| Developers: | Nvidia |
| Date of the premiere of the system: | 2022/03/25 |
| Branches: | Electrical and microelectronics |
| Technology: | Processors |
Main articles:
2022: Announcement of Hopper silicon architecture and its GPU H100
On March 25, 2022, NVidia announced a number of technologies focused on using in the field of artificial intelligence. First of all, this is the Hopper silicon architecture, based on its GPU H100 and the Grace CPU "superprocessor."
According to the company, the architecture and superprocessor are named after Rear Admiral of the US Navy Grace Hopper, creator of the first-ever computer programming language compiler and one of the first high-level COBOL programming languages.
The Hopper architecture and the H100 GPU are designed to optimize the process of learning machine models. The fourth generation NVLink bus allows you to connect up to 256 H100 GPUs, providing a nine-fold increase in throughput compared to previous generations.
The H100 processor itself contains 80 billion transistors. It supports PCIe Gen5 and HBM3; this brings memory throughput to 3 terabytes per second. Nvidia claims that the H100 is three to six times superior to the previous generation A100 processor in floating point calculations depending on the bit rate.
Depending on the form factor, the characteristics of the GPU will differ slightly. So, the older version has 15,872 CUDA cores.
Both versions have 80 GB of memory, and this is HBM3 with a bandwidth of 2 or 3 TB/s. The TDP of the older version is 700 watts, while the modification with PCIe 5.0 has half as much.
The performance of the SXM5 version reaches 30 TFLOPS (FP64), 60 TFLOPS (FP32), and the PCIe 5.0 version has indicators of 24 and 48 TFLOPS, respectively. It is also worth noting the NVLink interface, which in the same order provides a data rate of 900 or 600 GB/s.
As for the Grace CPU, these are actually two processors connected by the NVLink-C2C bus. Grace is designed for use in "hyperscale high-performance computing and AI" complete with H100 or independently. The processor has 144 Arm cores, and its throughput is 1 terabyte per second[1].
Hopper architecture brings to the market fourth-generation tensor cores, which are nine times faster than the previous generation in a number of machine learning tasks. Separately, Nvidia noted a set of DPX instructions that will help developers write code to accelerate dynamic programming algorithms in various industries, accelerating work processes for disease diagnosis, quantum modeling, graph analysis and routing optimization.
Hopper-based systems will be available to Nvidia partners in the third quarter[2].
Notes
| The site content is translated by machine translation software powered by PROMT. The machine-translated articles are not always perfect and may contain errors in vocabulary, syntax or grammar. Read original article If you find inaccuracies or errors in the results of machine translation, please write to editor@tadviser.ru. We will make every effort to correct them as soon as possible. |