Blaize AI: Now In Production And Trials

by | Aug 31, 2020 | AI and Machine Learning, In the News, Semiconductor

Last November I covered Blaize and its silicon and software strategy, and noted that the company’s fairly large team has been focused on early customer engagements to gain insights and accelerate adoption. Now the company, backed by industrial heavyweights such as Samsung, Daimler and Denso, is advancing from development into full production and customer deployment. Blaize claims its silicon is designed to be a “no-compromise” platform that delivers the performance customers need, while delivering it at low cost and low power. From what we now know, the company’s claims seem to have merit. Let’s take a look at what they have recently announced.

What did Blaize announce?

Blaize (formerly Thinci) previewed its Graph Streaming Processor (GSP) chip and Picasso software last fall when the company rebranded itself. The recent announcements disclosed availability (sampling now), efficiency (> 2 TOPS/Watt), pricing (starting at $299) and its new AI Studio software which complements the previously announced Picasso SDK. Blaize also announced an EDSFF card, which can be used with or without a server, and a half-height PCIe card that supports 1-4 GSPs, delivering up to 64 TOPS. This ready-to-deploy approach should simplify adoption and speed time to solution for Blaize’s clients and is a productization approach being adopted by most AI chip vendors (including NVIDIA who took it to the next level with DGX systems). Blaize also shared five on-going customer trials in industrial, smart city, sensor fusion, last mile delivery and retail applications.

Blaize is sampling its GSP chip in two flavors: an SDSFF card and a 1-4 chip PCIe Gen 3 card.  image: BLAIZE

The GSP looks to be quite efficient based on the recent disclosures, delivering 16 Trillion Operations Per Second (8-bit integer TOPS), and using 7 watts of power (presumably at the chip level). The chip is a fully programmable graph processor, so it can handle an entire job that would otherwise require some sort of host processor. Meanwhile, it keeps power and cost down to a minimum for applications such as smart cities. To my eye, this chip looks a bit like a smaller and simpler version of the Graphcore chip. Both are native graph parallel execution engines, with Graphcore targeting training at very high scale over its fabric, and Blaize focusing on edge processing at low power and cost.

Figure 2: Blaize contends its GSP does not compromise between performance, power and cost, although it did not disclose actual pricing. image: BLAIZE

Blaize was able to show several use cases where its chip is in trial, potentially for mass market adoption as soon as by the end of this year. The case that really got my attention was a smart city deployment where the Blaize Xplorer EDSFF card is paired with 5 video cameras to monitor safety and security at traffic intersection. The GSP handles the sensor fusion, DNN inference processing and communication of alerts, all without CPU involvement. So, for example, if someone is looking at a cell phone and steps out onto the street against traffic, the GSP can activate an audible alarm while turning all traffic signals to red. And if the worst should happen and a pedestrian is seen to be lying on the pavement, the system can again stop traffic and alert medical and emergency personnel to the site. The platform also recognizes license plates for security and policing support. Blaize shared other examples of edge intelligence which I will explore in-depth in a future blog.

Blaize also announced the new AI Studio, which enables “code-free” development of edge applications, complementing the updated Picasso software suite. The idea here is to simplify the task of the data scientist, who isn’t an expert coder, helping to get an application up and running quickly. Blaize claims AI Studio can reduce development time from months to days. The Picasso software enables programmers to optimize and deploy sophisticated models.


Blaize now boasts impressive AI technologies, including its enhanced software stack, that could open new doors for smart edge applications. Consequently, it has nurtured customer engagements and appear well positioned to begin deployments at scale in the next six to nine months.

So, has the promised Cambrian Explosion of AI chips and NVIDIA challengers finally begun? Not really, at least not as I had earlier anticipated. But Blaize is off to a strong start, and there are some interesting new technologies being discussed at the upcoming HotChips conference and the annual AI Hardware Summit (both virtual, of course). Meanwhile the market for AI at the edge is just emerging and there will soon be a lot of players jockeying for position. Based on these new announcements, Blaise looks like a real contender in this race and has the industrial backing to deliver exciting real-world solutions. Hope to “see” you at Hot Chips and the AI HW Summit!