NVIDIA GPUAI Computing
From professional graphics to data centers, from edge computing to AI inference, NVIDIA's full product portfolio delivers powerful computing power for every industry.
NVIDIA DGX B200
The NVIDIA DGX B200 is an enterprise-grade unified AI platform powered by 8 NVIDIA Blackwell GPUs connected via fifth-generation NVLink. It delivers 3x training performance and 15x inference performance compared to the previous-generation DGX H100, featuring FP4 compute of 144 PFLOPS (sparse) and a total GPU memory of 1,440GB HBM3e. Designed to support diverse workloads—including large language models, recommendation systems, and chatbots—it serves as the foundational building block for DGX BasePOD and DGX SuperPOD solutions. Preloaded with the NVIDIA AI software stack, it enables end-to-end AI transformation from development to deployment.
NVIDIA DGX B300
The NVIDIA DGX B300 is an AI infrastructure platform designed for generative AI and large model inference. It features 8 NVIDIA Blackwell Ultra SXM GPUs, delivering up to 144 PFLOPS of FP4 inference performance and 2.1 TB of total GPU memory. Compared to the DGX B200, it offers 1.5× denser FP4 compute performance and 2× faster attention processing. The new 10U chassis supports deployment in NVIDIA MGX racks for the first time, with AC/DC power options and pre-installed NVIDIA AI Enterprise full-stack software—enabling enterprises to run AI workloads at the scale of hyperscale cloud providers.
NVIDIA DGX GB300
The NVIDIA DGX GB300 is a rack-scale liquid-cooled AI system built on the NVIDIA Grace Blackwell Ultra superchip. It integrates 72 Blackwell Ultra GPUs and 36 Grace CPUs, delivering 2,592 Arm Neoverse V2 cores, up to 20TB of GPU memory (with bandwidth up to 576TB/s), and 37TB of total fast memory. With FP4 Tensor Core performance reaching 1440 PFLOPS (sparse), it offers up to 1.5x denser FP4 performance and 2x faster attention mechanisms compared to the DGX GB200. Designed with the NVIDIA MGX chassis, the DGX GB300 can scale to thousands of nodes when combined with DGX SuperPOD. Preloaded with NVIDIA Mission Control and AI Enterprise software, it is optimized for large model training, post-training, and test-time inference.
NVIDIA Vera Rubin NVL72
NVIDIA Vera Rubin NVL72 is a rack-scale agent AI supercomputer that unifies 72 NVIDIA Rubin GPUs, 36 NVIDIA Vera CPUs, NVIDIA ConnectX-9 SuperNICs, and BlueField-4 DPUs on a single platform. It scales vertically via sixth-generation NVLink and NVLink Switch, and horizontally through Quantum-X800 InfiniBand and Spectrum-X Ethernet. Built on the third-generation MGX NVL72 rack design with a cable-free modular tray architecture, it reduces GPU requirements for AI training to one-quarter of Blackwell's and lowers inference costs per million tokens to one-tenth.
NVIDIA GB300 NVL72
The NVIDIA GB300 NVL72 features a full liquid-cooled rack-scale architecture that integrates 72 NVIDIA Blackwell Ultra GPUs and 36 Arm-based NVIDIA Grace CPUs onto a single platform. Compared to Blackwell GPUs, it delivers up to 1.5× more dense FP4 Tensor Core performance and up to 2× better attention mechanism performance, purpose-built for test-time scaling inference and AI inference workloads. Paired with Quantum-X800 InfiniBand or Spectrum-X Ethernet, ConnectX-8 SuperNICs, and NVIDIA Mission Control management, the entire AI factory achieves up to 50× higher throughput than Hopper-based platforms.
NVIDIA GB200 NVL72
The NVIDIA GB200 NVL72 uses rack-scale liquid cooling to connect 36 Grace CPUs and 72 Blackwell GPUs. Its 72-GPU NVLink domain operates as a single large GPU, delivering 130TB/s ultra-low latency interconnects. Compared to the H100 platform, it accelerates real-time inference for trillion-parameter LLMs by 30x, LLM training by 4x, energy efficiency by 25x, and database query processing by 18x over CPUs. The GB200 Grace Blackwell superchip is the system's core building block, linking two Blackwell Tensor Core GPUs with one Grace CPU via NVLink-C2C.
NVIDIA H200
The NVIDIA H200 is a memory-enhanced data center GPU based on the H100, featuring 141GB of HBM3e memory and 4.8 TB/s bandwidth. It delivers significantly higher throughput and lower latency for large language model inference. Available in SXM and NVL form factors, it supports generative AI training and inference, as well as high-performance computing (HPC) workloads.
