AI Developer Groq Looking for Investments
08-04-2021 | By Robin Mitchell
Recently, AI solutions developer Groq announced that it is in the process of looking for investments that would put it on par with Nvidia. Who is Groq, what have they developed, and what will the investment round value them at?
Who is Groq?
Groq is a company that specialises in the development of AI hardware and software. The founder of Groq, Jonathan Ross, initially worked for Google developing their Tensor Processing Units and was responsible for designing the core elements. After working on Google’s AI hardware, Jonathan moved on to work on another Google project; Moonshots factory.
The company was founded in 2016, is located in Palo Alto, California, and now has an estimated 50 to 100 employees. To date, Groq has received an estimated total of $60 million in investment, and the company is still in its investment stage (i.e. it is not a commercial company as of yet).
How does Groq technology differ from other AI implementations?
The use of AI has skyrocketed thanks to its ability to adapt and learn to complex tasks that typically cannot be practically expressed in code. For example, speech-to-text has historically been very poor on computing systems, and this is a result of different accents, pitches, and languages. AI, however, can be made to listen to many millions of people and learn how to ignore irrelevant data in speech to hear what is important; the words.
While the development of AI opened the world to new possibilities, the nature of AI is fundamentally different to typical computer programs. As such, AI systems running on traditional CPUs would be time consuming and power-hungry. It did not take long for researchers to recognise that GPUs provide a much better platform to run AI algorithms on thanks to their large parallelism.
However, technology never stops, and the hardware being used to execute AI systems is fundamentally flawed; the hardware executing AI was never designed with AI in mind. Therefore, if AI is to continue in performance, the underlying hardware must be dedicated to the task.
This is the mindset that Groq has, and has gone on to develop some of the worlds best AI hardware and compiling software. Their current device can perform 1000TOp/s (Tensor operations), has 220MB of SRAM on-chip, and a memory bandwidth of 80TB/s. Furthermore, all silicon space on the chip is dedicated to AI execution, whereas GPUs have wasted space for rendering and CPUs have space on unneeded instructions.
The device developed by Groq has a total of 26.8 billion transistors, utilises a 14nm process, and has an INT8 performance of 1 PetaOp / second when operating at 1.25GHz. To further improve performance, Groq also implemented features into their compiling stage so that the compiler knows exactly how to assign tasks on the device to maximise hardware usage. The result is no speculative code execution and all instructions occur exactly when needed.
Groq Looking for More Investors
Groq is now looking to expand its operations and find more investors with a target of around $300 million. If the company can secure such funding, it would have a valuation of around $1 billion, and this would rival competitors such as Nvidia (keep in mind that the $1 billion valuations are solely in AI).
So far, Groq has been able to raise $250 million, and considering how global investment into semiconductor developers is booming, Groq may have found the right time to obtain investment. For example, the years between 2000 and 2015 saw little change in semiconductor investments, but since 2015, these numbers have more than doubled as the world realises the importance of the semiconductor.
Groq has stated that it is using its technology with customers in the autonomous vehicle sector but has not named any partners. A competitor of Groq, Graphcore, has already been valued at $2.5billion and is working with Dell and Microsoft. Since the automotive industry will be heavily dependent on AI in the future, securing automotive partners early could be essential for Groq.
Read More
