EEVblog 1535 - DeepPCB AI AutoRouting FAIL!

EEVblog
28 Mar 202326:16

TLDRIn this video, the host revisits a previous challenge comparing his PCB routing skills against an auto-router, this time testing DeepPCB's AI AutoRouter. Despite its promises of advanced AI and machine learning capabilities, the AI routing fails to outperform traditional methods, resulting in a messy and inefficient layout. The host expresses disappointment, as the AI does not seem to learn from its mistakes or prioritize routing effectively, ultimately suggesting that traditional auto-routers might still be the better choice for PCB design.

Takeaways

  • 😀 The video revisits a previous comparison between human PCB routing skills and an auto router, highlighting the presenter's superior manual routing over the auto router in a past project.
  • 🔍 The presenter, a former professional PCB design engineer, explains that auto routers are useful but require proper setup and constraints to work effectively.
  • 🤖 The video introduces DeepPCB, a company offering an AI-powered auto router, which is tested against traditional methods to see if it can improve the routing process.
  • 🚀 The DeepPCB AI router is described as fully automated and cloud-native, capable of handling complex boards, but is still in beta and only supports certain file formats like DSN.
  • 🔧 The presenter emphasizes the importance of component placement in PCB layout, noting that the AI router lacks the understanding of circuit schematics and priorities that a human designer would consider.
  • 🔄 The AI router's process involves 'rip up and retry' algorithms, attempting to find the best routing solution, but the presenter observes this process and finds it lacking.
  • 📉 The video shows the AI router failing to complete some routes and making poor routing decisions, such as jagged traces and inefficient path choices.
  • 🛑 The AI router's process eventually freezes, and the presenter contacts support for assistance, reflecting a less-than-ideal experience with the technology.
  • 📊 After 24 hours, the AI router claims to have succeeded, but the results are still unsatisfactory, with many connections unrouted and poor routing choices persisting.
  • 📚 The presenter compares the AI's routing with his own manual routing, demonstrating a clear preference for the human touch and the ability to prioritize and group components effectively.
  • 👎 The video concludes with a negative assessment of the AI router, suggesting that traditional auto routers with proper setup may be more reliable and effective.

Q & A

  • What was the main topic of EEVblog 1535 video?

    -The main topic of the video was to evaluate the performance of DeepPCB AI AutoRouter against manual routing and traditional auto routing methods on a double-sided PCB board design.

  • What was the presenter's previous experience with auto routers?

    -The presenter is a former professional PCB design engineer with decades of experience in routing PCBs, and he has previously compared his manual routing skills with Altium's auto router in a multi-part video series.

  • Why did the presenter believe auto routing is useful despite its limitations?

    -The presenter believes auto routing is useful because it can produce good results for complex PCBs like motherboards when set up with the right constraints and parameters, although it requires more time to configure properly.

  • What was the presenter's initial expectation of DeepPCB AI AutoRouter's performance?

    -The presenter expected the AI AutoRouter to perform better than traditional auto routers due to its machine learning capabilities and the hype around AI technology.

  • What was the result of the presenter's test with DeepPCB AI AutoRouter?

    -The test showed that DeepPCB AI AutoRouter did not perform better than traditional auto routers. It failed to route some connections and produced a messy layout with poor prioritization and routing decisions.

  • What was the presenter's opinion on the AI aspect of DeepPCB's auto router?

    -The presenter was disappointed with the AI aspect, as it did not seem to learn from its routing attempts and did not produce a result that was significantly different or better than a traditional auto router.

  • What was the presenter's suggestion for using auto routers effectively?

    -The presenter suggested using auto routers for less critical parts of the PCB layout, while manually routing the important and high-priority traces to ensure a clean and efficient design.

  • What was the issue with the AI AutoRouter's routing of high voltage lines?

    -The AI AutoRouter did not prioritize the high voltage lines, resulting in poor routing decisions that did not maintain necessary clearances and did not follow the best practices for high voltage traces.

