Edge AI chip startup Blaize has successfully listed on the NASDAQ after merging with special purpose acquisition company BurTech. While several of the new breed of AI chip startups that have emerged in the last 15 years have had exits, Blaize is the first to go public.

Blaize, founded in 2011 as ThinCI, is one of the most mature AI chip startups out there; the company will reach 14 years old this year. Semiconductor companies that IPO typically do so around 7-10 years after founding. Cerebras, another AI chip startup benefiting from strong GCC investments and customer base, was founded in 2015 and filed its S-1 9 years later, for example (it has not yet gone public).

Blaize had raised $335 million in funding, including $106 million in April 2024, after announcing plans to merge with BurTech. The company was expected to be valued at around $1.2 billion after the merger.

Blaize’s financials will do little to quiet industry criticism of AI chip startups as being overhyped. The company’s prospectus shows it made just $28,000 revenue on hardware sales in the first three quarters of 2024 (the bulk of the company’s $1.5 million in revenues for the same period is from engineering services, including software licensing).

Blaize’s chip, which launched in 2020 but hit the market in 2022, has made $909,000 revenue total. According to its prospectus, the majority of this revenue was generated in the United States across 11 design wins. At present, 23 proofs-of-concept have been completed for customers in North America, Japan, Korea, the EU and the GCC.

Blaize’s prospectus gives a little information about its pipeline, which it estimates at $458 million.

However, its contracts with strategic investors Denso and Mercedes Benz, which have generated revenues of $25 million and $7 million in engineering fees to date, respectively, will not result in chip sales until an automotive-grade version of the Blaize chip can be delivered. The prospectus put this date at 2028 at the earliest (Blaize’s chip has been designed with ASIL-D compliance in mind, the company told EE Times in 2021, but certification can take years). Mercedes-Benz and its suppliers are evaluating Blaize technology for a future L4 driving platform, which is due to launch by the end of the decade.

Blaize signed a multi-year MoU with strategic investor Mark AB Capital (a fund controlled by members of the Abu Dhabi Royal Family) in September 2023. This MoU had Blaize creating an AI data center and developing various smart city and security applications for the UAE government. Blaize and Mark AB Capital were also working together to create a training center for UAE citizens based on Blaize’s AI Studio—a no-code platform for developing AI applications on hardware including Blaize chips—with the aim of training a workforce of 5,000 UAE citizens.

Shahal Khan, speaking at the time on behalf of Mark AB Capital, told EE Times in September 2023 that this investment and MoU came out of the UAE’s desire to become the world’s first “smart nation.” (Khan is also CEO of BurTech, the SPAC company Blaize has merged with.)

“We first looked at Blaize for a strategic investment because the UAE could become a smart nation, not just a smart city,” he said, noting that the UAE is one of the most network-connected countries in the world because of its size. “There are various implementations in the infrastructure on the edge that we could do straight away with what [Blaize] has developed.”

Khan’s examples include national security, traffic monitoring, smart buildings and autonomous cars.

“The last piece of the puzzle for us is to train and create a development core platform for our citizens to get accredited on and hopefully create a workforce of AI software developers using AI Studio out of the UAE that could service not only the UAE but the GCC and other growing markets,” Kahn said. “Once we establish this as a good working model in the UAE, [Mark AB Capital and Blaize will] take it out to other governments in our geography and help them accelerate services created with Blaize technology into adoption.”

The companies said at the time that they expected to generate $50 million in orders annually from GCC companies as a result of this MoU over several years. So far, it seems that is still in the works. Blaize’s prospectus says the company has completed a drone detection and video surveillance proof-of-concept for a GCC government entity. The company has also received a purchase order with a UAE defense supplier which is worth “up to $105 million,” but is currently pending an updated proof-of-concept, due by April 2025.

While Blaize’s 16 TOPS chip is not currently covered by U.S. export restrictions, these restrictions can change quickly.

Graph streaming processor

Blaize’s first-generation graph streaming processor (GSP) chip launched in 2020, and if we assume a 2-year design cycle, is now 7-years-old. While the computer vision use cases Blaize specializes in were already fairly well established in 2018, 7 years is a long time in AI. Not having produced an updated second-generation chip in that time frame is therefore unusual.

Roberto Mijat, VP of marketing at Blaize, told EE Times that the company is working on a second-generation chip, but that the first-generation GSP can handle some smaller multi-modal networks like vision-language models.

