Blaize Unveils Graph Streaming Processor (GSP) for AI Workloads

TOKYO — AI accelerator company Blaize, formerly ThinCI, announced that its fully programmable Graph Streaming Processor (GSP) will enter volume production in Q2 2020.

The six‑year‑old startup has not disclosed detailed specifications, including power consumption and benchmark data. However, its test chip—taped out in mid‑2018 and run in a Linux‑based enclosure—has been fielded in 16 pilot projects around the globe for the past year, said CEO and co‑founder Dinakar Munagala.

Blaize characterizes the GSP as executing direct graph processing, managing and executing task graphs on‑chip, and enabling task‑level parallelism. In essence, the GSP addresses AI workloads that GPUs, CPUs, and DSPs have struggled to serve efficiently.

Industry analysts have noted that this proposition echoes familiar claims from other AI‑processor vendors.

Kevin Krewell, principal analyst at Tirias Research, said, “I know a bit about ThinCI, but never got the architecture pitch. I’m glad they changed the name though.”

The lack of technical details on the GSP architecture in its slide deck has fueled frustration and skepticism among analysts. Munagala, however, promised a comprehensive information release in Q1 2020.

Getting on a qualified vendor list

The good news for Blaize, according to Munagala, is that a growing roster of early adopters is already using its GSP. For the past year, Blaize has shipped a desktop unit equipped with the GSP. The unit can be plugged into a power socket and connected via Ethernet. Data scientists, software, and hardware developers are already evaluating system‑level functions enabled by the GSP.

Blaize, with $87 million in funding, is backed by early investors and partners including Denso, Daimler, and Magna. “We’ve also been generating revenue from the automotive segment for a few years,” said Munagala.

With a taped‑out chip in hand, many startups face a “What do we do now?” dilemma. Richard Terrill, vice president and strategic business development at Blaize, told EE Times, “We already passed that stage a year ago.”

Blaize has focused on expanding its infrastructure by bolstering an engineering team that now totals 325 members across California, India, and the U.K. It is relocating to new facilities and hiring field‑application engineers in Japan and EMEA. “We are maintaining our momentum,” said Munagala.

For Blaize, the GSP business is no longer about competing on specification sheets. It is about understanding how customers will use the GSP in real applications and measuring system‑level power consumption in those contexts.

Blaize has been busy finalizing logistics, obtaining automotive qualification, and ensuring internal processes and documentation meet certification standards. “We’ve already completed an auditing process and are on an approved and qualified vendor list of one automotive client,” said Munagala. This milestone is critical for carmakers and tier‑one suppliers that prefer partners with proven longevity.

Blaize hired 30 engineers in the U.K. (in Kings Langley and Leeds) to focus on automotive product development. The team, comprised of highly qualified individuals who previously worked at MIPS on MIPS‑based ASICs for Mobileye, brings deep domain expertise.

Graph computing

While AI spans many neural‑network types, “all neural networks are graph‑based,” explained Munagala. In theory, this enables developers to build multiple neural networks and entire workflows on a single architecture. Hence the GSP’s marketing slogan: “100 percent graph‑native.”

Blaize is not alone in the graph‑computing space. Competitors such as Graphcore, Mythic, and the recently shuttered Wave Computing have also discussed optimizing and compiling data‑flow graphs for AI processing.

Terrill noted, “Graph computing has a history of over 60 years.”

Blaize GSP distinguishes itself from other graph‑based data‑flow processors in three ways, according to Munagala. First, the GSP is fully programmable, capable of handling a wide range of tasks. Second, it is dynamically reprogrammable “on a single clock cycle.” Third, it integrates streaming to minimize latency. The data streaming mechanism reduces or eliminates non‑computational data movement, delivering a significant efficiency multiplier.

The graph‑native design of the GSP architecture minimizes data movement to external DRAM. Only the initial input and final output require external memory; intermediate data remains on‑chip, dramatically reducing memory bandwidth and power usage.

The stated objectives for Blaize systems are “the lowest possible latency, reduction in memory requirements, and energy demand at the chip, board, and system levels.”

Asked whether Blaize’s graph‑computing design is patent‑defensible, Munagala responded, “We are confident in our patent portfolio. We hold multiple granted patents and have additional applications pending, reflecting years of focused research.”