Another big fundraising deal is underway for an AI company today, this time for a start-up called Mythic getting $ 40 million because it seems like massive deals close to the left and right in the area.
Mythic focuses in particular on the inference side of AI operations – essentially doing the on-site calculation for something based on a largely formed model. The chips are designed to be low-power, small, and achieve the same kind of performance you expect from a GPU in terms of the ultra-fast operations that algorithms must perform to determine whether or not this thing is about hitting is a cat or just a text on the road. SoftBank Ventures led this latest round of financing, with a strategic investment also from Lockheed Martin Ventures. Rene Haas, a member of ARM management, will also join the board of directors of the company.
“The key to getting very high performance and very good energy efficiency is keeping everything on the chip,” said Henry. “The minute you have to get out of the chip in memory, you lose all performance and energy, it just comes out the window, and knowing that, we found that you can actually take advantage of flash memory in a very The limit is there, it’s for the inference only, but we’re just following the market of inference – it’s going to be huge, besides that, the challenge is to bring processors and memory as much as possible so you do not have to move the data on the chip. “
Mythic, like other startups, is looking to ease back and forth on processors to speed things up and reduce power consumption, and CEO Michael Henry says that the company understood how to basically do the operations – based in a mathematical domain called linear algebra – on the flash memory itself.
Mythic’s approach is designed to be what Henry calls more analog. To visualize how this might work, imagine a configuration in Minecraft, with a number of different strings of blocks leading to an end gate. If you flipped a switch to activate 50 of these strings with a value of unity, leaving the rest, and joined them at the end and saw the combined end result of the power, you would have accomplished something similar to an addition operation leading to a sum of 50 units. The Mythic chips are designed to do something not so different, by finding ways to supplement these types of analog operations for addition and multiplication in order to handle the computational requirements for an operation. 39; inference. The end result, says Henry, consumes less energy and dissipates less heat while getting just enough precision to get the right solution (more technically: the calculations are 8-bit results).
After that, the challenge is to stick a layer on top of it to give it the appearance and behavior of a normal chip to a developer. The goal is to, like the other players in the IA hardware space, to simply plug into frameworks like TensorFlow. These frameworks eliminate all the complicated tooling and optimization needed for such specific hardware and make it very accessible and easy for developers to start building machine learning projects. Andrew Feldman, CEO of Cerebras Systems, said at a conference on technology and the internet of Goldman Sachs last month that frameworks like TensorFlow had most of the value Nvidia had built an ecosystem for developers on his own system.
Henry, too, is a big fan of TensorFlow. And for good reason: thanks to frameworks such as TensorFlow, the ideas of new generation chips can even take off. These types of frameworks, which have become increasingly popular with developers, have eliminated the complexity of working with specific low-level hardware such as a Field Programmable Gate Array (FPGA) or GPU. This has made machine-based operations a lot easier for developers and has led to an explosion of activity in machine learning, as it works. recognition of speech or image among a number of other use cases.
“Things like TensorFlow make our lives so much easier,” said Henry. “Once you have a neural network described on TensorFlow, it is up to us to take it and translate it on our chip.We can summarize this difficulty by having an automatic compiler.”
While many of these companies are talking about getting massive performance gains on a GPU – and, of course, Henry hopes this will be the case – the short-term goal for Mythic is to 39, equaling the performance of a $ 1000 GPU showing that it can take up less space and consume less energy. There is a market for the card that customers can swap hot right away. Henry says the company is focusing on the use of a PCI-E interface, a very common plug-and-play system, and that’s all.
The challenge for Mythic, however, will go into the actual design of some of the materials that come out. It’s one thing to sell a bunch of cards that companies can stick in their existing hardware, but it’s another to incorporate them into the hardware components themselves – which is going to have to happen if she wants to be a real work tool. on the edge, like security cameras or things that manipulate voice recognition. This makes the buying cycle a little harder, but at the same time, there will be billions of devices that need advanced hardware to fuel their inference operations.
“If we can sell a PCI card, you buy it and you drop it off right away, but these are usually cheap products and high selling price,” Henry said. “The other customers that we serve inform you about the hardware products.It is a longer cycle, which can take more than a year.For this, the volumes are usually much higher.The good thing is that you are really very sticky.If they design you into a product, you are really sticky.We can go after both, we can go after the sales council, and then go after the design. “
There will probably be two big walls for Mythic, and even less for all the other players. The first is that none of these companies shipped a product . While Mythic, or other companies, might have a proof-of-concept chip that can fall on the table, getting a new generation silicon chip ready for production is a dramatic undertaking. Then there is the process of not only convincing people to buy the equipment, but also convincing them that they will have the systems in place to make sure that the developers are able to do it. will support this material. Mythic says that he plans to have a sample for customers by the end of the year, with a production product by 2019.
This also explains why Mythic, with these other startups, is able to raise huge sums of money – which means that there will be a lot of competition between them. Here is a quick list of what has happened so far: SambaNova Systems raised $ 56 million last week; Graphcore raised $ 50 million in November last year; The first round of Cerebras Systems was $ 25 million in December 2016; and this is not even counting a growing amount of activity occurring among companies in China. There is always a segment of investors who consider that the space is too hot (and there is, indeed, a ton of funding) or potentially useless if you do not need it. Efficiency or power of these products
And there are, of course, elephants in the room in the form of Nvidia and to a lesser extent Intel. The latter is betting big on FPGAs and other products, while Nvidia has won the bulk of the market thanks to GPUs that are much more effective in the maths needed for AI. The game for all these startups is that they can be faster, more efficient, or in the case of Mythic, less expensive than all these other options. It remains to be seen if they will overthrow Nvidia, but nevertheless there is a huge amount of funding coming in.
“The question is, will anyone be able to beat Nvidia when they have valuation and cash reserves,” said Henry. “But the thing is, we’re in a different market, we’re going after the edge, we’re going after things built into phones and cars and drones and robotics, for applications like AR and VR. and it’s really a really different market.When investors analyze us, they have to think about us differently.They do not think, is it whoever wins Nvidia, they think, are one or more of these markets “It’s a different conversation for us because we’re a high-tech company.”