IVA TPU

Main articles:

• Neuronets (neural networks)
• Machine learning (Machine Learning)

IVA TPU is the first completely Russian neural processor allowing to create the entrusted solutions in the field of telecommunications and radio electronics.

2020: Among the best world solutions on the version of MLPerf

The Russian IVA TPU processors were among the best world solutions on the version of MLPerf. On November 5, 2020 TAdviser was reported about it by IVA Technologies company.

The Russian tensor processors are recognized one of world-best. Photo source: rlocman.ru

On the international platform MLPerf held testing of the architecture of the accelerator of calculation of neural networks IVA TPU developed by IVA Technologies company. Product assessment will allow to raise trust degree to the developer, in particular from commercial and research foreign communities, and will provide it the worthy place among the international giants in the market of microelectronics in a support area of artificial intelligence.

Results of testing of architecture of IVA TPU implemented in FPGA are confirmed and published on the official page of the website MLPerf^[1] in sections of an inferens on the terminal units Edge, Closed Division (ID 0.7-156) and Edge, Open Division (ID 0.7-396). MLPerf is the international platform of certification and testing of the equipment in the field of machine learning and artificial intelligence. The solution was qualified in the most complex Closed Division group – the closed list of networks with restrictions for modifications, waiting time and accuracy.

This event without exaggeration otlichitelno for Russia. It is difficult to remember examples when domestic development in microelectronics, received the school-leaving certificate of hardware and program components at foreign experts. We showed that rather small business young company, at correctly built development strategy and workflows, is capable to implement the potential of the command, to create a product in one of the most difficult fields of microelectronics – microprocessors. noted the deputy CEO of IVA Technologies, the head Evgeny Terentyev.

Edge AI is the most fast-growing segment in the market, and in the Closed Division Edge group only 5 suppliers are provided: Nvidia, ARM, Centaur, Xilinx and, since September, 2020, IVA Technologies.

From the analysis of results it is visible that for one of the most widespread neural networks (Resnet-50), the architecture of IVA TPU provides a delay 3 times smaller, than the best solution of Xilinx. It is planned that when implementing the accelerator with architecture of IVA TPU in ASIC (are expected in the first quarter 2021), results will be better, than our closest competitor has Nvidia Jetson Xavier NX despite lag in technology (28 nanometers against 12 nanometers).

It was serious test and assessment of level of a maturity of nesting and program storages of architecture of IVA TPU which our command passed, having confirmed the high professionalism. The Russian development took the worthy place in the same row with the international giants with the sphere of microelectronics for support of artificial intelligence, added Evgeny Terentyev.

In the nearest future based on an IP core of IVA TPU two configurations of microprocessors – for terminal units (IVA E) and cloud services (IVA H) are put in production. Based on microprocessors several sales of modules which will be directly applied in electronic devices of customers are developed.

Thus IVA Technologies will be able to create a product line which will allow to satisfy the needs for computing resources for the accelerated calculation of neural networks, at absolutely different production and service companies. It is about producers a smart of devices (smart cameras, IoT, industrial and city security), about developers of independent machines (ADAS, drones), about large operators of services and DPC (processing photo and video this, the text, a voice, information security). Besides, in plans of the company the beginning of sales of computing resources for an inferens of neural networks through a cloud service.

2018: Presentation

On September 24-25, 2018 in Moscow the tensor IVA TPU processor was provided.

According to IVA Technologies company, the microprocessor is intended for application (inferens) of convolution neural networks based on FPGA Virtex Ultrascale + and finds practical application for detecting and pattern discrimination, object search, quality improvement and others.

The main features of IVA TPU, according to developers, are the compatibility with the ready open source (TensorFlow), high performance (more than 20 TOPS) at small energy consumption and considerably smaller time of processing, than GPU.

For September, 2018 the need for development of specialized computing devices caused by requirements imposed to computing powers during the training and application of neural networks is sharply felt. We successfully coped with this task — the implemented model of the tensor processor confirmed engineering calculations of TTH. We already have ready examples of solutions for detecting and image identification, segmentation and search of persons on images — Nikolay Ivenev, the chairman of the board of directors of IVA Technologies told.

Notes