NVIDIA T4 successor: L4, RTX PRO 2000 or RTX PRO 4000 SFF for 70 W slots

Which card replaces a T4

The NVIDIA L4 is the T4's successor for servers: a passive, low-profile, single-slot card powered from the slot, like the T4, with 24 GB of memory instead of 16 GB, FP8 support and NVIDIA vGPU. The RTX PRO 2000 and RTX PRO 4000 SFF Blackwell replace a T4 only in systems that have room for a dual-slot card with its own blower fan, which in practice means compact workstations and edge boxes. All three are supplied by us.

T4 cards sit in rack servers for inference and video, in virtual-desktop hosts and in cloud instances. The question of a successor arrives with dates: the end of new T4 shipments, the end of extended vGPU support and, for cloud users, the shutdown of T4 instances. The table sorts the choice by where the card runs today.

Our recommendation from NVIDIA's product pages and datasheets and NVIDIA's vGPU list, read on 24 October 2026; Google Cloud's T4 end-of-support page, last updated 13 October 2026.

Our comparison of the RTX PRO 4000 SFF, RTX PRO 2000 and L4 sets the three new cards against each other. This page starts from the installed T4 and covers what changes when it goes. We supply all three successors; describe the systems your T4 cards run in, and our team replies with the card each one takes.

Why T4 owners are deciding now

NVIDIA's list of GPUs supported by vGPU, last updated on 2 October 2026, gives the Tesla T4 the status "Extended Support", with "Dec 2024" as the date last shipped, "Dec 2027" as the end of extended support and the end of maintenance still "TBD". The L4 is in "Full Support" and listed as active. NVIDIA's separate page of end-of-software-support dates, updated the same day, still names the T4 among the GPUs that remain fully supported and gives no dates for it, so the vGPU list, the stricter of the two, is the source to plan with. NVIDIA's definitions of the phases explain why the difference matters little until December 2027: during Extended Full Support a GPU "is supported in the same way as during Phase I: Full Support", and only the maintenance phase that follows drops new features and technical support.

In the cloud the deadline is harder. Google Cloud's end-of-support page, last updated on 13 October 2026, states that from 1 August 2027 "you cannot create, launch, or access any Google Cloud resources that run NVIDIA T4 GPUs", that remaining T4 instances will be shut down, and that three-year commitments for T4 can no longer be bought or renewed. Google recommends the G2 machine series with the L4 (24 GB) or the G4 series with the RTX PRO 6000 (96 GB).

The software side is calmer. The CUDA 13.0 release notes state that it "supports all NVIDIA architectures from Turing through Grace Blackwell", and the T4 is a Turing card with compute capability 7.5. What CUDA 13 removed was Maxwell, Pascal and Volta. TensorRT 11.4.0, in its support matrix of 5 October 2026, "supports NVIDIA hardware with compute capability SM 7.5 or higher" and names the T4 in its GPU filter, and vLLM's installation guide lists compute capability 7.5 as its minimum, with the T4 as the first example. Current drivers list the T4 too: the release notes for data-centre driver 595.91.07 include "NVIDIA T4, T4G". A T4 server therefore keeps current software for now. The case for replacing it rests on memory, formats, vGPU support and the dates above.

T4, L4, RTX PRO 4000 SFF and RTX PRO 2000 compared

NVIDIA T4 product page and product brief PB-09256-001_v05 (April 2020); NVIDIA L4 product page; NVIDIA professional desktop GPU overview and RTX PRO datasheets (June 2026); NVIDIA Video Codec SDK support matrix; NVIDIA vGPU list of 2 October 2026; all read on 24 October 2026.

The T4 brief describes "a single-slot, low-profile, 6.6-inch PCI Express Gen3" card, half height and half length, "passively cooled", with 16 GB of GDDR6 shipped "with ECC and page retirement enabled by default", and air accepted from either side. The L4 keeps that slot format and adds 8 GB. The two RTX PRO cards have the same 2.7-inch height but take two slots and bring four Mini DisplayPort 2.1b outputs.

The tensor row compares unlike formats. On the same footing, dense INT8, the T4's 130 TOPS (NVIDIA's T4 page gives no sparse figure) stands against 242.5 TOPS on the L4, half of NVIDIA's sparse 485, which makes the L4 about 1.9 times faster on paper, by our arithmetic. FP16 gives the same ratio: 65 TFLOPS on the T4 against 121 dense on the L4.

Memory bandwidth: where the L4 is no faster

Language models generate each token by reading their weights from memory, so bandwidth caps the speed for a single user. The L4 has 300 GB/s, 6 per cent less than the T4's 320 GB/s. For an 8B model in FP8, 9.45 GB of weights, the ceiling is about 34 tokens per second on a T4, 32 on an L4, 46 on an RTX PRO 4000 SFF and 30 on an RTX PRO 2000, by our arithmetic: bandwidth divided by the bytes read per token. These are upper bounds, and real decoding reaches less.

The L4 wins where the T4 runs out of memory or compute: larger models, longer contexts, more conversations batched together and the prompt phase, which is bound by compute. A T4 that serves one small model to a few users at a time will not feel much faster on an L4. Where the speed of a single stream matters, the RTX PRO 4000 SFF is the card with more bandwidth, as long as the system can take it.