  • What was the presenter's final verdict on DeepPCB AI AutoRouter after the test?

    -The presenter concluded that DeepPCB AI AutoRouter was a failure and did not recommend its use, suggesting that traditional auto routers with proper setup might yield better results.

  • What was the presenter's advice for PCB layout designers regarding auto routers?

    -The presenter advised PCB layout designers to take pride in their work, manually route critical parts of the board, and use auto routers only for less critical areas with specific constraints.

Outlines

00:00

😀 PCB Routing Skills vs. Auto Router

The speaker revisits a previous video where they compared their manual PCB routing skills with Altium's auto router on a Nixie tube project. They highlight that while auto routers can be useful, especially for complex PCBs, they require proper setup and constraints to be effective. The speaker mentions their manual routing was superior in that instance. They also introduce a new AI auto router from Deep PCB, which they plan to test using the same project file for a fair comparison.

05:00

🤖 Testing Deep PCB's AI Auto Router

The speaker discusses the concept of AI in auto routing and mentions that while it's a current trend, similar technologies have been around for decades. They express skepticism but decide to test Deep PCB's AI auto router, which claims to offer fully automated routing without human intervention. The speaker notes the limitations of the tool, such as only supporting DOT DSN files, and the fact that their Altium license has expired, preventing them from exporting their original project. They proceed with a DSN file provided by others and observe the AI router's progress, noting that it's a live update, which is impressive.

10:01

🔍 AI Routing's Live Updates and Limitations

The speaker continues to observe the AI routing process, noting that it's live and dynamic, but also pointing out that it doesn't seem to understand certain PCB routing principles, such as prioritizing high voltage lines and maintaining clearance. They express concern about the AI's lack of understanding of the circuit's schematic and its inability to apply manual routing constraints. The AI router's process includes 'rip up and retry' algorithms, which the speaker explains, but they are ultimately disappointed with the AI's routing decisions, which they compare unfavorably to their own manual routing.

15:02

😕 AI Auto Router's Failure to Complete the Task

The speaker reports that the AI auto router has failed to complete the routing task, having frozen after a certain point. They express disappointment with the results, noting that the AI did not follow proper PCB routing practices and left many connections unrouted. Despite the AI's attempts at finding solutions, it failed to progress beyond a certain point, and the speaker emails Deep PCB support for an explanation.

20:03

🚫 AI Routing's Inability to Match Manual Skills

After 24 hours, the speaker reviews the AI's routing results, which are still incomplete and of poor quality. They compare the AI's work to their own manual routing and criticize the AI for not understanding basic PCB routing principles, such as layer prioritization and trace organization. The speaker emphasizes the importance of manual routing techniques and the inability of the AI to replicate a human designer's understanding and strategy.

25:04

👎 Disappointment with Deep PCB's AI Auto Router

The speaker concludes their experiment with Deep PCB's AI auto router by expressing strong disappointment. They find no advantage to using the AI over traditional auto routers and criticize the AI for its lack of learning and poor routing decisions. The speaker advises against using the AI for PCB routing and suggests that taking pride in one's work and manually laying out the PCB is a better approach.

📢 Final Thoughts and Availability

In the conclusion, the speaker invites viewers to engage with the content by liking the video, discussing it in the comments, and on the EV blog forum. They also mention their presence on various alternative channels and platforms, encouraging viewers to follow them for more content.

Mindmap

Keywords

💡PCB

PCB stands for Printed Circuit Board, which is the fundamental component used in electronics to mechanically support and electrically connect electronic components using conductive tracks, pads, and other features etched from copper sheets laminated onto a non-conductive substrate. In the video, the host discusses the process of routing or designing the conductive paths on a PCB, comparing manual routing skills with an AI-powered auto-routing system.

💡Auto router

An auto router is a feature in PCB design software that automatically creates the conductive pathways, or traces, between electronic components on a PCB based on predefined rules and constraints. The video's theme revolves around comparing the efficiency and quality of auto routing with manual routing, particularly highlighting the limitations of the AI auto router from Deep PCB.

