The NVIDIA L40S is the data-centre successor to the A40, built on the full Ada Lovelace die with GDDR6 ECC memory and a 900W ADA900 board. It combines inference throughput (FP8 Tensor Cores at up to 1,457 TOPS sparse), professional media (4× NVENC AV1 encode, 2× NVDEC) and ECC-protected 48 GB into a PCIe passive card designed for 24/7 rack duty.
packet.ai rents the L40S as a Dedicated single-tenant card for $0.92/hr in Virginia, billed hourly, 99% SLA. At roughly half the A100's price it is the obvious card for 13–34B serving, FP8-quantised 70B, SDXL and video generation pipelines that need more than 24 GB. Dynamic pricing is launching soon.
Why rent the L40S on packet.ai
48 GB ECC GDDR6 at 864 GB/s
Double the VRAM of a 4090, ECC-protected, with enough bandwidth for 13B and 34B models at FP16 and 70B at INT4. A single L40S can serve Llama 3.1 70B in 4-bit with a comfortable KV-cache headroom.
FP8 Tensor Cores
Ada's fourth-generation Tensor Cores natively process FP8, bringing roughly 2× the throughput of FP16 at the same batch size. For FP8-quantised inference the L40S trades bandwidth for compute, and does so more cheaply than an A100 or H100.
Professional media pipeline
4× NVENC (9th gen, AV1) and 2× NVDEC unlock hardware-accelerated video generation and processing at scale. For video diffusion models that output at real-time or faster, this matters.
ECC and passive cooling
Consumer cards (4090) have neither; the L40S is built for extended, unattended rack duty with error-correcting memory and passive dissipation. Fewer silent corruptions, lower tail-latency variance.
37% below the field
$0.92/hr Dedicated on packet.ai is 11% below RunPod Secure Cloud's $1.03, 37% below CoreWeave's $1.46 and 63% below Vast.ai Secure at $2.50. Only anonymous marketplace listings go lower, without SLA or single-tenancy.
L40S pricing on packet.ai
Three ways to buy the same card. Mix them on one account: Dynamic for development and bursty work, Dedicated for production, Clusters when a single node is not enough.
Dynamic
Shared card, scheduler-enforced isolation. Same peak performance and VRAM as Dedicated.
- Hourly billing, no commitment
- SSH-ready in under 5 minutes
- Never pre-empted
Dedicated
Whole card, exclusively yours. Zero scheduler interference, predictable p99.
- 99% SLA
- Single-tenant hardware
- 5–10 minute provisioning
Clusters
Multi-node, InfiniBand fabric, custom storage, named TAM.
- Bare-metal nodes and 8× configurations
- Up to 1,024+ GPUs
- 2–6 week provisioning
How it compares
Published on-demand starting rates for the same GPU at other providers, USD per GPU-hour. packet.ai's reference rate is $0.92/hr Dedicated.
| Provider | L40S rate |
|---|---|
| packet.ai | $0.92/hr Dedicated |
| RunPod Secure Cloud | $1.03/hrpacket.ai −11% |
| Lambda | $1.10/hrpacket.ai −16% |
| Hyperstack | $1.30/hrpacket.ai −29% |
| CoreWeave | $1.46/hrpacket.ai −37% |
| Vast.ai (Secure) | $2.50/hrpacket.ai −63% |
Specifications
| Feature | Value |
|---|---|
| GPU architecture | NVIDIA Ada Lovelace (AD102) |
| CUDA cores | 18,176 |
| Tensor Cores | 568 (4th generation, FP8 / FP16 / BF16 / TF32) |
| RT Cores | 142 (3rd generation) |
| Memory | 48 GB GDDR6 with ECC |
| Memory interface | 384-bit |
| Memory bandwidth | 864 GB/s |
| FP32 compute | 91.6 TFLOPS |
| FP16 / BF16 Tensor compute | 362 TFLOPS dense / 733 TFLOPS sparse |
| FP8 Tensor compute | 733 TFLOPS dense / 1,457 TFLOPS sparse |
| Interconnect | PCIe 4.0 ×16 (no NVLink) |
| Media engines | 4 × NVENC (9th gen, AV1), 2 × NVDEC (6th gen) |
| Board power | 350 W |
| Form factor | Dual-slot, passive cooling |
What you can run on 48 GB
Where the L40S is the right card, where it is comfortable, and where you should step up.
