L40

The L40 is a Ada Lovelace-generation NVIDIA GPU built on the AD102 architecture, manufactured on a TSMC 4N process node with 76.3 billion transistors. Released in 2022, it delivers 181 TFLOPS of FP16 throughput and 181 TFLOPS BF16 — the two precision formats most commonly used for transformer model training and inference. The AD102 architecture represents NVIDIA's approach to balancing compute throughput, memory bandwidth, and power efficiency for data center AI workloads. At 300W TDP, the L40 sits in the mid-range data center tier (300W), fitting standard GPU server form factors.

Memory capacity is 48GB of GDDR6 with 864 GB/s bandwidth. This determines which models can run without quantization: approximately 24B parameters at FP16 (2 bytes/param), 48B at INT8 (1 byte/param), or up to 96B parameters at INT4/GGUF quantization (0.5 bytes/param). These figures are theoretical maximums — actual capacity is reduced by KV cache, framework overhead, and activation memory, typically by 10–20% for inference and 30–40% for training. The arithmetic intensity ceiling is approximately 209 FLOP/byte (181 TFLOPS ÷ 864 GB/s). Most autoregressive LLM inference falls well below this threshold, making the 864 GB/s memory bandwidth the binding constraint on tokens-per-second throughput rather than raw TFLOPS.

The L40 uses PCIe 4.0 for host connectivity. Without NVLink, VRAM cannot be pooled across multiple cards — the single-card 48GB capacity is the hard ceiling for model size without model sharding over slower PCIe. For workloads that exceed 48GB, the alternative is pipeline parallelism (splitting model layers across cards) rather than tensor parallelism, which introduces inter-card communication overhead at each layer boundary. This makes the L40 best suited for workloads that fit within a single card's VRAM budget.

The primary workloads for the L40 are Visualization, Inference, Virtual workstations. Predecessor to L40S. Similar specs, slightly lower clocks. Good availability and pricing. Key limitations to factor into your evaluation: GDDR6 memory bandwidth (864 GB/s) far below HBM alternatives; No NVLink — single-card VRAM ceiling; Superseded by L40S — lower clock speeds. When comparing this GPU against alternatives at similar price points, the most important metrics are memory bandwidth (for inference throughput), VRAM capacity (for model size), and FP16/BF16 TFLOPS (for training speed). Raw TFLOPS figures can be misleading for inference — a GPU with lower TFLOPS but higher memory bandwidth will often outperform a higher-TFLOPS card on tokens-per-second for memory-bound autoregressive generation.

In the broader GPU market, the L40 is a 4-year-old design that is increasingly being replaced by newer architectures in cloud deployments, though it remains available at competitive rental rates. CUDA compatibility is a significant advantage: the L40 benefits from the full NVIDIA software ecosystem including cuDNN, TensorRT, FlashAttention, and all major inference frameworks (vLLM, TGI, TensorRT-LLM). CUDA's maturity means optimized kernels are available for virtually every model architecture. For cloud rental, availability varies significantly by provider — some specialize in this GPU tier while others may have limited stock. Compare on-demand and spot pricing across providers using the rental comparison table on this page, and factor in region availability if latency is a concern for your inference workload.