Why We Invested in Anari AI

Enabling AI users to design cloud chips for limitless value creation

Earlybird Venture Capital
Earlybird's view
Published in
4 min readApr 29, 2021

--

By Roland Manger, Earlybird Partner & Founder

Outgrowing GPUs

First, a bit of history: The graphics processing unit (GPU) made real-world Machine Learning (ML) possible. It provided significantly more compute power and offered a faster connection to memory than the traditional general purpose central processing unit (CPU). Data centers rapidly incorporated them into their offerings, and GPU vendors developed software to help their hardware be used effectively.

The use of Artificial Intelligence (AI) and Machine Learning (ML) is still in its early days; algorithms have been, and continue to, grow in size and complexity while the use of ML is progressing at a rate that is outstripping Moore’s Law (1), and in some ways more relevant, Dennard’s Law (2).

In other words, AI and ML are now evolving faster than fixed architectures like GPU and CPU can.

FPGAs haven’t cut it yet

There are processor technologies and paradigms that provide flexible architectures which can be adapted to specific problems or algorithms. Field-programmable gate arrays (FPGAs) represent one important approach. More specific and optimal processor configurations can be created by combining a number of parallel computational elements. These custom creations provide much higher processing capacity than GPUs and at the same time only require a quarter of a GPU power budget.

The most important aspect of FPGAs is their flexibility. With different kinds of resources (cores, IP, memory), their architecture can be adapted or reprogrammed in the field, which allows for very efficient processing of a wide range of different algorithms and computing approaches – including neural networks, important to ML. They can also be tied together into larger custom processor units consisting of several or even many individual FPGAs.

Very large enterprises like JP Morgan or Shell have started to use FPGAs for these reasons but also the large platform providers are starting to offer “FPGAs for rent” in their Clouds. AWS was the first to the party, but Azure, Alibaba and GCP will follow.

However, FPGAs do not provide a natural programming model that software developers are used to, so they require hardware engineers to program them. Hence, their utility has been limited. Previous attempts to abstract from the hardware level have tried to create general purpose development environments that would allow for the programming of a very wide range of applications. This broad approach, however, results in a lot of complexity that developers still need to deal with. As a result, hardware-oriented programming languages and paradigms are still necessary.

The founding team at Anari AI (L to R): Stefan Sredojevic, Jovan Stojanovic and Bogdan Vukobratovic

Enter Anari AI

Anari AI takes a vertical approach to this opportunity. It focuses on AI and ML only, with an initial further focus on 3-D image and graph model algorithms that cover a large and increasing share of ever more complicated processing workloads. Based on this specific domain, Anari AI is able to abstract to a level that accommodates programming in one of the most common high level languages today: Python, and hence gives software-only programmers the power to create custom cloud chips for the algorithms and models they have in mind.

Anari AI gives software-only programmers the power to create custom cloud chips for the algorithms and models they have in mind.

It’s worth noting: Virtual processor IP can be created 100 times faster than developing traditional custom hardware, and actual operating costs at the time of model deployment may be reduced to just one fifth of a corresponding GPU solution. The use of Python further dramatically increases the talent pool that would be capable of creating such solutions.

Taken together, Anari AI could provide the means for a large new wave of AI applications and the conditions for unprecedented non-linear gains in value creation.

This Serbian-American startup provides a new canvas for AI and Machine Learning, for daring ideas and unlimited imagination. CEO and co-founder Jovan Stojanovic, and all of us at Anari AI, are excited and curious to see what will develop. We see limitless possibilities. As the well-known naturalist and philosopher once said:

This world is but a canvas for our imagination.” — Henry David Thoreau

Find out more about Anari and join the team! 🚀. Keep up with more news and stories from the Earlybird portfolio on LinkedIn and Twitter.

References:

(1) https://www.investopedia.com/terms/m/mooreslaw.asp

(2) https://semiengineering.com/knowledge_centers/standards-laws/laws/dennards-law/

--

--

Earlybird Venture Capital
Earlybird's view

Earlybird is a venture capital investor focused on European technology companies. Read more at: https://medium.com/birds-view or www.earlybird.com