» Current | 2020 | 2019 | 2018 | Subscribe

Linley Newsletter

Huawei Delivers Powerful AI Training

October 8, 2019

Author: Bob Wheeler

The AI-training landscape is dominated by chip vendors and hyperscale data-center operators, but Huawei is first to deliver a training chip in a 7nm EUV process. The OEM recently announced accelerator cards and AI servers using the Ascend 910 AI processor, which is based on its DaVinci core. Those products are due in 1Q20, but the company is already shipping the DaVinci architecture in inferencing-accelerator cards, which are based on the earlier Ascend 310 chip, as well as in Kirin smartphone processors.

Like other recent training chips, the Ascend 910 integrates multiple die in a package, including four HBM2 stacks. It also uses two logic die, separating the processing and I/O functions. The Vitruvian die integrates 32 DaVinci cores with 256 FP16 teraflops per second (Tflop/s) of raw performance, 32MB of level-two memory, and quad HBM2 channels. An integrated host processor includes 16 Arm-compatible Taishan CPUs and dual DDR4 channels. The Nimbus I/O die provides a PCI Express host interface, 100Gbps Ethernet ports, and coherent interconnects. Including 32GB of HBM2, the Ascend 910 processor dissipates 350W (TDP).

Huawei intends to provide full-stack AI solutions, so it developed comprehensive software ranging from a low-level compiler to a high-level framework dubbed MindSpore. It also supports TensorFlow and PyTorch, but it claims MindSpore delivers greater performance than these popular frameworks. By offering complete systems and software, Huawei will compete directly with Nvidia’s DGX training systems. The Ascend 910 offers about twice the raw performance of the Nvidia’s Tesla V100, but early benchmarks show far less of an advantage when running neural networks.

Subscribers can view the full article in the Microprocessor Report.

Subscribe to the Microprocessor Report and always get the full story!

Purchase the full article

Events

Linley Spring Processor Conference 2020
Coming April 7-8, 2020
Hyatt Regency, Santa Clara, CA
Register Now!
Linley Fall Processor Conference 2020
Coming October 28-29, 2020
Hyatt Regency, Santa Clara, CA
More Events »

Newsletter

Linley Newsletter
Analysis of new developments in microprocessors and other semiconductor products
Subscribe to our Newsletter »