NVIDIA L40S vs RTX PRO 6000 Blackwell Server Edition: where Ada still wins

L40S and RTX PRO 6000: universal server GPUs of two generations

Both are the “universal” passive card of their generation: the card a server vendor qualifies first, and one that runs inference, graphics, video and virtual desktops from the same slot. NVIDIA positioned the L40S in this role in 2023, and the RTX PRO 6000 Blackwell Server Edition took it over in 2025. The RTX PRO 6000 Server Edition has twice the memory (96 against 48 GB), 1.85 times the bandwidth, FP4 and MIG, at up to 600 W against 350 W. Choose it for models above about 30B on one card or for isolated tenants; the L40S remains the practical choice where servers are qualified for 350 W cards and the work fits in 48 GB. We supply both cards. Describe your task and our team will help you choose between them.

L40S vs RTX PRO 6000 Server Edition specifications

Sources: NVIDIA product pages and product briefs for L40S (PB-11470) and RTX PRO 6000 Server Edition (SP-12355), Lenovo product guides LP1812 and LP2263. Blackwell PFLOPS figures are NVIDIA’s headline numbers; NVIDIA marks the workstation card’s figures as sparse but states no basis for the Server Edition’s; if they are sparse, dense is about half.

Two rows matter most. With 96 GB against 48 GB, a 70B model in FP8, about 68 GiB of weights, does not fit the L40S at all, while it fits the Blackwell card: by our users-per-card rule (0.9 × the 95.6 GiB the driver reports, less 3 GiB of overhead, less the weights) 15.3 GiB remain for the cache, enough for six concurrent 8K sessions of Llama 3.3 70B with a 16-bit cache and twelve with an FP8 cache. With 1,597 against 864 GB/s, single-stream generation, which is bandwidth-bound, can reach 1.85× the tokens per second before any Tensor Core arithmetic is counted.

What the RTX PRO 6000 adds over the L40S

FP4. Ada has no FP4 Tensor path; on the L40S 4-bit weights run as INT4 through AWQ or GPTQ, with FP16 or FP8 activations, and vLLM loads NVFP4 checkpoints only as weight-only compression with 16-bit arithmetic. Blackwell runs NVFP4 natively, which is why NVIDIA quotes 4 PFLOPS for the format and why a 70B model in NVFP4 (about 40 GiB, 42.7 GB) becomes a one-card deployment with 43.2 GiB left for the cache, 17 concurrent 8K sessions with a 16-bit cache and 34 with an FP8 cache by the same rule.

MIG. The L40S cannot be partitioned in hardware; you divide it with time-sliced vGPU and accept that one noisy tenant affects the others. The RTX PRO 6000 Server Edition splits into four isolated 24 GB instances, each with its own memory and compute, so a hosting team can assign one card to four departments with hardware isolation, which the L40S cannot do.

Density of seats. vGPU 19 lets one Blackwell card with MIG carry up to 48 virtual machines, time-sliced inside the MIG instances; on vSphere this depends on the hypervisor release, so check NVIDIA’s and Broadcom’s release notes for MIG-backed, time-sliced vGPU before planning 48 VMs. The L40S tops out at 32 with the smallest 1 GB profile. For VDI and CAD environments, this number sets the card count.

Video. The Blackwell card has four encoders and four decoders against three and three on the L40S. NVIDIA also lists four JPEG engines for it; for the L40S we found no JPEG engine count in NVIDIA’s specifications. If the workload is transcoding or a vision pipeline with many camera streams, the fourth engine adds a third more throughput from the same slot.

PCIe Gen5. Doubling the link matters when several cards exchange tensors without NVLink, and neither of these cards has NVLink. On a Gen4 platform the link would run at Gen4 x16 and the advantage is gone, and we found no server maker that lists the Blackwell card in a Gen4-generation server, so confirm support per chassis.

When the L40S is still the better choice

Power envelope. The L40S is rated at 350 W, against 600 W by default for the RTX PRO 6000. A Lenovo SR650a V4 takes four L40S cards but only two RTX PRO 6000 Server Edition at 600 W, or four capped to 450 W. Across a rack of existing 2U hosts, this can decide whether you add cards or replace servers.

Airflow. Both cards are passive and depend on chassis fans. At the same inlet temperature a 600 W card needs roughly 70 per cent more air than a 350 W card, which is why HPE recommends a fan upgrade for 600 W cards in the DL380a Gen12 and why Lenovo limits its SR650a V4 to a 25 °C inlet with RTX PRO 6000 cards. In an edge room, a factory cabinet or a hall that runs warm, the older card has margin the new one does not.

Qualification breadth. NVIDIA’s certified-systems list carries the L40S in Dell PowerEdge R750xa, R760 and R760xa, HPE DL380a and DL385 Gen11, Lenovo SR650 V3, SR655 V3, SR665 V3 and SR675 V3, and the Supermicro 2U and 4U GPU lines. If the chassis is already on the floor, the L40S is the card it was built for.

Homogeneity. An existing environment of L40S hosts is simplest to extend with more L40S cards, because it keeps one vGPU profile set, one power budget and one spare pool. The driver is not the problem (vGPU 19 and 20 support both cards from the same host driver), but a multi-GPU job across mixed cards runs at the pace of the slower one.

Workloads that fit in 48 GB. Diffusion models, most vision models, CAD and VDI seats, video pipelines and language models up to about 14B in FP8 fit the L40S with room for many users; a 32B model in FP8 fits as well, but with cache for only about two concurrent 8K conversations, five with an FP8 cache. For the smaller workloads, the extra memory is unused capacity.

For servers already qualified for 350 W cards, we supply the L40S. Tell us the server model and how many cards it should take, and we check the vendor’s support, power and airflow before you order.

When to choose the RTX PRO 6000 Server Edition

Choose the RTX PRO 6000 Server Edition for any model above roughly 30B parameters that has to run on one card, and for any deployment that requires four isolated tenants per card. It is also the card for a new GPU server. The inference node among our AI server configurations carries four RTX PRO 6000 Server Edition cards, 384 GB in total, and 3U to 5U machines such as the Dell XE7745, HPE DL380a Gen12, Lenovo SR675 V3 and Supermicro’s 5U GPU systems take eight or more; building a new multi-GPU server around 2023 silicon makes little sense. The same applies where NVFP4 is planned: NVIDIA’s datasheet says the card’s Tensor Cores “deliver up to 5x the performance of the previous generation for LLM inference with support for FP4 precision”; we found no published test conditions, and the L40S has no FP4.

Lifecycle and support

We found no NVIDIA end-of-life notice for either card as of October 2026. NVIDIA’s vGPU software lifecycle page, last updated on 2 October 2026, lists the L40S, the L40 and the RTX PRO 6000 Server Edition as fully supported, and the L40S remains in current vendor guides including the Xeon 6 generation of Lenovo servers.

What we supply

We supply both cards, the L40S and the RTX PRO 6000 Server Edition, on their own for servers you already run or in AI servers built to order, among them the four-card inference node, on one EU contract and invoice with manufacturer warranty. We size the choice from the workload: model size, precision, concurrency and context length, then the chassis with its power and airflow. vGPU and NVIDIA AI Enterprise licences, where the setup needs them, come in the same quote.

FAQ