Building a Better ASIC Company

Zach Herbert
Obelisk Blog
Published in
7 min readJan 31, 2019

How we continuously improve Obelisk’s processes to deliver better quality cryptocurrency mining hardware in less time.

When we launched Obelisk in summer 2017, we were completely new to hardware — none of the team had ever mass-produced a hardware product. Today we’ve shipped over 7,000 ASIC miners for Siacoin and Decred, with another 3,500 shipping over the next few weeks. We also launched a sale for our next-generation miner, the Obelisk GRN1 for Grin.

Our team is very proud of our progress. 18 months after launch, we have developed a repeatable process for designing, manufacturing, and shipping thousands of miners to a worldwide customer base. We are now confident that, once we release a chip for production (“tape-out”), we can be shipping miners to customers in less than 6 months.

We have focused on four key areas to improve our processes and reduce timelines:

  1. Establishing direct relationships with chip foundries.
  2. Building mechanical, electrical, and software reference designs — and iterating on these designs.
  3. Trialing manufacturers and building strong relationships with the best.
  4. Understanding common manufacturing delays and updating our designs accordingly.

This post will detail how we focused on each of these areas to build Obelisk into a high-quality designer and manufacturer of ASIC mining hardware.

Foundry Relationships

When we were working on our first generation of miners, we hired an outside company to design ASIC chips that could mine Siacoin and Decred. In addition to design, this outside team was also responsible for selecting the foundry and managing the foundry relationship. This means that Obelisk had no interaction with the foundry that was producing our chips, and we trusted the ASIC design house to effectively manage the relationship as an intermediary.

TSMC Facility

Our design house selected a foundry in China to produce our chips. As we grew close to tape-out, the foundry dropped us without warning. Because we had no direct interaction or relationship with the foundry, we were forced to rely on the design house to resolve the problem. Ultimately they made the decision to switch foundries, causing a delay of multiple months.

Additionally, Obelisk was unable to influence production schedules or negotiate pricing. We later learned that the design house was using an intermediary instead of directly communicating with the new foundry. This ultimately caused significant delays, both due to capacity constraints and ordering errors. All this would have been easily avoidable if we were involved in the foundry relationship.

Now, Obelisk is involved directly in the foundry relationship. We have met with senior leadership at multiple foundries and have a strong understanding of production timelines and pricing across the various process nodes. While we still use outside ASIC design houses, we are directly involved in foundry conversations and presented with clear timelines and pricing.

For the GRN1 miner, we are confident in the timelines being communicated by the foundry, and confident that we will avoid needless delays from capacity constraints and ordering errors. Moreover, we are working with a new design house that has very close relationships with the major foundries.

Reference Designs

When we set out to design our first generation miners, we had no previous designs to reference — we had to design a hashing board, control board, enclosure, and firmware from scratch. We looked at competitor designs for inspiration, but designing the first generation of hardware simply takes time.

Now we no longer have to start from scratch. Obelisk has designed and manufactured two generations of hardware, and we are currently designing our third generation of hardware for the GRN1. Moreover, we realized significant improvements in performance, quality, time, and cost between first and second generation hardware.

Hashing Boards
Our Gen 1 hashing board design was highly conservative. Completed in 6 months, it contains 15 ASICs per board and draws slightly over 200 W.

Hashing Board comparison, Gen 1 vs Gen 2

By contrast, our Gen 2 hashing board was engineered for high performance. Completed in only 3 months at 25% the design cost of the Gen 1 board, it packs 32 ASICs drawing over 400 watts of power. And, on a per-chip basis, it is 40% less expensive to produce.

On our Gen 3 hashing board for the GRN1, we will continue to improve performance and cost while reducing design timelines. We are no longer starting from scratch — we are building off our existing high-performance designs.

Control Board
We’ve made similar improvements to our firmware and control boards across generations. Our Gen 2 control board is 33% less expensive to produce, has a significantly smaller footprint, and relies on easily-procurable parts instead of expensive SOMs. This control board is a fantastic reference design that we will continue to build upon. For our GRN1 units, we expect to once again significantly reduce costs and complexity.