“[Unlike others], our choice has been programmability,” he said. “Our emphasis and investment has been on software and the tools to make it easier and optimal to deploy that on the hardware…there is a lot more longevity in [our first generation chip], customers can take it and deploy it and then they can keep it for a long time because the emphasis is on upgrading software. So we haven’t really been in a situation where we’ve been forced to get new hardware out there.”

At the AI Hardware Summit in September, Blaize demonstrated generative AI inference running on a single Blaize chip (the 300M vision-language model Florence-2, which describes input videos in text).

“You can do useful things with small models now,” Mijat said.

Blaize also demonstrated its four-chip PCIe cards for EE Times. Six 4-chip cards (24 chips) fit into a 1U air-cooled server.

“To do the equivalent amount of work to what this unit can deliver, you probably need a 4U or 8U unit equipped with GPUs, because they need more space and more cooling,” Mijat said at the time. “Our value proposition is built around TCO. Our deployment is smaller and cheaper—it’s about $30,000 to put together a box like this based on an off-the-shelf Supermicro blade.”

In one of Blaize’s demos, 1,000 video streams were being processed concurrently in a scenario that would suit central processing of camera footage in retail stores or airports, Mijat added. The company also has a 4U combination training and inference (GPU and Blaize-based) server in the works with Supermicro, and is working closely with other server OEMs.

Blaize has partnerships with more than a dozen ISPs, Mijat added, including companies like security camera system front-end provider Milestone Systems and defense middleware and software provider Vantiq.

First generation

Blaize’s GSP chip offers 16 TOPS (INT8) performance in a power envelope of 7 W, which the company’s claims of performance and TCO advantages are based on.

The architecture is designed for graph processing, whether that is neural networks or vision pre-processing algorithms, Val Cook, chief software architect at Blaize, told EE Times during an earlier interview.

“We don’t compile for our chip,” he said. “What I mean by that is, the chip gets a linked structure of control blocks and each of those blocks represents a node in the data flow graph, and all of the scheduling occurs on chip.”

Most other graph-based architectures use software to statically map the graph to the silicon, assigning hardware resources to portions or layers of the neural network calculation.

“With Blaize, it doesn’t work that way—[initially], all [the chip’s resources] are on deck getting that first layer moving, but as soon as the mathematical data dependencies are ready, we’ll start peeling off processors one by one to start servicing the next layer while the first layer is still dominant,” Cook said. “Pretty soon, the machine will dynamically load balance, to the point where you’re matching the peak throughput to minimize the latency for that particular graph, and that’s why it works so well.”

The chip effectively compiles the graph at runtime, Cook said. One of the things this allows Blaize to take advantage of is data sparsity—in Cook’s example, a car driving uphill where most of the image will be sky.

“Dynamic execution ends up being a lower energy solution for higher throughput,” he said.

Blaize CEO Dinakar Munagala, in an earlier interview with EE Times, said that a significant part of graph processing is today done on the host CPU.

“In the case of Blaize, that’s all moved to hardware, thereby you get significant efficiency,” he said. “In real customer workloads we’ve seen for autonomous driving, we’ve seen 80-plus points per frame where control is handed back between the CPU and GPU. [With Blaize], that’s been cut to one, and as workloads become more complex, that means more advantage for Blaize.”

Blaize has two parallel software stacks. One of these stacks, AI Studio, supports Blaize and other hardware targets and is offered as a standalone product. Software licensing is not broken out as a separate line in the Blaize prospectus, so it is difficult to say how successful this product has been since its launch in 2020.

AI Studio is intended for no-code development of AI applications, and includes lifecycle management software for deployment of AI models at scale, data management and curation before and after deployment (including patented data and model drift detection), and enables retraining when necessary, Dmitri Zakharchenko, VP of research and product development, told EE Times in an earlier interview.

“AI Studio is designed to do a lot more than just retraining, it does data drift and model drift detection,” he said. “AI Studio continuously monitors applications in production and when it detects drift we’re able to retrain the model on the new data and put it back into production…it’s redeployed until the next time something changes. This is a very important part a lot of other companies tend to skip.”

The company’s other software toolchain, Picasso, is a developer-focused SDK, including high level libraries for various use cases and end markets.

As the first AI chip company to go public, semiconductor investors will no doubt be watching the performance of Blaize stock in the coming months. While SPACs surged in popularity in 2020-2021 as a faster route to an IPO, they fell out of favor in the investment community after many stocks had disappointing performance.

Blaize shares were initially trading at $17.50, but had dropped to $7.55 at the time of writing, a week after listing.

By: DocMemory
Copyright © 2023 CST, Inc. All Rights Reserved