XENON and NVIDIA: Certified across the entire stack

The NVIDIA L4 Tensor Core GPU is a versatile, energy-efficient accelerator for AI inference, video and virtual workstations, in the cloud, in the data centre and at the edge.

• NVIDIA Ada Lovelace architecture in a 72W, single-slot, low-profile PCIe card that fits virtually any server
• 24GB GDDR6 memory with 300GB/s bandwidth and PCIe Gen4 connectivity
• Up to 120x higher AI video performance than CPU-based infrastructure, with up to 99% better energy efficiency
• Full AV1 encode and decode (2x NVENC, 4x NVDEC, 4x JPEG decoders): over 1,000 concurrent 720p30 AV1 streams per server
• Complete NVIDIA enterprise stack support: NVIDIA AI Enterprise plus vGPU software for virtual PCs and RTX Virtual Workstations

The NVIDIA L40S is the most powerful universal GPU for the data centre, accelerating everything from generative AI inference and model fine-tuning to 3D graphics, rendering, Omniverse and video applications on a single card.

• NVIDIA Ada Lovelace architecture with 18,176 CUDA cores, 568 fourth-generation Tensor Cores and 142 third-generation RT Cores
• 1,466 TFLOPS of FP8 Tensor performance (with sparsity) with the FP8 Transformer Engine for generative AI inference and fine-tuning
• 48GB GDDR6 memory with ECC and 864GB/s bandwidth
• 91.6 TFLOPS FP32 and 212 TFLOPS ray tracing performance for rendering, digital twins and video workloads
• Standard dual-slot 350W PCIe Gen4 form factor, deployable in mainstream NVIDIA-Certified servers without specialised infrastructure

The NVIDIA H200 supercharges generative AI and HPC with the Hopper architecture and the largest, fastest memory of its generation: the first GPU with HBM3e.

• 141GB of HBM3e memory at 4.8TB/s: nearly double the capacity of the H100 with 1.4x more memory bandwidth
• Up to 3,958 TFLOPS of FP8 Tensor performance (SXM, with sparsity) for large language model inference and training
• Up to 1.9x faster Llama 2 70B inference and 1.6x faster GPT-3 175B inference than the H100, within the same power profile
• Fourth-generation NVLink with 900GB/s of GPU-to-GPU bandwidth, plus up to 7 Multi-Instance GPU (MIG) partitions and confidential computing
• Available as H200 SXM for HGX systems or H200 NVL, a dual-slot air-cooled PCIe card for mainstream enterprise servers, with a 5-year NVIDIA AI Enterprise subscription included

Powered by the NVIDIA Hopper architecture, the H100 delivers proven, order-of-magnitude leaps in AI performance and remains the workhorse GPU for large-scale training, inference and HPC.

• Transformer Engine with FP8 precision: up to 30x faster AI inference and up to 4x faster training on large language models than the prior-generation A100
• Up to 3,958 TFLOPS of FP8 Tensor performance (H100 SXM, with sparsity) and 34 TFLOPS of FP64 for HPC
• 80GB HBM3 at 3.35TB/s (SXM), or 94GB HBM3 at 3.9TB/s in the PCIe air-cooled H100 NVL
• Fourth-generation NVLink with 900GB/s of GPU-to-GPU bandwidth (SXM) for multi-GPU scale
• Up to 7 Multi-Instance GPU (MIG) partitions, and the first GPU with built-in confidential computing; H100 NVL includes a 5-year NVIDIA AI Enterprise subscription

The universal data centre GPU of the Blackwell generation: one card for agentic AI, physical AI, scientific computing, rendering and video, and NVIDIA’s designated successor to the L40S with up to 6x faster inference.

• NVIDIA Blackwell architecture with 24,064 CUDA cores, 752 fifth-generation Tensor Cores and 188 fourth-generation RT Cores
• Up to 4 petaFLOPS of FP4 AI performance (with sparsity) via the second-generation Transformer Engine, for multimodal and agentic AI inference
• 96GB GDDR7 memory with ECC and 1.6TB/s bandwidth for large models and massive 3D scenes
• Universal MIG: up to 4 fully isolated 24GB instances running AI and graphics concurrently, with confidential computing and secure boot
• Passive dual-slot PCIe Gen5 design, configurable up to 600W, available now in NVIDIA-Certified servers from 1 to 8 GPUs