TL;DR:
- CentML, an AI startup, has raised $27 million in an extended seed round.
- Investors include Gradient Ventures, TR Ventures, Nvidia, and Microsoft Azure AI VP, Misha Bilenko.
- The funding brings CentML’s total raised to $30.5 million, which will be used for product development and expansion.
- CentML’s mission is to reduce machine learning costs and address chip shortages.
- The company’s software optimizes model training and promises up to an 80% cost reduction without compromising speed or accuracy.
- The AI chip supply problem has driven major companies to explore custom chip solutions.
- CentML distinguishes itself by not compromising model accuracy and offering a high-performance compiler.
- The startup plans to expand into optimizing inference for AI models.
- The growing AI-focused chip market is expected to reach $53 billion this year, with significant growth predicted.
Main AI News:
In the world of AI, large seed rounds are far from obsolete, and CentML is here to prove it. This innovative startup, dedicated to creating tools that drive down the cost while boosting the performance of deploying machine learning models, recently announced a substantial injection of $27 million in an extended seed round. Notable contributors to this funding include Gradient Ventures, TR Ventures, Nvidia, and Microsoft Azure AI VP, Misha Bilenko.
CentML’s initial seed round was concluded in 2022, but the company decided to extend it over the past few months as interest in its pioneering product continued to surge. This extension brought their total funding to an impressive $30.5 million. The newly acquired capital is earmarked for strengthening CentML’s product development and research initiatives, as well as expanding its workforce, which currently consists of 30 individuals spread across the United States and Canada.
According to Gennady Pekhimenko, Co-Founder and CEO of CentML, the startup’s mission is to address the challenges posed by rising machine learning costs, talent shortages, and chip supply issues. Pekhimenko, an associate professor at the University of Toronto, founded CentML alongside Akbar Nurlybayev and PhD students Shang Wang and Anand Jayarajan. Their shared vision was to develop technology that would enhance access to computing power in the face of a growing AI chip supply problem.
Pekhimenko highlighted the common predicaments faced by AI and machine learning companies, including the unavailability of high-end chips due to soaring demand from both enterprises and startups. This scarcity often forces companies to compromise either on the scale of their deployed models or on the inference latencies they experience.
The backbone of many model training processes, particularly those involving generative AI models like ChatGPT and Stable Diffusion, relies heavily on GPU-based hardware. The parallel processing capabilities of GPUs make them an ideal choice for training cutting-edge AI. However, the scarcity of these chips has become a pressing issue, with companies like Microsoft and Nvidia facing significant hardware shortages.
Microsoft recently cautioned about the possibility of service disruptions due to a severe shortage of server hardware required for AI operations, while Nvidia’s top-performing AI cards are reportedly unavailable until 2024. In response, major players like OpenAI, Google, AWS, Meta, and Microsoft have explored developing custom chips for model training, but this strategy has encountered its own set of challenges.
Pekhimenko recognized the urgency of launching software capable of optimizing existing hardware to run AI models more efficiently, especially in a market where AI-focused chip spending is projected to reach $53 billion this year, with further exponential growth predicted by Gartner over the next four years.
With CentML’s optimization technology, Pekhimenko claims that expenses can be reduced by up to 80% without compromising speed or accuracy—an ambitious promise that has caught the attention of the AI community.
CentML’s software is designed to identify bottlenecks during model training and predict the time and cost required to deploy a model accurately. Furthermore, it offers access to a compiler—a component that translates programming language into machine code, making it compatible with GPUs. This compiler automatically optimizes model training workloads to deliver peak performance on target hardware.
Pekhimenko emphasized that CentML’s software does not compromise model accuracy and requires minimal effort from engineers to implement. She cited an example where CentML optimized a customer’s Llama 2 model to operate three times faster using Nvidia A10 GPU cards.
While CentML faces competition from other software-based model optimization solutions, such as MosaicML and OctoML, Pekhimenko asserts that CentML’s techniques do not sacrifice model accuracy, unlike some of its rivals. Additionally, CentML’s compiler is described as a “newer generation” and more performant than that of OctoML.
Looking ahead, CentML plans to expand its focus from model training to optimizing inference, the process of running models after training. GPUs are widely used for inference, and Pekhimenko sees this as a potential avenue for the company’s future growth. She explained that the CentML platform can run any model, producing optimized code for various GPUs and reducing the memory requirements for deploying models. This flexibility enables teams to deploy on smaller and more affordable GPUs, further enhancing efficiency in the AI landscape.
Conclusion:
CentML’s successful funding round and its mission to make AI model deployment more efficient reflect the continued growth and investment in the AI sector. As companies grapple with chip shortages and escalating costs, CentML’s optimization technology could have a significant impact on reducing expenses while maintaining high-performance AI models. This development underscores the market’s need for innovative solutions to tackle pressing challenges and improve the accessibility of AI technologies.