or up to six compute vGPUs of 4 GB each (L4-4C) under NVIDIA AI Enterprise. Both need NVIDIA licences. None of the three supports MIG. For virtual desktops in low-profile servers, the L4 is the only choice among them.

We build virtual-desktop servers with L4 cards to order, sized by seats per card. Tell us the number of users and the hypervisor, and we reply within one business day with a configuration and a quote.

Which card for which machine

The Blackwell cards also improve on their own predecessors in the same format. Puget Systems measured the RTX PRO 2000 44 per cent faster than the RTX 2000 Ada in V-Ray GPU and 31 per cent faster in Blender, and our comparison of the RTX PRO 4000 and the RTX 4000 Ada covers the SFF pair.

What comes after the L4

We found no NVIDIA document that names a successor to the L4. NVIDIA now calls it “previous-generation”, and as of October 2026 its list of data-centre products has no 70 W Blackwell card to replace it. The card NVIDIA compares with the L4 is the RTX PRO 4500 Blackwell Server Edition: 32 GB at 800 GB/s, 165 W, passive, in a single-slot full-height, full-length format, which NVIDIA says delivers “over 5x the performance of the previous-generation L4 GPU” in small-model AI inference, a claim for which we found no published test conditions. For virtual workstations, NVIDIA’s own sizing guide puts the gain over the L4 at 90 per cent in SPECviewperf 15. The RTX PRO 4500 Server Edition can take the L4’s place in servers with full-size slots and power to spare, not in the low-profile slots the L4 was made for. Where the slot is low-profile and the budget is 72 W, the L4 remains the data-centre option. Our comparison of the L4 and the RTX PRO 4500 Server Edition covers that choice per server.

What we supply

We supply the RTX PRO 4000 SFF, the RTX PRO 2000 and the NVIDIA L4 as cards, in configured workstations or in servers built to order, delivered EU-wide on one EU contract and invoice with manufacturer warranty, alongside the rest of the professional GPU range. Before you order, we confirm that the chassis takes the card in that slot and can cool it. For virtual desktops on the L4, the vGPU licences come in the same quote.

FAQ