How to read this page. Each section starts Simple, then goes to a Deep dive: stop wherever you have what you need. The small numbers are sources: click one to open the original document. Where the maker has not published a figure -- a price, a die size, a factory address -- this page says so rather than estimate it.
1.At a glance
SimpleStart here
The L40S is a single NVIDIA GPU built to do three jobs at once: AI training and inference, 3D graphics rendering, and video processing. NVIDIA announced it 8 August 2023 at SIGGRAPH, calling it “a powerful, universal data center processor” 1.
Deep diveThe technical detail
Architecture Ada Lovelace. 48GB of GDDR6 memory, no HBM, no NVLink — a deliberate trade-off for a GPU meant to slot into standard PCIe servers rather than a dedicated AI training rack 2. Early cloud availability was announced the same day via CoreWeave 1.
2.Launch and history
SimpleStart here
L40S arrived as data centers were retooling for generative AI but most servers still had standard PCIe slots, not the exotic NVLink backplanes that NVIDIA’s top-end training chips need. NVIDIA positioned it as the GPU for that gap: powerful enough for AI, and equally at home rendering 3D scenes or encoding video.
Deep diveThe technical detail
Announced 8 August 2023 alongside a wave of NVIDIA OVX systems from ASUS, Dell, GIGABYTE, HPE, Lenovo, QCT and Supermicro 1. NVIDIA’s VP of Professional Visualization, Bob Pette, framed it as covering “AI, graphics and video processing workloads” in one card 1. It quickly became one of the most widely stocked GPUs across independent GPU clouds, precisely because it needs no special rack infrastructure.
3.What’s inside it
SimpleStart here
L40S uses NVIDIA’s Ada Lovelace architecture — the same generation as GeForce RTX 40-series gaming cards — with third-generation RT cores for ray tracing and fourth-generation Tensor cores for AI 2.
Deep diveThe technical detail
142 third-gen RT cores, 568 fourth-gen Tensor cores, 18,176 CUDA cores 2. NVIDIA’s Ada architecture whitepaper describes fourth-gen Tensor cores as delivering “more than 2x” the FP16/BF16/TF32/INT8/INT4 throughput of the prior Ampere generation, plus Shader Execution Reordering and an 8th-generation NVENC encoder with AV1 support 3. NVIDIA’s L40S materials do not themselves name the underlying die; the whitepaper covers the Ada family generally rather than confirming L40S’s exact die by name.
4.Full spec table
SimpleStart here
18,176 CUDA cores, 48GB of GDDR6 memory (not HBM), a 350W power limit, and no NVLink — it connects to a server over a standard PCIe Gen4 slot 2.
Deep diveThe technical detail
| Spec | NVIDIA L40S |
|---|
| CUDA cores | 18,176 |
|---|
| RT cores | 142 (3rd gen) |
|---|
| Tensor cores | 568 (4th gen) |
|---|
| Memory | 48GB GDDR6 with ECC |
|---|
| Memory bandwidth | 864 GB/s |
|---|
| FP32 | 91.6 TFLOPS |
|---|
| TF32 Tensor (dense/sparse) | 183 / 366 TFLOPS |
|---|
| FP16 Tensor (dense/sparse) | 362 / 733 TFLOPS |
|---|
| FP8 Tensor (dense/sparse) | 733 / 1,466 TFLOPS |
|---|
| INT8 Tensor (dense/sparse) | 733 / 1,466 TOPS |
|---|
| Interconnect | PCIe Gen4 x16 (no NVLink, no MIG) |
|---|
| Form factor | Dual-slot, passive cooling, 4.4" x 10.5" |
|---|
| Power | 350W max, 16-pin connector |
|---|
| Video | 3x NVENC (incl. AV1) / 3x NVDEC, 4x DisplayPort 1.4a |
|---|
Figures per NVIDIA’s own L40S product page 2. Sparse figures use structural sparsity; dense figures are roughly half.