Firmware
Firmware challenges were one of the major sources of delays during Gen 1 production. Now, with two generations of firmware under our belts, we’ve significantly mitigated these challenges.

As we design and manufacture new products, like the GRN1, we only need to make modifications to our existing firmware. And as we make improvements, we can carry those forward to each new generation.

Additionally, Obelisk has adopted industry-leading methodologies for firmware development. For the GRN1, we will be simulating a string of ASICs using an FPGA board. This will allow us to write firmware for the on-board microprocessor before we have even received test chips from the foundry. Therefore, when chips arrive, we only have to test and refine the firmware. The hard work will already be done.

Enclosure
We are proud of our Gen 1 enclosure, but our Gen 2 enclosure is better in every way. Our team reduced costs by 30%, shrunk horizontal footprint by 50%, and created a single-screw design to make it incredibly easy to access the hashing board, control board, and fans. We implemented research by Northshore and stacked 60mm fans, which provides more uniform airflow.

Obelisk Gen 2 enclosure with single top-screw

Obelisk is now capable of designing and prototyping new enclosures, such as the SC1 Immersion enclosure, in just 1-2 months.

For the GRN1, we will continue to build off our reference designs and reduce cost and footprint. We are also exploring an integrated power supply.

Manufacturers

Obelisk originally planned to build our first generation hardware in China. We selected a US-based contract manufacturer (CM) that owned a facility in China. However, right when we were supposed to begin purchasing parts, the CM dropped us without warning, and we had to scramble to rebuild our supply chain.

We ultimately split production across three US-based CMs, so that components could be built and assembled in parallel. In addition to moving rapidly, this allowed us to trial each CM, assess quality and turnaround times, and decide whether to move forward with future business.

Today, we use one of those CMs as our primary manufacturer and another as a backup. Because we’ve established strong relationships with multiple CMs in the USA, we are able to quickly launch new projects and source components without worrying about relationship and quality issues.

We plan to build the GRN1 at our primary, tried-and-tested CM, and we will continue to expand our supply chain to handle growth.

Manufacturing Delays

After manufacturing over 5,000 Gen 1 units, we understood the largest sources of delays and were able to update Gen 2 designs for easier manufacturing. This process is called DFM, or design for manufacturability.

Our largest delay was related to heatsinks. At the time, most ASIC miners utilized individual heatsinks on each chip. We decided to emulate that approach and hired a reputable heatsink company to design a solution consisting of a plastic base, a metal spring clip, and a heatsink. The plastic base was supposed to easily adhere to the hashing board using 3M double-sided tape, and had been successfully utilized by the heatsink company in other products.

Unfortunately, tolerancing issues and production imperfections forced us to remove the tape from every single plastic base — 150,000 of them — and glue them down instead with adhesive. This caused a major, unavoidable slowdown in production. And, once shipped to customers, heatsink failure rates were high.

Gen 2 hashing board with single heatsink

To fix this, we designed our Gen 2 miners to use a single heatsink that spans across all chips. It uses a thermal gap pad for electrical isolation and mounts directly to the hashing board with screws. This makes manufacturing significantly easier. We’ve seen Whatsminer also move to this single-heatsink design.

This example demonstrates how Obelisk has improved our designs to reduce manufacturing timelines and costs. We’ve taken many other steps, such as eliminating the need for testing fixtures by pre-programming control boards and hashing boards to self-test upon startup. As we design the GRN1, we will continue to make these DFM improvements.

Over the last 18 months, Obelisk has significantly improved our operations to bring ASIC miners to market faster — with far less risk. By pursuing the above focuses, as well as countless other small improvements, we have put our company on a path to repeatedly deliver miners within 6 months of tape-out. We will continue to improve our processes to further reduce timelines and risks. Obelisk aims to be the leading ASIC miner manufacturer. Thank you for joining us on this journey!

Zach Herbert runs operations for Obelisk. Learn more about our product offerings at obelisk.tech.

--

--