| Workload | Fit | Notes |
|---|---|---|
| 13–34B LLM inference at FP16 | Single card | Qwen 2.5 14B, 32B; Llama 3.1 8B and 13B |
| 70B LLM inference (FP8 / 4-bit) | Single card at FP8; 2 cards at FP16 | Llama 3.1 70B, Qwen 72B |
| SDXL / FLUX image generation | Full precision plus LoRA stacks | 4× NVENC AV1 for video output |
| Video diffusion (Wan, CogVideo) | Large frames fit, AV1 hardware encode | Better than any GDDR card for video IO |
| Embedding and reranking | Very large batches | BGE, E5, ColBERT at high QPS |
| LoRA / QLoRA fine-tuning to 34B | Single card | Hourly billing suits 4–40 hour runs |
| 70B+ full FP16 / pre-training | Not enough VRAM or bandwidth | Use A100 or B200 |
Deploy in minutes
- Create an accountSign up at dash.packet.ai. No credit card required to start; billing is usage-based.
- Pick the L40S and a tierChoose Dedicated (Dynamic launching soon), hourly or monthly billing, and the region. Multi-GPU and bare metal are on request.
- DeployFrom the dashboard, the API, or the packet CLI. Dedicated instances are typically SSH-ready in five to ten minutes.
- SSH in and runUbuntu, current NVIDIA drivers and CUDA are in place, with local NVMe attached. Bring your containers or install directly.
# after the instance shows Running in the dashboard ssh ubuntu@<instance-ip> nvidia-smi # confirm the L40S and driver # example: serve a model with vLLM pip install vllm python -m vllm.entrypoints.openai.api_server --model <hf-model-id> --port 8000
CLI and API reference, including how to launch and stop instances programmatically, is in the packet.ai docs.
FAQ
How much is the L40S on packet.ai?
$0.92 per GPU-hour on Dedicated capacity, billed hourly, no minimum term. About $672 per month at 730 hours; monthly commits take up to 20% off. Dynamic pricing is launching soon.
Is the L40S the same as the RTX 6000 Ada?
Similar die but different product: both use the AD102 at different TDPs and configurations. The L40S is the server SKU with passive cooling and ECC, slightly higher FP32 at 91.6 vs 91.1 TFLOPS, and the same 48 GB GDDR6 ECC. For cloud GPU purposes they are functionally equivalent; the L40S on packet.ai is the server-grade variant.
How does the L40S compare to the A100 80GB?
The L40S has more CUDA cores and higher FP32 throughput, but less VRAM (48 GB vs 80 GB) and much lower memory bandwidth (864 GB/s vs 2,039 GB/s). For batch inference and generation the L40S often wins on throughput per dollar; for memory-bound serving at high concurrency or models that need 80 GB, the A100 wins on latency.
Can I run 70B models?
Yes, in FP8 or 4-bit on a single L40S (model weights fit in 48 GB at those precisions). For FP16 you need two cards. The FP8 path on Ada is native and fast; FP8 70B inference is the L40S sweet spot at this price.
Which region?
US East (Virginia) today. Use the pricing page region filter for the live list.
What is pre-installed?
Ubuntu with current NVIDIA drivers and CUDA, SSH access and local NVMe. Bring your containers or install frameworks directly.
Hidden fees?
No platform fees, no ingress. Egress $0.04/GB. Up to 2 TB local NVMe included; object storage from $0.018/GB-month.
Same silicon. Smarter economics.
L40S from $0.92/hr, billed hourly, live in US East · Virginia. No quota approval, no credit card to start.