Eight NVIDIA L40S or eight NVIDIA RTX PRO 4500 Blackwell Server Edition in one server

Which eight-card server to choose

With eight cards in each server, eight NVIDIA L40S give 384 GB of GPU memory at 2.8 kW of board power and eight NVIDIA RTX PRO 4500 Blackwell Server Edition give 256 GB at 1.32 kW. Choose the L40S server when each card must hold more, a 32B model in FP8 or long prompts for many users, when vision throughput per server counts or when the VDI estate runs on the long-term vGPU branch. Choose the Server Edition server for MIG isolation, FP4, more work per watt and a server that stays on one 230 V 16 A feed.

NVIDIA L40S and RTX PRO 4500 Blackwell Server Edition product pages, read on 24 October 2026; NVIDIA vGPU documentation index for the branches; usable memory, sums and wall power are our arithmetic, explained below.

The single cards are compared in our article on the L40S and the RTX PRO 4500 Server Edition, which also compares four L40S with eight Server Edition cards. This page compares eight against eight, as in a 4U server that lists either card.

384 GB at 2.8 kW or 256 GB at 1.32 kW

Our users-per-card rule takes 90 per cent of the memory the driver reports, less 3 GiB per card, for weights and KV cache. For the L40S we scale the 97,887 MiB a 96 GB card reports to about 47.8 GiB and deduct the 6.25 per cent that NVIDIA's CUDA Best Practices Guide gives for ECC on GDDR memory, which leaves 37.3 GiB per card, or 40.0 GiB with ECC off. For the Server Edition the same scaling gives 31.9 GiB and 25.7 GiB usable; NVIDIA's vPC sizing guide states that Blackwell's GDDR7 has ECC built in that is "always enabled without reducing usable VRAM". Eight cards make 298.4 GiB against 205.6 GiB.

Our estimates by the users-per-card rule, not measurements: concurrent conversations of 8,192 tokens with an FP8 KV cache; L40S with ECC on, in brackets ECC off. Server Edition figures as in our page on eight RTX PRO 4500 Server Edition. NVFP4 runs on the L40S only as weight-only compression.

The extra 16 GB per card changes what fits. A 32B model in FP8, 32.0 GiB of weights, fits one L40S with cache for five conversations and does not fit a 32 GB card, so on the Server Edition it runs as pairs. Llama 3.3 70B in FP8 needs four Server Edition cards per copy and leaves 35.1 GiB for cache, where four L40S leave 81.5 GiB; two such copies hold about 130 conversations on the L40S and 56 on the Server Edition. With the example values of our guide to a private ChatGPT server by company size, 80 requests in flight correspond to about 2,000 staff and 40 to about 1,000, so the L40S server covers the larger company with that model and the Server Edition server the smaller one.

The cost of the memory is power. Per GB of GPU memory the L40S draws about 7.3 W of board power and the Server Edition about 5.2 W, by our arithmetic. For one user both cards read the weights at similar speed: 864 against 800 GB/s per card.

Per-card results from StorageReview, 14 October 2026

StorageReview tested one RTX PRO 4500 Server Edition, one L40S and other cards with vLLM 0.27.1 (NGC container 26.08) in a Dell PowerEdge R770 with two Xeon 6787P and published the results on 14 October 2026. The table shows output tokens per second at 32 concurrent requests.

StorageReview, NVIDIA RTX PRO 4500 Blackwell Server Edition review, 14 October 2026, read on 24 October 2026; 512/512 and 8K/1K are input and output tokens per request. StorageReview did not meter power; its efficiency figures divide by each card's rated limit, 165 W and 350 W.

At short prompts the two cards are level. With 8,192 input tokens the Server Edition flattens at about 180 tokens per second on the 14B model from 8 concurrent requests, while the L40S climbs to about 230 at 16, and StorageReview wrote that "The flattening is KV cache running out, and 32GB sets that limit". That is what our rule predicts: the 48 GB card keeps more cache per copy.

The NVFP4 row runs the other way from what FP4 Tensor Cores suggest. StorageReview explains that on the L40S vLLM dequantises NVFP4 weights and computes in higher precision, and that "The NVFP4 path in this vLLM build does not scale with memory bandwidth on any of them". Its matrix-multiply test gives the Server Edition 568.6 TFLOPS in NVFP4, 1.5 times the L40S's 368.6 TFLOPS in FP8, so the FP4 advantage depends on serving software that uses it. In BF16 and FP8 the L40S scored higher, 220.0 and 368.6 TFLOPS against 135.9 and 280.1, which counts for long prompts and fine-tuning.

By our arithmetic eight independent copies of the 14B model would reach at most about 8,856 output tokens per second on the Server Edition server and 7,960 on the L40S server at short prompts, and about 1,440 against 1,840 at 8K, if the host does not become the limit. These are upper bounds from one card; StorageReview tested no multi-card setup. Our round-ups of RTX PRO 4500 benchmarks and L40S benchmarks collect the other published figures.

We build both servers to order; send us your models, prompt lengths and peak requests, and our team replies with the card that fits and the conversations per layout.

