Upgrading an L40S server to the NVIDIA RTX PRO 6000 Blackwell Server Edition: the checks

Can an L40S server take the RTX PRO 6000 Server Edition?

An L40S server can take the NVIDIA RTX PRO 6000 Blackwell Server Edition only if its manufacturer lists the new card for that exact model, slot, cable and fan set, and many servers bought with L40S cards in 2023 and 2024 are not on those lists. Where the swap is allowed, each card doubles its memory to 96 GB and draws up to 600 W instead of 350 W, and the power feed, cables and airflow decide whether the server keeps its place in the rack.

Not listed are, among others, Dell's sixteenth-generation PowerEdge servers and the HPE ProLiant DL380a Gen11; listed are, for example, the Lenovo SR675 V3 with 5th Gen AMD EPYC processors and the HPE ProLiant DL385 Gen11. Both cards are passive, dual-slot, full-height and full-length, with one 16-pin power connector, so the new card fits the same bay. Almost everything else changes.

NVIDIA L40S product page and RTX PRO 6000 Blackwell Server Edition product brief SP-12355-001_v02, as quoted in our L40S and RTX PRO 6000 articles (read 24 October 2026). Usable memory by our users-per-card rule: 0.9 × the memory the driver reports, less 3 GiB per card.

Our comparison of the L40S and RTX PRO 6000 Server Edition decides which card to buy, and our list of RTX PRO 6000 Server Edition compatible servers covers new servers. This guide starts from the L40S server you already own.

L40S-era servers on the makers' lists

NVIDIA's certified-systems list carried the L40S in a broad range of servers: Dell PowerEdge R750xa, R760 and R760xa, HPE ProLiant DL380a and DL385 Gen11, Lenovo SR650 V3, SR655 V3, SR665 V3 and SR675 V3, and Supermicro's 2U and 4U GPU lines. For the RTX PRO 6000 Server Edition, the makers' lists are shorter.

Lenovo Press LP2263 (updated 28 July 2026) and Supermicro's RTX PRO 6000 systems page, read on 24 October 2026; Dell PowerEdge Server GPU Matrix (August 2026), NVIDIA Accelerators for HPE QuickSpecs (version 83), HPE DL385 Gen11 QuickSpecs, Dell XE7745 technical guide Rev. A09 and Cisco's C845A M8 data sheet (24 August 2026), as read for our compatible-servers article on 9 and 24 October 2026. The L40S list is NVIDIA's certified-systems list as quoted in our L40S comparison.

Three entries need a closer look. Lenovo's footnote for the SR675 V3 reads "Supported with 5th Gen AMD EPYC processors only", so an SR675 V3 bought with 4th Gen EPYC processors is outside the list even though the chassis is the same. Lenovo's guide marks the other ThinkSystem servers in its tables, apart from the SR650a V4 and the SR675 V3, as not supported, and states that the card "is only supported on SR650a V4" with its machine types. Supermicro's page shows the AMD-based AS-4125GS-TNRT with up to four cards and the TNRT2 with up to eight; the Intel-based SYS-421GE-TNRT and SYS-521GE-TNRT do not appear on it.

The servers on these lists are current platforms. We found no server of the PCIe Gen4 generation, with AMD EPYC 7003 or 3rd Gen Xeon Scalable processors, on any maker's list for the card. Send us the exact model number of your L40S server, its processors and its power supplies; ask us to check the swap, and our team replies with what the maker lists and what has to change.

Power per card and per server

The new card draws up to 600 W against the 350 W of the L40S, 250 W more per card at full power. NVIDIA's product brief gives the Server Edition two power modes, 600 W or 450 W maximum, each with a 300 W minimum. At the 450 W cap the difference shrinks to 100 W per card; StorageReview found that the card at 450 W "gave up 2 to 5 percent at short context" and "4 to 9 percent at long context", measured on a Workstation Edition with the same GPU and memory.

Our estimate: board power from NVIDIA's L40S page and the Server Edition brief, 280 W per EPYC 9355 and 360 W per EPYC 9555, plus 15 per cent for memory, drives, fans and conversion losses; a measurement on the delivered system confirms or corrects it.