5.Where it’s made
SimpleStart here
NVIDIA does not state a fab or process node on the L40S product page itself. Its Ada architecture whitepaper says Ada GPUs are built on a custom TSMC process called 4N — but that is an architecture-family statement, not an L40S-specific confirmation 3.
Deep diveThe technical detail
NVIDIA’s L40S datasheet does not disclose a die name, fab, or process node. The closest official figure is architecture-level: the Ada whitepaper states Ada Lovelace GPUs use “TSMC’s custom 4N process” 3. This page does not assert a specific die (such as AD102) for L40S, because NVIDIA has not confirmed one by name.
6.Which systems use it
SimpleStart here
L40S ships in standard servers, not a dedicated NVIDIA system. NVIDIA named ASUS, Dell, GIGABYTE, HPE, Lenovo, QCT and Supermicro as OVX system partners at launch 1.
Deep diveThe technical detail
Dell’s own PowerEdge GPU compatibility matrix confirms L40S support across R760xa, R770, R7625, R7615, R660, R650, R750, XE7745 and other models, with the maximum GPU count varying by chassis 4. On the cloud side, CoreWeave, AWS (EC2 G6e) and Oracle Cloud Infrastructure (BM.GPU.L40S.4 bare-metal shape) each offer L40S instances; Google Cloud’s equivalent G2 tier uses the smaller L4 GPU, not L40S.
7.Official pricing
SimpleStart here
NVIDIA does not publish a unit price for L40S. CoreWeave’s own pricing page lists an 8-GPU L40S node at $18.00/hour on-demand — $2.25/hour per GPU — or $7.88/hour spot 5.
Deep diveThe technical detail
CoreWeave’s official pricing page (checked September 2026) lists on-demand L40S at $18.00/hour for an 8-GPU node ($2.25/hour per GPU) and spot pricing at $7.88/hour, the same rate quoted for both its North America and Europe regions 5. AWS confirms its EC2 G6e instances use “up to 8 NVIDIA L40S Tensor Core GPUs,” and Oracle Cloud lists an official L40S bare-metal shape, but neither publishes a simple headline hourly rate on the page this site cites, so this page links to CoreWeave’s figure rather than an unconfirmed one.
9.What came before, what came next
SimpleStart here
Came before: the L40, and NVIDIA names the A40 as the “previous-generation” comparison point for L40S specifically 2. Came after: NVIDIA’s own blog calls the RTX PRO 6000 Blackwell Server Edition the successor 6.
Deep diveThe technical detail
L40S’s direct sibling is the earlier L40 (300W, launched January 2023 7) — L40S is the higher-power (350W), AI-optimized variant of the same class of card. NVIDIA’s own product page names the A40 as the generation L40S is compared against 2. Going forward, NVIDIA’s own announcement blog states plainly: “Compared to the previous-generation NVIDIA Ada Lovelace architecture L40S GPU, the RTX PRO 6000 Blackwell Server Edition GPU will deliver a multifold increase in performance” 6. NVIDIA does not name the T4 as a predecessor to L40S anywhere in its own materials; that lineage runs through the separate L4 product line instead.
10.Hidden in plain sight
SimpleStart here
A GPU NVIDIA markets mostly for graphics, video and enterprise AI sits, by U.S. federal filing, in the very same export-control tier as NVIDIA’s flagship H100 AI-training chip.
Deep diveThe technical detail
NVIDIA’s own 10-K filing for fiscal year 2024 names L40S in the same sentence as A100, A800, H100 and H800 as a product subject to U.S. export-licensing requirements for China and certain other country groups: “exports to China and Country Groups D1, D4, and D5 … of our products exceeding certain performance thresholds, including A100, A800, H100, H800, L4, L40, L40S and RTX 4090” 8. A card sold as a “universal” graphics-and-inference GPU turns out to be legally classified alongside NVIDIA’s most powerful AI-training silicon.