deploybase

L40S on CoreWeave: Pricing, Specs & How to Rent

Deploybase · April 17, 2025 · GPU Pricing

Contents

L40S GPU Specifications

The NVIDIA L40S represents a professional-grade GPU optimized for graphics, visualization, and AI inference tasks. With 48GB of GDDR6 memory, the L40S serves different workload patterns than compute-focused alternatives like A100 or H100 GPUs.

Technical specifications:

  • 48GB GDDR6 memory
  • 18,176 CUDA cores
  • 864 GB/s memory bandwidth
  • 366 TFLOPS TF32 peak performance (91.6 TFLOPS FP32)
  • Real-time ray tracing capabilities
  • Tensor Float 32 (TF32) support
  • PCIe 4.0 interconnect (no NVLink)
  • Maximum power consumption: 350W
  • Supports NVIDIA CUDA and cuBLAS libraries

The L40S excels at rendering, graphics, video processing, and inference tasks requiring high memory capacity. Unlike datacenter-only GPUs, the L40S maintains legacy support for professional graphics workflows.

CoreWeave L40S Pricing

CoreWeave specializes in bundled GPU clusters, typically offering systems with multiple cards rather than individual GPU rentals. Their pricing model emphasizes stable, production-ready infrastructure.

L40S pricing on CoreWeave:

  • 8x L40S cluster: $18/hour as of March 2026
  • Per-GPU cost: approximately $2.25/hour
  • No setup fees
  • Dedicated infrastructure provisioning
  • Fixed-rate billing with no surprises
  • Support for long-term commitments with discounts

CoreWeave's approach differs from spot markets like Vast.ai GPU pricing, prioritizing stability and predictability. The bundled 8-GPU configuration targets professional rendering studios and large-scale inference deployments.

Compare this to RunPod L40S pricing at $0.79/hour for single GPU access, which costs substantially less for smaller workloads. CoreWeave suits teams requiring production reliability at the expense of per-unit cost.

How to Rent L40S on CoreWeave

Renting L40S GPUs through CoreWeave involves account setup and cluster configuration:

  1. Create CoreWeave account
  2. Complete verification process
  3. Add payment method
  4. Access the CoreWeave dashboard
  5. Select L40S GPU cluster (8x minimum)
  6. Choose region for deployment
  7. Configure storage (local NVMe or persistent block storage)
  8. Set up networking (public or private)
  9. Deploy container or custom image
  10. Access via SSH or web console

CoreWeave provides pre-built templates for common workloads like rendering engines, LLM inference servers, and video processing pipelines. The platform handles driver installation and CUDA optimization automatically.

Most deployments begin with container images from Docker Hub or custom container registries. CoreWeave manages underlying infrastructure, allowing users to focus on application logic.

L40S vs. Other GPUs

The L40S fits a specific niche within the GPU ecosystem. Comparing to alternatives clarifies when to choose this option.

Vs. A100: A100 specifications focus on tensor performance for training. L40S offers higher memory bandwidth and better graphics support. A100 is better for training, L40S for inference and rendering.

Vs. H100: H100 specifications represent the top performance tier for training. L40S trades compute for memory efficiency, making it more cost-effective for inference-only workloads.

Vs. L40: The original L40 contains 48GB GDDR6. The L40S adds better ray tracing and minor performance improvements. L40S is preferred for new deployments.

Vs. RTX 4090: Consumer-grade RTX 4090 lacks the L40S's reliability certifications and professional support. L40S is built for datacenter deployment.

teams running mixed inference and rendering workloads benefit from L40S cost-efficiency. High memory capacity (48GB) handles large batch sizes or big models without overflow.

Real-World L40S Performance

L40S performance varies by workload type, with particularly strong results in inference scenarios.

Inference benchmarks:

  • BERT base: 800 sequences/second per GPU
  • LLaMA 13B: 25-35 tokens/second per GPU
  • Stable Diffusion: 8-12 images per minute at 512x512
  • ONNX Runtime: 3-5x faster than CPU inference

Video processing:

  • H.264 encoding: 4K real-time (30fps) per GPU
  • Video decoding: hardware-accelerated, multiple streams
  • NVIDIA NVENC: fully supported for live streaming

Rendering (graphics workloads):

  • Ray tracing: 60-100 FPS at 1440p in professional apps
  • Path tracing: 30-50 FPS for complex scenes
  • Real-time visualization: sub-5ms latency in most cases

Memory advantages matter for inference at scale. The 48GB capacity allows loading larger models or processing bigger batches than smaller-memory GPUs, sometimes eliminating need for multi-GPU setups.

FAQ

Is L40S good for LLM inference? Yes, L40S works well for inference. The 48GB memory handles models up to 70B parameters without quantization. Consider fine-tuning guide for optimization techniques.

How many L40S GPUs do I need? Depends on throughput requirements. Single L40S handles 25-40 requests/second for typical LLM inference. 8-GPU cluster handles production-scale traffic.

Can I use L40S for training? L40S is not optimized for training. Its tensor performance lags A100/H100. Use L40S only for inference and graphics tasks.

Does CoreWeave offer smaller L40S configurations? CoreWeave typically bundles 8 L40S GPUs minimum. For single-GPU rental, check RunPod GPU pricing or spot markets.

What's the advantage of CoreWeave over spot markets? CoreWeave provides dedicated, production-grade infrastructure with SLA guarantees. Spot markets offer lower cost but with availability risks.

Sources