Troubled AI Processor Developer Graphcore Finds a Purchaser: SoftBank
After months of looking for a purchaser, troubled U.Okay.-based AI processor designer Graphcore stated on Friday that it has been acquired by SoftBank. The corporate will function as a completely owned subsidiary of SoftBank and can presumably collaborate with Arm, however what stays to be seen what occurs to the distinctive structure of Graphcore’s intelligence processing items (IPUs).
Graphcore will retain its title as it’ll turn into a completely owned subsidiary of SoftBank, which paid both $400 million (in line with EE Times) or $500 million (in line with BBC) for the corporate. Over its lifetime, Graphcore has acquired a complete of $700 million of investments from Microsoft and Sequoia Capital, and at its peak in late 2020, was valued at $2.8 billion. Nigel Toon will stay on the helm of Graphcore, which is able to rent new workers in its UK workplaces and proceed to be headquartered in Bristol, with further workplaces in Cambridge, London, Gdansk (Poland), and Hsinchu (China).
“It is a super endorsement of our staff and their capacity to construct really transformative AI applied sciences at scale, in addition to a fantastic final result for our firm,” stated Nigel Toon. “Demand for AI compute is huge and continues to develop. There stays a lot to do to enhance effectivity, resilience, and computational energy to unlock the total potential of AI. In SoftBank, we’ve got a associate that may allow the Graphcore staff to redefine the panorama for AI know-how.”
Though Graphcore says that it had gained contracts with main high-tech corporations and deployed its IPUs, it couldn’t compete in opposition to NVIDIA and different prêt-à-porter AI processor distributors as a result of inadequate funding. Within the latest years the corporate’s issues have been so extreme that it needed to lay off 20% of its workers, bringing its headcount to round 500. These cuts additionally noticed workplace closures in Norway, Japan, and South Korea, which made it even tougher to compete in opposition to massive gamers.
Graphcore definitely hopes that with SoftBank’s deep pockets and willingness to spend money on AI applied sciences normally and AI processors specifically, it’ll lastly be capable of compete head-to-head with established gamers like NVIDIA.
When requested whether or not Graphcore will work with SoftBank’s Arm, Nigel Toon stated that he was trying ahead to work with all corporations managed by its dad or mum, together with Arm. In the meantime, SoftBank itself is reportedly trying ahead to construct its personal AI processor enterprise referred to as Project Izanagi to compete in opposition to NVIDIA, whereas Arm is reportedly developing AI processors that will work in datacenters owned by SoftBank. Subsequently, it stays to be seen the place does Graphcore slot in.
For now, the most effective processor that Graphcore has is its Colossus MK2 IPU, which is constructed utilizing 59.4 billion transistors and packs in 1,472 impartial cores with simultaneous multithreading (SMT) able to dealing with 8,832 parallel threads. As an alternative of utilizing HBM or different forms of exterior reminiscence, the chip integrates 900 MB of SRAM, offering an aggregated bandwidth of 47.5 TB/s per chip. Moreover, it options 10 IPU hyperlinks to scale with different MK2 processors. With regards to efficiency, the MK2 C600 delivers 560 TFLOPS FP8, 280 TFLOPS FP16, and 70 TFLOPS of FP32 efficiency at 185W. To place the numbers into context, NVIDIA’s A100 delivers 312 FP16 TFLOPS with out sparsity in addition to 19.5 FP32 TFLOPS, whereas NVIDIA’s H100 card presents 3,341 FP8 TFLOPS.