The table decides the feed. We plan 2.94 kW per 230 V 16 A single-phase feed and 5.89 kW per 32 A feed, 80 per cent of the rating. A four-card L40S server at about 2.3 kW stays on a 16 A feed; with four Server Edition cards at 600 W it needs 32 A or three-phase, at the 450 W cap it stays on 16 A. An eight-card L40S server at about 4.0 kW runs on 32 A single-phase; at 600 W per card it moves to a 400 V three-phase position, at 450 W it stays on 32 A. With redundant supplies on A and B feeds, each side has to carry the whole server. Our guide to GPU server power on 16 A, 32 A and three-phase works through the feeds.

The power supplies inside the server are the next limit. HPE's DL380a Gen12 QuickSpecs, for example, require 2,400 W or 3,200 W supplies whenever the RTX PRO 6000 is configured, with five supplies for two or four double-wide cards. An L40S server configured for 350 W cards may have smaller or fewer supplies, so check the maker's power calculator for the new configuration with your exact supply model.

The 16-pin cable and its sense pins

Both cards take power through a 16-pin connector, and the same plug does not mean the same cable. The cable signals through two sense pins how much power the supply side delivers, and the Server Edition reads them at power-up. NVIDIA's brief states that if the cable signals less than the card's default cap, "the card will not boot". The server itself starts normally, and the card stays dark.

The makers treat the cable as part of the configuration. Lenovo states that the card's option part number "does not ship with auxiliary power cables", that "Cables are server-specific due to length requirements", and that "For field upgrades, cables will need to be ordered separately". For the SR675 V3 it names a 235 mm 16-pin cable, 4X97A84510, and for the SR650a V4 a 320 mm 16-pin cable, 4X97B04860. We found no maker that documents reusing an L40S cable for the Server Edition, so order the cable the maker lists for the new card and slot.

A power cap set later does not replace the cable. A cap set with nvidia-smi must be set again after every driver load, and one set out of band through SMBPBI persists, according to the brief. Our guide to the RTX PRO 6000 Server Edition at 600, 450 or 300 W explains the modes and the commands.

Airflow and inlet temperature

A passive 600 W card needs more air than a 350 W one. NVIDIA's brief asks for 43 CFM through each Server Edition card at a 25 °C heat-sink inlet, rising to 125 CFM at 45 °C. For the whole server, our planning figure of 160 CFM per kW at an 11 °C rise gives about 360 CFM for the four-card L40S server and about 540 CFM with four Server Edition cards at 600 W; for eight cards, about 650 CFM against about 1,000 CFM.

The makers lower the inlet limit with the new card. Dell's XE7745 technical guide allows 35 °C at the inlet with eight L40S and 30 °C with eight RTX PRO 6000 Server Edition. Lenovo limits the SR650a V4 to 25 °C with the RTX PRO 6000 cards. HPE supports the Maximum Performance fan kit in the DL385 Gen11 only when 600 W GPUs are in the configuration. Measure the inlet temperature at the rack position before you order: a room that ran four L40S comfortably at 32 °C is above Dell's limit for the new card. Our guide to GPU server inlet temperature covers where to measure.

PCIe link, driver and CUDA

The L40S uses PCIe Gen4 x16, which NVIDIA gives as 64 GB/s bidirectional; the Server Edition uses Gen5 x16, 128 GB/s. On a listed server the cards get Gen5, which matters because neither card has NVLink, and a model split over several cards exchanges its tensors over PCIe. By default vLLM disables its custom all-reduce on more than two GPUs without NVLink; vLLM 0.30.0 (release notes, 22 September 2026) added an opt-in FlashInfer PCIe IPC all-reduce for such systems (#53576). Our article on RTX PRO 6000 without NVLink explains the layouts.

The software has to move before the cards go in. The Server Edition's brief sets driver R575 and CUDA 12.9 as minimums, NVIDIA requires its open kernel modules for Blackwell, and the card has compute capability 12.0, so containers and frameworks need builds that include it. An L40S host still on the R535 branch, which ended in June 2026, needs a newer branch in any case. As of October 2026, R580 is the long-term branch with support to June 2028, and NVIDIA's security bulletin of 30 September 2026 names 580.178.04 as the first fixed R580 driver. Bring the BIOS and management controller firmware to the level the maker lists for the card. Our guide to Blackwell RTX software support for SM120 lists the framework versions.

vGPU hosts: new profile names and ESXi 8.0 U3g

A VDI or compute host changes more than its driver. NVIDIA added the Server Edition in vGPU release 19.0, and on vSphere its support matrix lists the card for ESXi 8.0 Update 3g or later and for VMware Cloud Foundation 9.0.1 or later. The card ships in Display Off mode, which NVIDIA's brief names as required for vGPU.

