For Twigfarm, a Seoul-based AI translation startup, success is a balancing act. The company’s ambitious and important mission depends on its ability to deliver high-accuracy results as quickly as possible. At the same time, as a fast-growing business, operational expenses need to be carefully managed.
When Twigfarm recently transitioned from GPUs to Graphcore’s Intelligence Processing Units (IPUs) no price-performance trade-off was needed. Quite the opposite—switching to IPUs delivered a 10x performance gain, while simultaneously lowering costs.
Communication without borders
Twigfarm is the brainchild of entrepreneur Sun Ho Baek, who founded the company in 2016 with the goal of harnessing AI to help remove language barriers for people and organisations around the world. This vision stemmed not only from Sun Ho’s professional background as a software engineer, but also from his own first-hand struggles: having spent extended early career stints in Israel and Nepal, he had come to realise just how much potential could be unlocked if communication could be made even a little more seamless.
This realisation inspired Sun Ho to develop what would become Twigfarm’s cornerstone technology for text-based translation—a proprietary hybrid translator and natural language processing (NLP) engine. In Twigfarm’s first six years of life, the company has turned this technology into an array of services including LETR, a neural network-based language processing engine; Gcon Studio, a bespoke B2B translation platform for enterprises; and heybunny, an B2C enterprise solution that translates newsletters and emails.
Managing cost-intensive training workloads
As Twigfarm’s products continue to receive awards, technical accreditations, industry attention, and plaudits from customers, Sun Ho remains keenly aware of the need to prevent the company’s operational costs from growing in tandem with the company’s level of success and recognition. As he explains, “Startups are particularly sensitive about expenses, as our businesses need to operate with very limited resources.”
This is especially challenging given the time and cost-intensive nature of the services Twigfarm provides for its customers. In addition to gathering and preprocessing data, Twigfarm’s models need to be retrained continuously based on customer feedback. “We have to train the models as quickly as possible and deliver them to our customers,” notes Sun Ho. “This means minimising the time spent on training.”
Until recently, Twigfarm had been training AI models using a cloud-based solution powered by GPUs, but this entailed a prohibitive cost burden, leading Sun Ho to search for compute alternatives better suited to Twigfarm’s products. As Sun Ho recalls, “The costs would sometime reach tens of thousands of dollars, so we were trying to work out how to achieve better performance in a cost-effective way.”
Looking beyond GPUs
The search for a performance and efficiency improvement over cloud-based GPUs quickly led Sun Ho to Graphcore’s IPUs. IPUs were ideally suited to solve multiple problems at once: improved efficiency, lower total cost of ownership (TCO), and faster time to result. The competitive advantage of IPUs is particularly clear in NLP applications, as demonstrated by the impressive performance of IPUs in the most recent MLPerf results for the BERT NLP model.
Before unlocking these improvements, however, Sun Ho needed to ensure that any transition from GPUs to Graphcore’s IPUs would be seamless and involve relatively light levels of technical training, code changes, and optimisations.
In this regard, he feels Graphcore excelled. “When migrating to other systems, most of the time you’re left on your own or you have to follow whatever documentation has been made publicly available,” Sun Ho says. “By contrast, Graphcore provided us with related documents as well as technical support. In terms of code changes, they didn’t just explain it—they also went over the code with us in depth and resolved any issues we encountered.”
A growing partnership
Twigfarm was able to complete its migration smoothly, and has been operating on Graphcore’s cloud-based IPUs and Poplar software since November 2021.
The company’s operations have already benefited substantially from the shift, as Sun Ho attests: “We found that the IPU outperformed all other systems we had used before. The bottom line is that Graphcore’s IPU enables us to get a 10x performance gain while significantly reducing costs.”
Looking to the future, Twigfarm plans to develop additional services using Graphcore’s IPU technology. Sun Ho sees a bright future for the partnership: “Graphcore is a global company and a growing startup,” he says. “With their support and advice, we hope to grow move forward together onto the global stage.”