NVIDIA L40S 48GB
The 48GB Ada workhorse: 18,176 CUDA cores, 864GB/s, FP8 tensor throughput at 733 TFLOPS and 3x NVENC in a 350W dual-slot card. It is the practical entry point to 48GB of professional VRAM for fine-tuning and Omniverse-class rendering, and it is the card most often pulled from decommissioned AI nodes. Used and OEM-pull stock, tested and shipped from Hong Kong.
Fully Tested & Documented · Hong Kong Warehouse Stock · Worldwide Shipping 3–7 Days · 1-Year Warranty
Hardware Gallery
The L40S, Up Close
Official NVIDIA Render
The L40S in its reference form: dual-slot, full-height, passively cooled, fed by a single 16-pin connector. Unlike the low-profile L4 it needs real chassis airflow and a 350W power budget, so confirm your node is on the supported-GPU list before ordering.
Dual-Slot, Full-Height
Angle view showing the full-length board, the bracket end and the finned passive shroud. Because Ada-generation cards have no NVLink and no MIG, scaling out means multiple cards rather than one partitioned card — plan PCIe slots and airflow accordingly.
48
GB GDDR6 with ECC
864
GB/s memory bandwidth
733
TFLOPS FP8 tensor
350
W TDP, dual-slot
Key Specifications
| Component | Specification |
|---|---|
| Model | NVIDIA L40S 48GB PCIe; OEM-branded Dell / Lenovo / HP boards are the same GPU |
| Architecture | Ada Lovelace, 18,176 CUDA cores, 568 fourth-gen Tensor cores, 142 third-gen RT cores |
| Memory | 48GB GDDR6 with ECC, 864 GB/s |
| Interface | PCIe Gen4 x16, dual-slot full-height, passive |
| Power | 350W TDP, 16-pin (CEM5) power connector |
| Compute | FP8 733 TFLOPS / FP16-BF16 362 TFLOPS / TF32 183 TFLOPS / FP32 91.6 TFLOPS |
| Media | 3x NVENC / 3x NVDEC — AV1 encode and decode |
| Virtualization | vGPU supported; NVIDIA AI Enterprise ready |
| NVLink / MIG | No NVLink; MIG not supported on this board |
The L40S sits between the inference-optimised L4 and the HBM-class A100 40GB: it has the GDDR6 capacity for 13B-34B fine-tunes and 4K rendering that a 24GB L4 cannot hold, without HBM pricing. Be sure the host can feed it — our CPU for AI inference server guide covers PCIe lane count and host-CPU selection before you buy.
Transparent Pricing
2026 Reference Pricing by Configuration
Tracked Sep 20, 2026 — 11 eBay asking prices for the L40S 48GB, all used, refurbished or OEM-pull stock. Two further listings were excluded from the range (a Parts Only board and a junk-price listing), and the exclusions are shown below rather than hidden.
Pre-Owned / Refurbished — Core Band
$5,999-$7,450
Seven tracked listings: $5,999.99 Pre-Owned, $6,900 Very Good - Refurbished, $7,300 and three at $7,450. This is where most L40S stock sits — cards pulled from decommissioned nodes, with no OEM-brand premium.
OEM-Branded (Lenovo / HP) — Upper Band
$8,150-$8,588
Two tracked listings at $8,150.87 and $8,588.00, one Lenovo and one HP variant among them (band shown rounded to whole dollars, as the index reports it). Same 48GB board, higher ask — usually because OEM-branded cards come with a documented service history and a purchase channel the buyer already trusts.
Parts Only — Excluded From Range
$3,898.98
A $3,898.98 Parts Only listing was deliberately excluded: incomplete or cannibalised boards are not a comparable product. It is shown here so you can recognise it — if an L40S looks far cheaper than $5,999, read the condition line first.
Asking prices, not transaction prices — volume deals typically close 5-15% below asking. Sources: eBay seller listings (NVIDIA, Lenovo and HP variants; sorted by lowest price plus shipping, first page). Collected Sep 20, 2026. Full dataset: Used Server Price Index.
Origin
Hong Kong Warehouse
Free port — full export documentation
Transit
3–7 Days Air Freight
Tracked, insured, worldwide
Warranty
1-Year Warranty
Diagnostic report in every box
Quality Process
What We Test Before an L40S Ships
Board and vBIOS Verification
Board markings, vBIOS string and device ID checked against the L40S reference — re-badged or flashed cards are a known grey-market risk on Ada-generation boards.
48GB Memory ECC Burn-In
The full 48GB of GDDR6 is exercised with ECC counters logged. On ex-datacenter pulls, memory health degrades before compute does, so any error history is disclosed before you pay.
Sustained-Load Thermal Log
The 350W board runs a sustained load in a 2U chassis with core, memory and hotspot temperatures logged — a passively cooled card in weak airflow throttles long before it fails.
FP8 Tensor Throughput Check
FP8 and FP16 GEMM throughput compared against a reference L40S. Cards that pass POST but deliver 60% of rated tensor throughput are rejected, not discounted.
16-Pin Connector and Provenance Check
Power inlet inspected for heat damage, and the serial traced to its source fleet. OEM-branded pulls ship declared as such, with the OEM part number on the invoice.
Report in the Box
Every unit ships with its per-machine diagnostic report and warranty. Full process: inspection guide.
Frequently Asked Questions
How much is a used NVIDIA L40S 48GB in 2026?
The tracked range is $5,999-$8,588 across 11 eBay asking prices (collected Sep 20, 2026). Most cards sit in the $5,999-$7,450 core band; OEM-branded Lenovo and HP variants ask $8,150-$8,588. A Parts Only listing at $3,898.98 was excluded from the range. Volumes quote below asking.
NVIDIA L40S vs L4 — which Ada card do I need?
Same Ada Lovelace generation, different jobs. The L40S gives you 48GB, 864GB/s, 18,176 CUDA cores, FP8 733 TFLOPS and 350W in a dual-slot full-height card. The L4 gives you 24GB, 300GB/s, 72W and a single-slot low-profile passive card that runs on slot power alone. Buy L40S when the model or the render does not fit in 24GB and when you fine-tune; buy L4 when you are deploying tens of inference cards per rack and perf-per-watt is the whole business case.
Can the L40S serve 70B-class models?
Not comfortably on one card. 48GB of GDDR6 holds 13B-34B fine-tunes and quantized serving in that class well, and lower-precision 70B work is possible with quantization and tight context, but 864GB/s of GDDR6 bandwidth — not capacity — becomes the bottleneck for long-context decode. For sustained 70B serving the answer is a 80GB-class HBM card such as the H100, or a multi-card L40S node with tensor parallelism.
What condition are HKCHL L40S cards in?
Used, refurbished and OEM-pull stock, graded and priced accordingly. Every card ships with a burn-in report (48GB ECC memory test, sustained thermal log, FP8 tensor throughput check) and carries a 1-year warranty. OEM-branded pulls are declared as OEM on the invoice, never sold as NVIDIA retail-boxed.
Does the L40S need a special chassis or power connector?
Yes, on both counts. It is a full-height dual-slot passively cooled card that needs front-to-back server airflow, and it draws 350W through a 16-pin (CEM5) connector — a workstation with a spare x16 slot and a 6-pin cable will not run it. Tell us your chassis and PSU and we will confirm fit, cabling and the supported-GPU list before quoting.
Get a Current NVIDIA L40S Quote
Tell us quantity, whether you need NVIDIA or OEM-branded boards, and the chassis they are going into — we confirm live graded stock, quote volume pricing below the tracked band, and ship worldwide from Hong Kong, usually within one business day.
Request a Quote