The Russian company launched mass production of neuroprocessors - competitors Nvidia

Neuroprocessor of STC "Module"

The Russian Scientific and Technical Center "Module" has developed a new neural processor NM6408, and the creation of a chip cost 480 million rubles. Now it is launched into mass production. “Several thousand processors are already in our warehouse, ready for shipment, we plan to order them as needed,” the company said.

The 21-core chip is designed for a 28 nm topology, as reported by Cnews. It was created as a computing accelerator for servers with a peak performance of 512 Gflops. It is designed to process large data streams in real time, allowing, according to the developers, to solve a variety of tasks, not only neural networks.

The most promising area of ​​application for the new chip is the processing by neural network applications of large arrays of video data, for example, from high-resolution cameras. Potential customers of the chip are state and commercial transport companies, organizations from the cyber security field, energy companies, etc.
')
"Since our chips have advanced technical characteristics (for example, a wide temperature regime), then we can talk about devices operating in severe weather conditions, for example, in the High North," the developers said.

Processors are made in Taiwan. The problem is that in Russia there are no industries that allow you to create processors of 28 nm process technology. The developers promise that when the necessary technologies appear in the Russian Federation, production will be transferred to Russia.

In “Modula” they believe that now in Russia there are no direct competitors for the new chip, since no other domestic developer has completely ready neural processors. “This is a neural processor with the highest performance on the Russian market (all other processors are also with us, they are less productive) and, according to their characteristics, can compete with world leaders,” the company reports.

Of course, there are world analogues, but here Modul’s products have an advantage - the price of the product is about $ 70 versus $ 100 for comparable foreign chips. Of the competing chips, the representatives of the Module mention Nvidia Jetson Xavier.

“We are approaching them in terms of performance. Moreover, during tests, it turned out that our processors are more efficient precisely for the operation of neural networks, that is, they achieve greater results with lower hardware costs; The efficiency varies by several times, ”the company said.

Based on the new process, the “Module” has developed a high-performance computing complex MC127.04. It is designed to work as part of embedded computing technology, as well as specialized high-performance systems. Plus, the chip can be used to create digital signal processing systems and machine vision.

As mentioned above, the development of the module cost the domestic manufacturer 480 million rubles. All work was carried out in the framework of the execution of the state contract with the Ministry of Industry and Trade. The contract was signed by the parties in March 2014, the contract expired on December 26, 2017.

“OCD was delivered on time,” say the developers. “During 2018, mass production was mastered: an instrumental board was developed, testing was conducted, the appropriate software was written to work on the processor, etc.”

Also, developers and technological details comment: “The“ Module ”develops a line of processors based on its own NMC core - the NeuroMatrix Core, a unique microprocessor architecture”. The new chip belongs to the fourth generation of the development of ideas that were first realized in the 1990s. The first chip, as the developers claim, “created a furore at the Cebit exhibition in German Hanover in 1998 and was licensed by Fiujitsu”.

The processor includes one 32-bit universal control RISC processor with a reduced instruction set of ARM Cortex-A5, plus four clusters, each of which contains one ARM Cortex-A5 RISC processor and four NMC4 processor cores. The chip has five interfaces with external memory such as DDR3, interfaces PCIe2.0, SPI, Ethernet IEEE Std 802.3-2012, GPIO, JTAG and high-speed interfaces for interprocessor exchange.

Typical power consumption of the chip - 20 watts, maximum - 35 watts.