Artificial Intelligence has been extolled as the next big thing in human progress, primed to transform industries. Adoption of AI has blown up in the last few years, making its way into the mainstream in countless business applications. This increasing penetration of AI’s influence in business is driving powerful change, automating processes and streamlining operations.
Abacus Semiconductor Corporation is re-imaging the application of Artificial Intelligence. As a fabless semiconductor company, Abacus Semi is focused on creating a line of processors, accelerators and AI-driven designs that address the gap between the theoretical peak performance and real-life performance of today’s supercomputers.
For its unique designs and beyond-von-Neumann and beyond-Harvard CPU architecture, Abacus Semiconductor Corporation is positioned at the leading edge of innovation, making a big impact on the semiconductor market.
The CEO Views spoke to Axel Kloth, CEO of Abacus Semiconductor Corporation to delve into his remarkable narrative and uncover how the company is making strides in AI.
Today, everybody talks about AI and in particular, generative AI. Why is that, and how do you think that companies will have to react to this?
Any CEO and CIO today needs to evaluate how they can deploy Artificial Intelligence (AI) in their business today. There are likely many situations in which AI can already be deployed today in production and office environments, taking over routine and boring tasks that employees increasingly refuse to carry out. The earlier a company starts doing this, the more quickly employees get used to this new technology, and the less resistance the executive management will feel when AI is introduced on a broader scale. If AI is not at least being tested today, I would ask the question if the company will be able to fend off competitors that do.
Would not that require vast additional computational resources?
This does not require large-scale AI or Artificial General Intelligence (AGI). It can be accomplished using special-purpose, very targeted AI models that can be created in-house, thereby alleviating the danger of disclosing the inner workings of the company or any other secrets. These foundational models do not require the excessive computational performance that GPT4 needed, and as such generating those foundational models can be developed on a budget in an on-premises data center. Those are much more useful and much easier to adapt to the business needs of a company and of a business unit of companies.
For that reason, Abacus Semi believes that the modest infrastructure needed for special purpose and targeted foundational models will become corporate commonplace and will be added to the data centers that are currently being re-homed from the cloud. That means that both inference and training will be executed on premises (or short on-prem), and that also means that the risk of leaking corporate secrets is reduced over cloud-based use of AI. We are convinced that AI will play an ever-increasing role in all kinds of companies of all sizes, and that it will become an essential part of doing business.
Do we really need to start using AI right now?
While some people may feel that the use of AI is not yet warranted, the matter of fact is that the genie is out of the bottle, and we are not going to be able to put it back in. What we as a society, as developers of AI hardware and AI solutions can and should do though is to define a set of rules for the ethics and the morals of AI systems and most importantly, the models, under deployment. Without that, AI could become detrimental instead of helpful, and we of course also have to acknowledge and understand and fend off AI being used by any adversaries, be it business, political or national adversaries. As such, we have to be able to identify adversarial use of AI, and develop defenses against it. However, those are societal and political tasks, and not technological challenges.
What do we have to do, and how do we achieve that? What does the industry need to do to make AI useful and safe and secure?
I agree that the industry needs to do much more to make AI useful and safe, fair, unbiased and secure. With that being said, the use of AI is going to be universal and ubiquitous. In order to make it work on a budget for the necessary AI hardware acquisition and for the longer-term operation of that hardware, AI for training and for inference must be vastly more energy-efficient than what we observe today. Large-scale AI hardware must scale more linearly over the number of processors and accelerator cores deployed, and its power consumption must be substantially improved over existing solutions. That implies that a new generation of AI hardware architecture must be deployed as current solutions will not work. At Abacus Semi, we have developed such solutions, and we are working with the industry to standardize them. We are also very active in making sure that standards are being developed to ensure that model generation is free of bias to the degree that we can do that. We try to foresee the known unknowns, but what will always elude us are the unknown (and unknowable) unknowns.
How do you envision the future? What do you think it takes to get to a point where the computational resources for AI use, including the generation of foundational models, can be reasonably done within each company that has decided to go for it?
We see a number of trends. The first one is the re-homing of IT into on-premises data centers. The second is that more and more users, especially CTOs, VPs of Engineering, and CIOs, understand that accelerated compute is the future, not CPU-based compute.
What we focus on is a near-linear scale-out of performance over the number of cores. We all know that computational challenges are not going to become smaller in the future. They will become bigger, and in fact, we believe they will become much bigger since we see exponential growth, not linear growth. If that is the case, then a low-latency and high-bandwidth interconnect between processor cores and accelerator cores as well as smart shared memory is essential, not the performance of a single core.
That is where our expertise is, and in specific accelerators, such as those numerical accelerators needed for the math in the generation of foundational models. That math is largely based on vector math, matrix math, tensor math, and on different types of transforms. The founder and design engineers have a long tradition working with advanced math, and we have made sure to optimize both performance and power consumption when executing these operations. All of our technology is developed in-house, and we have a large number of patents pending covering our intellectual property. As such, we can bring down the long-term cost of an AI data center.
We believe that the Total Cost of Ownership (TCO) of nearly any data center today is too high, largely due to the excessive power consumption of prevailing solutions, and associated with that power consumption is of course the incidental heat generation which must be removed from the data center, again costing electric energy and thus money. While the acquisition cost might not differ much between our solutions and the existing prevailing ones, we can scale out to higher performance at lower power consumption, thereby reducing TCO over the typical life of a data center.
“All of our technology is developed in-house, and we have a large number of patents pending covering our intellectual property.”