NVIDIA H100
The NVIDIA H100 is a data center GPU based on the Hopper architecture, available in SXM, PCIe, and H100 NVL form factors. It features up to 80 GB of HBM3 memory (with 4 GB per card for the NVL version), offers fourth-generation NVLink bandwidth of 900 GB/s, supports Multi-Instance GPU (MIG), and serves as a leading compute platform for large model training, HPC, and large-scale inference.
NVIDIA GH200 Grace Hopper Superchip
The NVIDIA GH200 Grace Hopper Superchip connects the 72-core Arm Grace CPU and Hopper GPU via a unified memory architecture through NVLink-C2C (900 GB/s). It is available with 96GB or 144GB of HBM3/HBM3e memory, with up to 480GB of total CPU memory LPDDR5X, making it ideal for large model training and inference, high-performance computing (HPC), and recommendation systems.
NVIDIA L40S
The NVIDIA L40S is a general-purpose data center GPU based on the Ada Lovelace architecture, featuring 48GB GDDR6 ECC memory. It combines third-generation RT Cores and fourth-generation Tensor Cores to support generative AI fine-tuning and inference, graphics rendering, and video processing. It comes in a 350W dual-slot PCIe form factor.
NVIDIA L40
NVIDIA L40 is a data center GPU based on the Ada Lovelace architecture, designed for professional graphics and AI workloads. It features 48GB GDDR6 ECC memory, third-generation RT Cores, and is ideal for AI image generation, ray tracing rendering, virtual workstations, and video processing. The card has a 300W dual-slot PCIe form factor.
NVIDIA L4
The NVIDIA L4 is a lightweight data center GPU based on the Ada Lovelace architecture. It features a 72W single-slot low-profile design and 24GB GDDR6 memory, with hardware AV1 encoding and decoding support. Ideal for video transcoding, lightweight AI inference, cloud gaming, and virtual desktops, it delivers exceptional performance per watt and high deployment density.
NVIDIA A100 80GB
The NVIDIA A100 80GB is a flagship Ampere-based training GPU featuring 80GB HBM2e memory, third-generation Tensor Cores, and Multi-Instance GPU (MIG) technology. Available in both SXM and PCIe form factors, it serves as the go-to compute platform for large-scale model training, fine-tuning, and HPC. In the domestic market, the A800 variant offers NVLink bandwidth adjusted to 400 GB/s.
NVIDIA A40
The NVIDIA A40 is a data center GPU based on the Ampere architecture, combining professional graphics rendering with AI computing power. It features 48GB GDDR6 ECC memory, second-generation RT Cores, and third-generation Tensor Cores, supports dual-card NVLink bridging, and is ideal for graphics rendering, virtual workstations (vGPU), and AI inference.
NVIDIA A30
The NVIDIA A30 is a mainstream compute GPU based on the Ampere architecture, featuring 24GB HBM2e memory and 933 GB/s bandwidth. It supports MIG (up to 4 instances) and FP64 double-precision computing, making it ideal for AI inference, model fine-tuning, and database acceleration.
NVIDIA A16
NVIDIA A16 integrates 4 Ampere architecture GPUs (4x 16GB GDDR6, totaling 64GB) in a dual-slot, 250W form factor. It features multi-channel NVENC/NVDEC codecs and is purpose-built for high-density virtual desktop infrastructure (VDI) and multi-user virtual workstations (vGPU).
NVIDIA A10
The NVIDIA A10 is a single-slot data center GPU based on the Ampere architecture, featuring 24GB of GDDR6 memory and a 150W TDP. It delivers balanced performance for AI inference, cloud gaming, virtual workstations, and video processing; the A10G is a variant customized for cloud service providers.
NVIDIA A2
The NVIDIA A2 is an entry-level data center GPU based on the Ampere architecture, featuring 16GB of GDDR6 memory. Its low-profile, single-slot, low-power design makes it ideal for entry-level AI inference, video transcoding, and virtual desktops, enabling easy large-scale deployment in standard servers.
NVIDIA DGX Spark
Desktop AI supercomputer powered by the GB10 Grace Blackwell superchip, featuring a 20-core Grace CPU and Blackwell GPU with coherent unified 12 8GB memory. Delivers up to 1 PFLOPS of FP4 AI performance, supporting local development and inference for large-parameter models, with scalability up to 4 nodes in a cluster.
Need NVIDIA solutions?
As an NVIDIA partner, Huaying Technology provides end-to-end professional services—from selection consulting to deployment implementation.