💡AI Auto router

The AI Auto router refers to an automatic PCB routing system that employs artificial intelligence algorithms to improve the efficiency and effectiveness of the routing process. The script discusses a specific AI auto router from a company called Deep PCB, which is purported to use advanced AI technology for PCB routing, but the host finds it lacking compared to traditional methods.

💡Nixie tube project

The Nixie tube project mentioned in the script is an electronics design challenge that the host had previously tackled, which involved creating a circuit using Nixie tubes—electronic display devices that were popular before the advent of LCD and LED displays. The project serves as a basis for the comparison between manual and auto-routing techniques.

💡Placement

In PCB design, placement refers to the strategic positioning of electronic components on the board to optimize routing, minimize signal paths, and ensure the circuit functions correctly. The script emphasizes that proper placement is crucial to successful PCB design, often accounting for 90% of the layout's success.

💡Constraints

Constraints in PCB design are rules and limitations set to guide the auto router in creating an efficient and functional circuit layout. They can include trace width, spacing, and other design parameters. The video script describes how setting up constraints is a time-consuming process that is essential for the auto router to function effectively.

💡DRC checked

DRC stands for Design Rule Check, which is a process used in PCB design to verify that the layout adheres to the specified design rules, ensuring manufacturability and functionality. The script mentions that the AI auto router's results are DRC checked, indicating an automated validation of the design's adherence to these rules.

💡Reinforcement learning

Reinforcement learning is a type of machine learning where an agent learns to make decisions by performing actions in an environment to maximize some notion of cumulative reward. The script mentions reinforcement learning as part of the AI technology used by Deep PCB's auto router, suggesting that the system learns from its routing attempts to improve over time.

💡Rip up and retry

Rip up and retry is an industry term for a process where an auto router will remove and reroute certain paths if it encounters a routing problem or 'dead end'. The script describes the AI auto router's use of this technique, indicating that it attempts to find better routing solutions by undoing and redoing parts of the layout.

💡Cloud infrastructure

Cloud infrastructure refers to the resources and services provided over the internet, allowing for scalable computing power and storage. The script mentions that Deep PCB's AI auto router leverages cloud infrastructure, suggesting that the routing process is performed on remote servers, potentially offering faster processing times and more computational power.

💡Solution

In the context of the script, a solution refers to a specific routing outcome or attempt generated by the AI auto router. The video shows the process of the AI generating multiple solutions, with the host evaluating the quality of each solution and ultimately finding them unsatisfactory.

Highlights

EEVblog revisits the challenge of PCB routing, comparing human skills with AI AutoRouting.

In a previous video, Dave Jones demonstrated his routing skills against Altium's AutoRouter.

AutoRouters require proper setup and constraints to function effectively, contrary to common misconceptions.

AI AutoRouting by DeepPCB is tested against traditional methods in a Nixie tube project.

DeepPCB's AI AutoRouter is in beta and claims to utilize cloud-based machine learning for PCB routing.

The AI AutoRouter is expected to handle complex boards and learn from constraints and schematics.

Dave Jones expresses skepticism about the AI's ability to understand schematic relationships without explicit instructions.

The AI AutoRouter fails to produce a clean or efficient routing, similar to traditional auto routers.

Live updates during the routing process show the AI making poor routing decisions.

The AI struggles with high voltage trace routing, lacking the understanding of human designers.

Dave Jones manually routes the board as a comparison, showcasing a cleaner and more organized result.

The AI AutoRouter's final result is deemed a failure, with many connections unrouted and poor trace management.

Dave questions the AI's ability to learn from its routing attempts, as it does not show improvement.

The video concludes that traditional auto routers may still be superior for certain tasks.

AI AutoRouting is criticized for not living up to its hype and failing to outperform human routing in this test.

The video ends with a recommendation to take pride in manual PCB layout for optimal results.