Demonstrating the speed at which AI can accelerate workloads, two engineers recently designed and verified a chip design in under two weeks. But what exactly did the engineers build, and what does this say about the role of AI in engineering work?
Engineers Use AI to Create Chip in Two Weeks
A US startup called Architect Labs recently claimed that two engineers used an AI-based workflow to take its Redwood chip from an architectural specification through RTL design, firmware and verification in roughly two weeks.
According to the company, the AI workflow performed most of the low-level design and code-generation work, allowing the engineers to focus on higher-level tasks. More interestingly, the system was able to automatically check the resulting design, respond to changes in the specification, regenerate the hardware and then rerun verification.
This effectively creates a feedback loop where engineers can change what they want the chip to do, with the AI then updating the underlying hardware and checking that the changes still work.
Architect Labs claims that as many as 115 hardware changes were merged in a single day, while the resulting module achieved 95% coverage. However, these figures have not been independently verified. The Redwood chip has already been tested on an AMD Versal FPGA, where it successfully ran multiple open-source AI models, including Meta's Llama and Alibaba's Qwen.
However, it should be noted that Redwood is not yet a manufactured ASIC or taped-out chip. The FPGA demonstration shows that the design can function on programmable hardware, but it does not prove that the same architecture will achieve the same performance, power consumption or efficiency once manufactured.
Architect Labs currently projects that a future version of Redwood manufactured on Samsung's 8nm process could deliver 1.75× the inference throughput of Nvidia's Jetson Orin Nano while using around 47% of its power. These figures remain projections based partly on FPGA results and process extrapolation.
There is also the question of exactly how much human engineering was involved. The two engineers consisted of a CTO who guided the AI workflow and an engineer responsible for implementing firmware. Furthermore, Redwood is based on pre-existing architectures rather than being a completely new architecture created from nothing.
This means that the real test will come when AI is asked to design more complex hardware, identify subtle errors and account for the countless edge cases that can appear in a real chip.
And this is where Architect Labs is exploring something particularly interesting. The company is investigating a feedback loop where AI models running on the hardware can identify performance bottlenecks and feed this information back into the design process. The architecture could then be modified based on the problems discovered during actual operation, creating a cycle of model execution, bottleneck identification and hardware redesign.
If this can work reliably, chip development could become an iterative process rather than something where engineers spend years designing a processor before discovering problems after fabrication. Of course, there are still significant hurdles. FPGA performance does not directly translate to ASIC performance, and an error that makes it through AI-assisted verification could become extremely expensive once a chip has been manufactured.
Redwood therefore does not yet demonstrate that AI can replace established chip development platforms or ecosystems. The real test will be whether these AI-generated designs can repeatedly reach tape-out, manufacture successfully, meet their power and performance targets and support a viable software ecosystem.
What Does this Say About the Role of AI in Engineering?
Many believe that AI is still nowhere near ready for real commercial and practical engineering work, being more of a gimmick than a solution.
However, while this may be true for some areas of engineering, such as CAD, the increasing capabilities of AI are showing that its engineering capabilities are improving rapidly. Interestingly, chip design is an area where AI has a significant advantage.
Most modern chip design revolves around hardware description languages such as VHDL and Verilog. At their core, these are effectively language problems. Engineers describe what they want a piece of hardware to do using text, and that description can then be transformed into the logic required to create the final design. This makes chip design particularly suitable for AI.
An AI system can take a specification, generate RTL, modify that RTL when requirements change and then repeatedly test the result. It can also perform many of the repetitive tasks that would otherwise require engineers to spend hours writing and checking code.
This means that AI is more likely to replace people in the complex design of chips than it is in applications such as PCB design.
A PCB is fundamentally a physical system. The placement of components, routing of traces, thermal considerations and mechanical constraints all interact with one another, and many of these decisions require engineering intuition that is difficult to describe as a simple language problem. Chip design, by comparison, can be represented almost entirely through code and formal descriptions.
Right now, AI is still in its infancy, so its use in engineering remains limited. But this is no reason for engineers to refuse to learn and use these tools. In fact, this could be one of the most important periods for engineers to start experimenting with AI.
As these systems become increasingly capable, the role of the engineer could shift away from manually writing code, placing footprints and performing repetitive design tasks, and towards defining the architecture, requirements and overall function of a system.
The engineer would not necessarily disappear. Instead, the engineer could become the person directing the AI, checking its decisions and determining what the final system needs to accomplish.
And if that is where engineering is heading, then learning how to work alongside AI now could be considerably more valuable than waiting until the technology is already capable of doing the job.
You may also like
