-1）GeForce RTX 4090

GeForce RTX 4090

0) GeForce RTX 3090 graphics card, it seems that there is no single or double precision data?

1)

A40:

The NVIDIA A40 accelerates the most demanding visual computing workloads from the data center, combining the latest NVIDIA Ampere architecture RT Cores, Tensor Cores, and CUDA® Cores with 48 GB of graphics memory. From powerful virtual workstations accessible from anywhere to dedicated render nodes, NVIDIA A40 brings nextgeneration NVIDIA RTX™ technology to the data center for the most advanced professional visualization workloads.

There is no single or double precision.

Neither does bandwidth.

2)

A30：

Built for AI inference at scale, the same compute resource can rapidly re-train AI models with TF32, as well as accelerate high-performance computing (HPC) applications using FP64 Tensor Cores. Multi-Instance GPU (MIG) and FP64 Tensor Cores combine with fast 933 gigabytes per second (GB/s) of memory bandwidth in a low 165W power envelope, all running on a PCIe card optimal for mainstream servers.

 3）A100

　NVIDIA A100 GPU adopts the super-large core GA100 of the new Ampere architecture, 7nm process, 54.2 billion transistors, an area of ​​826 square millimeters, 6912 cores, equipped with 5120-bit 40/80GB HBM2 memory, a bandwidth of nearly 1.6TB/s, and a power consumption of 400W.

NVIDIA A100 Tensor Core GPUs deliver outstanding acceleration for AI, data analytics, and high-performance computing (HPC) applications at every scale, powering high-performance, elastic data centers around the world. As the engine of NVIDIA's data center platform, the A100 delivers up to 20x higher performance than the previous-generation NVIDIA Volta™. Efficiently scalable, the A100 can also be partitioned into seven independent GPU instances with Multi-Instance GPU (MIG) technology, providing a unified platform that enables elastic data centers to dynamically adjust to changing workload demands. NVIDIA A100 Tensor Core technology supports a wide range of mathematical precision, enabling a single accelerator for each workload. The latest generation of A100 80GB doubles the GPU memory, providing 2TB/s of global ultra-fast memory bandwidth, which can accelerate the processing of very large models and massive data sets. The A100 is part of a complete NVIDIA data center solution consisting of a stack of hardware, networking, software, libraries, and optimized AI models and applications from the NGC™ catalog. As a powerful end-to-end AI and HPC platform for the data center, the A100 empowers researchers to achieve real-world results and deploy solutions into production environments at scale.

It's a pity that it is not allowed to be sold in China, the evil ghost.

 4）V100

The NVIDIA® V100 Tensor Core GPU is a powerful accelerator for deep learning, machine learning, high-performance computing (HPC), and graphics computing. Powered by the NVIDIA Volta™ architecture, the V100 Tensor Core GPU can deliver the performance of nearly 32 CPUs in a single GPU, enabling researchers to tackle previously inaccessible challenges. The V100 has already topped MLPerf, the industry's first AI benchmark, proving it to be the world's most powerful computing platform with massive scalability and versatility.

Not worse than A100, but better than A30.

American brains are also funny

5）A800

 NVIDIA will launch a new A800 GPU for Chinese users to replace the A100, which is in line with the US government's export control policy. According to the policy of the US government, the data transfer rate of chips for export shall not exceed 600GB/s, and NVIDIA A800 is set at 400GB/s. But other specifications are unclear, such as the number of cores, operating frequency, video memory, power consumption, etc.

, Nvidia and AMD said that products including Nvidia's data center chips A100 and H100 were included in the export control list by the US Department of Commerce.

And according to Nvidia, the new A800 can replace the A100, both of which are GPU (graphics processing unit) processors. According to the NVIDIA A800 GPU information introduced on the official website of chip dealer OMNISKY Rongtian, the data transfer rate of the new chip is 400GB per second, which is lower than the 600GB per second of A100, which represents a significant decline in the performance of the data center. Moreover, the A800 supports a memory bandwidth of up to 2TB/s, with little change in other parameters.