Vision models and video engines

Both cards carry three NVENC and three NVDEC engines, 24 of each per server. NVIDIA states that the Blackwell encoders add 4:2:2 H.264 and HEVC encoding, which the L40S lacks. For analysis, NVIDIA's DeepStream 9.2 documentation lists both cards in one table, measured end to end with TensorRT.

NVIDIA DeepStream 9.2 documentation, performance of dGPU pretrained models, one GPU each, read on 24 October 2026.

Per card the L40S delivers 1.5 to 1.7 times the frames per second. Per rated watt the order turns: 1,114 ÷ 350 gives 3.2 PeopleNet frames per second per watt on the L40S and 741 ÷ 165 gives 4.5 on the Server Edition, by our arithmetic. Eight cards give 8,912 against 5,928 frames per second, about 297 against 198 cameras analysed at 30 frames per second before headroom. Our guide to a GPU server for video analytics turns these figures into camera counts with decoder limits.

MIG and vGPU on eight cards

The L40S has no MIG and is shared only by time-sliced vGPU. The Server Edition splits into two MIG instances of 16 GB, so eight cards give 16 isolated GPUs, each with its own memory, compute and one video decoder and encoder. A department, a student lab or a customer then has its own share of memory and compute, which time slicing on the L40S does not give.

NVIDIA Virtual GPU Types Reference (29 September 2026) for L40S maxima per card; NVIDIA vPC sizing guide (19 August 2026) for the Server Edition; NVIDIA AI Enterprise vGPU types for compute VMs; NVIDIA vGPU documentation index for branch end dates. All are maximum densities; per-server totals are our arithmetic.

At the same 2 GB profile, eight L40S host half as many desktops again as eight Server Edition cards, and NVIDIA's vPC sizing guide marks 1 GB profiles "Not supported on Blackwell". Per desktop the Server Edition draws less: 1,320 W ÷ 128 is about 10 W of board power, 2,800 W ÷ 192 about 15 W. The branch matters as much as the count. The Server Edition runs vGPU only from release 20, whose R595 branch NVIDIA supports until March 2027, so a host bought now moves to a later branch within its first year; the L40S also runs on vGPU 19, the long-term branch supported until July 2028. Desktop licences are sold per concurrent user or per GPU, and NVIDIA AI Enterprise is licensed per GPU, so either server needs eight. Our guides to L40S vGPU profiles and MIG and vGPU seats per card give the details.

Power, feeds, slots and chassis

With two AMD EPYC 9355 at 280 W each and 15 per cent for the rest of the server, we estimate about 3.9 kW at the wall for eight L40S (2,800 + 560 W, plus 15 per cent) and about 2.2 kW for eight Server Edition cards (1,320 + 560 W, plus 15 per cent). We plan a 230 V 16 A feed at 80 per cent, 2.94 kW. The Server Edition server stays on one such feed per side; the L40S server needs a 32 A single-phase or a three-phase feed. NVIDIA's product brief lets the Server Edition be capped between 100 and 165 W, about 1.6 kW at the wall at 100 W, while for the L40S we found no minimum board power in NVIDIA's brief. By the 160 CFM per kW of our guide to GPU rack power and cooling, the two servers need roughly 620 and 350 CFM of air.

Cisco UCS C845A M8 data sheet (24 August 2026), Dell XE7745 technical guide (Rev. A09, September 2026, table 23, which lists the single-width "NVIDIA RTX Pro 4500" without naming the edition), Supermicro RTX PRO Blackwell Server Edition systems page and Puget Systems configurator, read on 24 October 2026.

The L40S is a dual-slot board on PCIe 4.0 x16, the Server Edition a single-slot board on PCIe 5.0 x16, both passive and full length with one 16-pin connector. Eight Server Edition cards therefore leave slot space for network cards and drives, and in a Gen5 server a model split over cards gets twice the link bandwidth per card. Neither card has NVLink. Supermicro's largest 2U listing for the Server Edition is six cards, so neither eight-card server is a 2U machine in the listings we read. The maker's configurator decides for any chassis. We check the feed, the inlet temperature and the chassis listing before we quote; tell us the rack position and feed, and we say which of the two servers fits it.

Recommendation per use case

Our reading of the figures above.

For one or two 70B-class models for many users, neither server is the best fit: four RTX PRO 6000 Server Edition cards hold one copy per card without splitting a copy over four cards, as our page to buy a 4× RTX PRO 6000 Server Edition server shows, and our comparison of eight RTX PRO 4500 or four RTX PRO 6000 Server Edition sets that layout against the eight small cards.

What we supply

You can buy either system from us: our team sizes the configuration to your workload and budget and replies within one business day. We build servers with eight L40S or eight RTX PRO 4500 Server Edition cards to order in chassis whose makers list them, assembled and tested before delivery, on one EU contract and invoice with manufacturer warranty on every component; NVIDIA vGPU and AI Enterprise licences come on the same invoice. The configuration and ordering steps are on our pages to buy an L40S server and to buy an 8× RTX PRO 4500 Server Edition server, and what each runs on eight L40S and eight RTX PRO 4500 Server Edition.

FAQ