Every VM needs a new vGPU type. The L40S types are named L40S-1Q to L40S-48Q, the Server Edition types begin with DC, such as DC-24Q, and NVIDIA's vWS sizing guide notes that "1 GB profiles are not available on Blackwell GPUs". A host that ran 32 office desktops on L40S-1B therefore moves them to a 2 GB type: NVIDIA's vPC sizing guide gives 32 per card on the whole card and 48 with MIG-backed time-sliced vGPU, the latter on vSphere only from VCF 9.1.

NVIDIA vGPU types reference (29 September 2026) and NVIDIA AI Enterprise 8.2 vGPU references for Ada and Blackwell (2 September 2026), as quoted in our L40S and RTX PRO 6000 Server Edition vGPU profile guides.

The licences follow the series letter, as on the L40S: Q types need vWS, B types vPC, and C types NVIDIA AI Enterprise per GPU; neither card comes with an AI Enterprise subscription. Our guides to L40S vGPU profiles and RTX PRO 6000 Server Edition vGPU profiles list every type. Mixing L40S and Server Edition cards in one host is possible where the maker lists both, but we advise against it: two profile sets, two power budgets, and a multi-GPU job runs at the pace of the slower card.

What the swap changes: a four-card example

Take a four-card L40S server on a 230 V 16 A feed that serves an internal assistant. Its 192 GB give 149.2 GiB by our rule. Llama 3.3 70B in FP8, 67.7 GiB of weights, runs on it only as one copy over all four cards, with cache for 65 conversations of 8K with an FP8 cache.

With four Server Edition cards at the 450 W cap, the server stays on its 16 A feed at about 2.7 kW by our estimate and holds 332.2 GiB. The same model then runs as one copy per card, 48 conversations with no traffic between the cards, or as one copy over four cards with 211, as our guide to the four-card RTX PRO 6000 Server Edition server works out. Models that never fitted the L40S server, such as gpt-oss-120b as one copy per card, become possible.

The published results show the scale of the change on the one benchmark both cards ran. In MLPerf® Inference: Datacenter, closed division, a Dell PowerEdge XE7745 with eight L40S reached 3,481.5 tokens per second on Llama 2 70B offline (v5.0, entry 5.0-0018, retrieved from mlcommons.org on 24 September 2026), and a Dell PowerEdge XE7740 with eight RTX PRO 6000 Server Edition 29,796.7 (v6.1, entry 6.1-0036, retrieved from mlcommons.org on 24 October 2026), results verified by MLCommons Association. That is 8.6 times by our division, not an MLPerf metric, across two rounds, two server models and different software. The MLPerf name and logo are registered and unregistered trademarks of MLCommons Association in the United States and other countries. All rights reserved. Unauthorized use strictly prohibited. See www.mlcommons.org for more information.

Swap the cards or replace the server

The order of the checks matters, because an early failure saves the later ones:

1. The maker lists the card for your exact machine type, processor generation and slots.
2. The power supplies carry the new total, with redundancy, by the maker's calculator.
3. The rack feed carries the new wall power on both A and B sides.
4. The maker's cable for the new card and slot is orderable.
5. The inlet temperature at the rack position is below the maker's limit for the new card.
6. The driver branch, the hypervisor and the vGPU release support the card, and the VMs have a new vGPU type each.

If the first check fails, the answer is a new server, and the L40S cards can stay where they are, or move to another listed host for VDI, video or smaller models. NVIDIA's vGPU lifecycle page lists the L40S as fully supported as of 2 October 2026. A new server with RTX PRO 6000 Server Edition cards avoids the PCIe Gen4 link and the older power supplies, and our reference configurations pair two AMD EPYC processors with two or four cards. We supply the cards for listed servers and build new ones to order; send us the server model, the feed and the workload, and we say which of the two routes fits.

What we supply

We supply the NVIDIA RTX PRO 6000 Blackwell Server Edition as a card for servers you already run, after a check of the platform, power and cooling, and in AI servers built to order with two to eight cards, assembled and tested before delivery. Both come on one EU contract and invoice, with manufacturer warranty on every component, handled through us; quantities and lead time are confirmed in writing with the quote. vGPU and NVIDIA AI Enterprise licences come on the same invoice. The card's specifications and editions are on our RTX PRO 6000 page, and our guide to GPU server acceptance testing describes the checks to run once the new cards are in.

FAQ