Blaize describes its GSP as capable of "direct graph processing, on-chip task graph management and execution, and task parallelism." In short, Blaize designed the GSP to meet processing demands that AI, GPUs, CPUs, or DSPs could not previously meet.
For many industry analysts covering embedded AI processors, this is a topic they've heard before.
"I knew about ThinCI, but never the architecture. I'm glad they changed the name," said Kevin Krewell, principal analyst at Tirias Research.
The lack of technical details about GSP's architecture in the GSP slides has caused frustration and skepticism in the tech analyst community. However, Mungagala promised to release the information in the first quarter of 2020.
The GSP architecture consists of a series of graphics stream processors, dedicated math processors, hardware control and various types of data caches. The company claims that GSP can provide: "True task-level parallelism, minimal use of off-chip memory, depth-first hardware graph scheduling, and a fully programmable architecture."
Obtain a list of qualified suppliers
The good news for Blaize, in Munagala's view, is the large number of early customers already using GSP. For a year now, Blaize has been shipping desktops with GSP. It can simply be plugged into a power outlet and connected to Ethernet. Data scientists, software and hardware developers are already evaluating the system-level capabilities enabled by GSP, Munagala said.
Blaize, which has $87 million in funding, is backed by early investors and partners including Denso, Japan's Daimler and Magna. "A few years ago, we were still making money in the automotive space," Munagala said.
With tapes in hand, many startups face the "what are we going to do now?" dilemma. "We were past that stage a year ago," Richard Terrill, Blaize's vice president and manager of strategic business development, told EE Times.
Blaize has shifted focus to building infrastructure by beefing up an engineering team (now 325 employees) and expanding to California, India and the UK. Blaize is moving to a new facility and has started hiring Field Application Engineers in Japan and EMEA. "We're keeping the momentum going," Munagala said.
For Blaize, its GSP business is no longer about competing with competitors in the Powerpoint presentation specification. It's about figuring out how customers will use GSP for which applications, and how much power is being consumed "at the system level" in a particular use.
Blaize has been busy figuring out its logistics, bringing its products up to automotive industry standards and ensuring internal processes and documentation are certified. "We have gone through the audit process and are on an approved list of qualified suppliers," Munagala said. It's a process that automakers and tiers of automakers have to go through, and they'd rather avoid startups that don't last long enough to deliver a product.
Blaize employs around 30 engineers in the UK, based in Kings Langley and Leeds, working on product development for the car. When Imagination dropped MIPS, they were a close-knit team of engineers. “These are a group of highly qualified people who worked together on MIPS to bring a MIPS-based ASIC to Mobileye’s automotive standards,” Munagala explained.
graph computing
While embedded AI appears in many different types of neural networks, "all neural networks are graph-based," Munagala explained . In theory, this allows developers to leverage graph-native structures to build multiple neural networks and entire workflows on a single architecture. Hence, the company's new marketing strategy for its GSP is "100% Graphics Localization".
However, Blaize is not exactly a unicorn in the field of graph computing. Graphcore, Mythic, and the now-failed Wave Computing all talk about "optimization and compilation of dataflow graphs" in embedded AI processing.
"Certainly, graph computing has been around for over 60 years," Terrill said.
Blaize GSP claims to be different from other graph-based dataflow processors in three ways, Munagala said. First, "our GSP is fully programmable" and capable of performing "a wide variety of tasks."
Second, it is "dynamically reprogrammable in a single clock cycle".
Third, "we provide streaming integration," which makes it possible to minimize latency. The huge efficiency multiplier, he explained, is delivered through a "dataflow mechanism," in which the movement of non-computational data is minimized or eliminated.
The graphics-native nature of the GSP architecture minimizes data movement to and from external DRAM. The outer only needs the first input and last output, and everything else in between is just temporary intermediate data. This results in greatly reduced memory bandwidth and power consumption.
The stated goal of the Blaize system is to "achieve the lowest possible latency at the chip, board, and system levels, reducing memory requirements and energy requirements."
When asked if Blaize's graphics computing design was patentable, Mungala said: "We are very confident in our patent portfolio. We have several patents, some granted and others pending, but we have Been doing it for years."
Embedded Internet of Things needs to learn a lot. Don't learn the wrong route and content, which will cause your salary to go up!
Share a data package with everyone, about 150 G. The learning content, face-to-face scriptures, and projects in it are relatively new and complete! (Click to find a small assistant to receive)
