If they succeed, the start-up company can skip digital processors and cores from companies such as Intel or other IP providers and many wealthy Chinese start-up companies; the goals of these processors are to station A generation of security cameras, drones, factory equipment and other embedded systems that try to catch up with the artificial intelligence (AI) boom, and even future self-driving vehicles.
'As we learned from the Institute, mixed signal processing is very suitable for this kind of application;' David Fick, who founded Mythic with his colleague Mike Henry at the University of Michigan, said: 'You need to use its adjustable threshold voltage To store large amounts of data and flash memory - each transistor is very attractive.
The flash memory array developed by Mythic basically eliminates the need to move data from external memory into and out of the memory, resulting in significant power savings; Fick said that his advisors David Blaauw and Dennis Sylvester have already conducted some flash memory research. We also have some expertise in technology so it can be very easy to start a project quickly.
However, it is a tricky task to implement the decades-old concept of analog processors in memory. Fick said: 'You must consider many analog effects—mismath, noise, temperature, and memory cells. There are many similar significant effects; ' Unlike digital computers with well-defined memory, processing and storage subsystems, analog computers used in machine learning are basically an integrated and massive project.
'You need to design everything together at the same time, so you need to understand people in overlapping areas, such as understanding component-side design and neural network design engineers in each other's fields;' Fick explained: 'We have outperformed everyone else in this area because We have a super team that can complete the entire task. '
Indeed, the company just completed round B financing and received a sum of up to $50 million in funding, partly because they have a diverse team of director-level experts, including analogies from Texas Instruments. Expert, Microchip's Flash Memory Design Director, and Netronome's Physical Design Specialist.
Dave Fick and his pet dog Ellie at the Austin office; Ellie is director of unofficial emotional support at Mythic (Source: EE Times)
Mythic also won the favor of investors by gradually demonstrating its technological progress through a series of prototype films. Fick won a lot of prestige in his VLSI design performance in schools; he said: 'When you design chips as an academic student, including Memory, synthesis, DRC variations... all steps have to be done by yourself; and if you are directly into the industry, you may never see the entire design process, so many independent start-ups from the school will It is easier to successfully mass-produce. '
The two founders of this company have been “geek” since childhood. Fick’s first job in high school was a web development engineer. When he studied at the institute, he entered AMD, IBM and Intel, etc. Born in. Henry is for fun, loves to participate in a variety of fast writing program contests.
Large and small competitors and software obstacles to overcome
These days, Mythic's duo encounters large and small competitors. At least 40 veteran and new IP providers or chip makers have expressed their intention to launch or plan some form of client AI accelerator chip. These competitors also include several wealthy start-up companies in China. Horizon Robotics, for example, is one of the most promising ones. It has introduced a low-power client AI accelerator with a more traditional digital architecture.
There is also a US startup company, Syntiant, who, like Mythic, is developing a processor-in-memory architecture that utilizes flash memory. The company's team includes several former Broadcom engineering managers and has received Intel Capital. In addition, IBM Research is also studying machine learning accelerators based on resistive RAM (ReRAM), but Fick believes that the company used the wrong method.
He said: 'They are trying to make everything easy with the perfect memory, but we are leading through designing everything together... even if they find the ideal memory, there will always be a less than perfect memory to support Lower power consumption or faster speed. '
Innovative parallel memories have historically always failed because of too difficult programming. Emerging in-memory processor chips will certainly face the same problem, because machine learning itself requires a new, still developing programming model. Although Mythic's tools are To develop the platform, but to play a role like a compiler, it can convert the neural network described in the TensorFlow database into a machine language for its chips.
Fick said that the development platform uses PCI Express and chip connectivity to provide hints on how to obtain additional performance from the chip, as well as optimized networking examples for some common applications. Customers who want to use a framework other than TensorFlow will obviously Need to use ONNX format to translate its tasks; ONNX (Open Neural Network Exchange) is one of the few emerging tools used to translate several different AI software architectures.
Fick is also fully aware of the software barriers facing his customers: 'In order to enter this field, you need to employ several deep learning scientists, but such experts are very scarce, and the cost is very expensive... Establishing data sets and neural nets Roads and training are very time consuming and costly... these are all the limits to venture into and invest in this area.'
The good news is that, compared with competing solutions, the memory array of Mythic chips should be able to handle more diverse convolutional or recurrent neural networks, and its performance improvement is expected to be achieved in power consumption. Restricted edge systems perform more complex models.
Mythic has several test chips to date (Source: EE Times)
Mythic has several heavyweight partners, such as Lockheed Martin, who hopes that future drones can use the company's chips, and Fujitsu, the company's flash memory supplier. So far, there are two The application seems to be beyond its scope. One is a smart speaker with a budget of only a few dollars. Compared with Mythic's target application, it is too costly to control; the other is a self-driving vehicle. Because of the need for vehicle grade specifications, the company is currently unable to Born.
The start-up is expected to introduce a 40nm process chip later this year. This node supports embedded flash memory cell design and also meets low-cost targets. Fick pointed out that its flash memory cell has passed 28nm process quality certification. This will be the company's next step; after that, the foundry industry is developing embedded MRAN and ReRAM units.
Fick said: “There is no reason to stop us from moving forward with the smallest node. We can benefit from the process miniaturization; 'And if Mythic succeeds, it’s not because Moore’s Law or those digital processors make them It's welcome, but it's because they push computing technology to a whole new direction.
Compilation: Judith Cheng