While ARM is focusing heavily on artificial intelligence (AI) as a growth driver for its processor IP, its former rival MIPS is heading in the same direction after being acquired by an AI start-up, Wave Computing of Campbell, California.
Reporters at EETimes discovered that Wave, which has developed a massively parallel dataflow architecture for AI, has acquired MIPS, and aims to harness its technology to expand its applications beyond its current market. For now, it focuses on supporting AI training in data centers, but wants to develop a unified platform that will power inference and analytics as well as training the model, and can be implemented anywhere from data center to edge. It also hopes, in time, to challenge specialized machine learning processors such as Google’s TPU.
MIPS was once a processor IP giant alongside ARM and PowerPC, but has lost ground in recent years partly because of its failure to challenge ARM in the mobile space. It was acquired by graphics processor IP supplier – and fellow ARM competitor – Imagination Technologies, but the latter was forced to divest several units, and put the whole firm up for sale, after its largest customer, Apple, switched to an inhouse GPU design.
MIPS was offloaded to Tallwood Venture Capital for a paltry $65m while the main Imagination company went to Canyon Bridge Capital Partners for $742.5m. In January, MIPS’ new owners and management relaunched the company, boasting that it had returned to its original Silicon Valley home after five years as a UK-owned outfit.
It was refocused around the IoT, and “processor innovations for a new generation of intelligent, connected platforms”, as Tallwood put it. As well as seeking to prolong the life of its processor designs in their traditional heartlands of set-top boxes and broadband routers, MIPS would now target robotics, AI, automotive, wearables, broader IoT opportunities, and even cellular modems.
But six months later, MIPS has another new parent, with a more specific AI focus. Wave Computing’s CEO Derek Meyer told EETimes that the purchase would help his firm to merge its scalable dataflow technology with MIPS’s RISC processors, and then harness the MIPS partner and customer ecosystem, to address a far larger market than it could do alone, in embedded systems.
Wave and MIPS are a similar size these days, each with about 100 staff, and both are partly funded by Dado Banatao, a managing partner at Tallwood and MIPS’ chairman. There are other connections – Meyer was once a VP for sales and marketing at MIPS, while his VP of operations, Mike Uhler, is a former MIPS CTO, and Darren Jones, VP of engineering, formerly performed the same role at MIPS.
It is unclear how an AI start-up will support and enhance products for MIPS’ far broader customer base – despite its waning market share, it still has large users such as Microchip, Intel’s Mobileye machine vision chip unit, MediaTek, and Japan’s Denso. That could be a challenge – even if Wave winds down the legacy business, it will be a long and delicate process beset by contractual arrangements. For now, it insists that will not happen and that MIPS will retain its brand and operate as a separate business unit which continues to license all the IP. Wave believes many CPU makers need an AI/ML roadmap and so some of MIPS’ licensees may welcome the integrated solution. However, it said licensees which choose to use their own, or another third party, AI accelerator will be free to continue to buy the MIPS CPU IP alone. Denso, for instance, has developed AI accelerators based on ThinCI technology.
As for the new products, Wave insists the integration will not be a steep hill to climb, and it has already worked to build hooks to connect Wave’s Dataflow Processing Units (DPUs) into a CPU core like MIPS’, to create AI-ready CPU cores.
Wave will deliver its first AI systems to early customers later this month and, at the same time, will announce a roadmap for the common MIPS CPU/Wave DPU platform, aiming to have that sampling by year end.
Meyer said this is needed by the industry, which is stuck with a “dual-architecture approach” in which deep learning runs in data centers on repurposed GPUs, while inference sits at the edge running on “a completely different hardware platform”. Wave says that its DPUs, like MIPS’ cores, are scalable and reconfigurable so can be used in a wide variety of scenarios from large servers to edge compute nodes to embedded devices, allowing AI to be supported wherever it is needed, using the same platform.
The challenge for Wave/MIPS will be to secure sufficient support for its integrated platform before the big names – from Google with its Tensor Processing Unit to Nvidia with its Xavier AI-driven system-on-chip, not to mention AI-centric cores from ARM and even Imagination – seize the space.