Formats: FP8 and FP4 against INT8 and FP16

The T4 computes in FP16, INT8 and INT4. FP8 arrived with Ada and Hopper. vLLM's FP8 documentation of 24 September 2026 states that "FP8 computation is supported on NVIDIA GPUs with compute capability >= 8.9 (Ada Lovelace, Hopper, Blackwell)", while on Turing an FP8 model runs "as weight-only W8A16, utilizing FP8 Marlin": the weights are stored in 8 bits and computed in 16. On an L4 the same checkpoint computes in FP8. The RTX PRO cards add NVFP4, the 4-bit format of Blackwell Tensor Cores.

Services that run in INT8 through TensorRT keep that path on the L4, which NVIDIA rates at 485 INT8 TOPS with sparsity, and TensorRT 11.4 supports both cards. Teams that move to FP8 or NVFP4 with the new card should plan an accuracy check on their own data, since the change of format alters the numbers the model computes with.

Language models in 16 GB and 24 GB

Our users-per-card rule takes 90 per cent of the memory the driver reports and subtracts 3 GiB for CUDA graphs and activations; what remains after the weights holds the KV cache. GDDR6 with ECC on gives up 6.25 per cent of its capacity, and the T4 ships with ECC on; on GDDR7 NVIDIA states ECC costs no capacity. By that rule the T4 has 10.4 GiB for weights and cache, the L4 17.2 GiB with ECC on, the RTX PRO 4000 SFF 18.5 GiB and the RTX PRO 2000 11.3 GiB.

Our estimate by the users-per-card rule: driver-visible memory scaled from 97,887 MiB for a 96 GB card, 0.9 × that less 3 GiB; Qwen3-8B FP8 weights 8.8 GiB; cache per 8,192-token conversation 1.125 GiB in 16 bits and 0.5625 GiB in FP8.

The T4 holds an 8B model in FP8 with room for barely one long conversation, so on a T4 such a model needs short contexts or 4-bit weights. On a 24 GB card the same model serves 14 to 17 conversations at once. A 14B model in FP8 or a 30B mixture-of-experts model in 4-bit needs the 24 GB cards, as our guides to small language models on the L4 and RTX PRO 2000 and 30B MoE models on 24 GB GPUs show.

Server slots: what fits where the T4 was

A server that holds T4 cards has low-profile slots, power for about 70 W per slot from the board and fans sized for passive cards. The L4 matches all three: 72 W maximum from the slot with no auxiliary connector, a low-profile single-slot passive card and, as NVIDIA's documents for both cards state, air accepted in either direction. It connects at PCIe 4.0 x16. What decides whether it may go into a given server is the maker's support list: Lenovo's L4 product guide LP1717, for example, lists eight L4 for the ThinkSystem SR650 V3 and four for the SR635 V3, and marks tower models such as the ST50 V3 as not supported.

The RTX PRO 2000 and RTX PRO 4000 SFF need two slot widths and blow their own air. In a 1U server with single-width low-profile risers they do not fit, and in a 2U server they occupy two positions and depend on the maker listing them. Both cards draw power from the slot only; NVIDIA's quick start guide states that they "do not use a power cable". NVIDIA's power guidelines say a PCIe slot is "typically" rated for 75 W, so the slot rating in a compact PC needs checking before the order.

Video and camera streams

NVIDIA's T4 page states that the card "can decode up to 38 full-HD video streams". Its single encoder is a 7th-generation NVENC without AV1, according to NVIDIA's Video Codec SDK support matrix. The L4 has two 8th-generation encoders with AV1, four decoders and four JPEG decoders, and NVIDIA quotes up to 1,040 concurrent AV1 streams at 720p30 for a server with eight L4 cards. For camera analytics and transcoding, the L4 is the T4's successor in the same slot, and our article on the GPU server for video analytics sizes streams per card.

The RTX PRO 4000 SFF has two 9th-generation encoders and the RTX PRO 2000 one, both with AV1 and with 4:2:2 encoding and decoding for H.264 and HEVC, which broadcast formats use. They suit a video workstation with monitors attached more than a stream server.

Virtual desktops on T4 hosts

A T4 vGPU host moves to the L4. Of the three successors only the L4 is on NVIDIA's vGPU list, supported since release 15.2 and in Full Support. Its profiles run from 1 to 24 GB, up to 24 desktops per card with 1 GB profiles; our guide to L4 vGPU profiles lists them. The licence editions do not change with the card. VM migration between hosts needs the same GPU type on both, so a mixed cluster is planned as separate pools.

A four-card T4 server, card for card

Replacing four T4 with four L4 keeps the card power almost unchanged, 280 W against 288 W, and raises GPU memory from 64 GB to 96 GB, by our arithmetic. Eight watts across four cards is a small change for the power supply; whether the maker supports the L4 in that model decides. Where the T4 servers are old enough to be out of warranty, a new server built around four or eight L4 is the alternative, as our pages to buy an L4 server and on the eight-card L4 server describe.

For services that outgrow 24 GB per card, the step up in a full-height slot is the RTX PRO 4500 Server Edition, 32 GB at 165 W, which our L4 vs RTX PRO 4500 Server Edition comparison covers, or the L40S with 48 GB in our L4 vs L40S comparison. Teams moving from Google's T4 instances to their own hardware can mirror Google's own advice: an L4 server for the G2 class, or RTX PRO 6000 Server Edition cards for the G4 class.

What we supply

We supply the NVIDIA L4, the RTX PRO 4000 SFF and the RTX PRO 2000 as cards or in servers and workstations built to order, on one EU contract and invoice, with manufacturer warranty on every component. When the cards go into servers you already own, we check platform, power and cooling first and say in advance if something will not work. For vGPU hosts, the licences come in the same quote. Tell us the server model and how many T4 cards it holds, and we reply with the successor that the maker supports in that chassis.

FAQ