After months of searching for a buyer, troubled U.K.-based AI processor designer Graphcore announced on Friday that it has been acquired by SoftBank. The company will operate as a wholly owned subsidiary of SoftBank and will possibly collaborate with Arm, but it remains to be seen what will happen to the unique architecture of Graphcore’s intelligence processing units (IPUs).
Graphcore will retain its name as it becomes a wholly owned subsidiary of SoftBank, which paid either $400 million (according to EE Times) or $500 million (according to BBC) for the company. Over its lifetime, Graphcore has received a total of $700 million in investments from Microsoft and Sequoia Capital, and at its peak in late 2020, was valued at $2.8 billion. Nigel Toon will remain at the helm of Graphcore, which will hire new staff in its UK offices and continue to be headquartered in Bristol, with additional offices in Cambridge, London, Gdansk (Poland), and Hsinchu (China).
“This is a tremendous endorsement of our team and their ability to build truly transformative AI technologies at scale, as well as a great outcome for our company,” said Nigel Toon. “Demand for AI compute is vast and continues to grow. There remains much to do to improve efficiency, resilience, and computational power to unlock the full potential of AI. In SoftBank, we have a partner that can enable the Graphcore team to redefine the landscape for AI technology.”
Although Graphcore claims to have won contracts with major high-tech companies and deployed its IPUs, it could not compete against NVIDIA and other prêt-à-porter AI processor vendors due to insufficient funding. In recent years, the company’s problems were so severe that it had to lay off 20% of its staff, bringing its headcount to around 500. These cuts also saw office closures in Norway, Japan, and South Korea, making it even harder to compete against big players.
Graphcore certainly hopes that, with SoftBank’s deep pockets and willingness to invest in AI technologies in general and AI processors in particular, it will finally be able to compete head-to-head with established players like NVIDIA.
When asked whether Graphcore will work with SoftBank’s Arm, Nigel Toon said that he looked forward to working with all companies controlled by its parent, including Arm. Meanwhile, SoftBank itself is reportedly looking forward to building its own AI processor venture called Project Izanagi to compete against NVIDIA, whereas Arm is reportedly developing AI processors that will work in datacenters owned by SoftBank. Therefore, it remains to be seen where Graphcore fits in.
For now, the best processor that Graphcore has is its Colossus MK2 IPU, which is built using 59.4 billion transistors and packs 1,472 independent cores with simultaneous multithreading (SMT) capable of handling 8,832 parallel threads. Instead of using HBM or other types of external memory, the chip integrates 900 MB of SRAM, providing an aggregated bandwidth of 47.5 TB/s per chip. Additionally, it features 10 IPU links to scale with other MK2 processors. When it comes to performance, the MK2 C600 delivers 560 TFLOPS FP8, 280 TFLOPS FP16, and 70 TFLOPS of FP32 performance at 185W. To put the numbers into context, NVIDIA’s A100 delivers 312 FP16 TFLOPS without sparsity as well as 19.5 FP32 TFLOPS, whereas NVIDIA’s H100 card offers 3,341 FP8 TFLOPS.
