When we set out with the goal to bring a new generation of Bitcoin miner to market, we wanted to push the envelope and innovate in a way that would stimulate further competition. Making BTC miners however, is surprisingly difficult because they they need to achieve extreme conditions. We chose the route of developing a 10nm chip because it was the newest technology available at the time with promising advantages. However, after 14nm the physics change and introduce harder engineering challenges due to the small scale. One of the common misconceptions is that the smaller the nm size, the greater efficiency one will gain. In reality, gains are tempered by losses in other areas due to the new physics.
The DragonMint T1558 chip is the first-ever 10nm Bitcoin mining chip. The tapeout was in April 2017 and first chips came back between late August and early September 2017, various manufacturing and engineering challenges were worked through as first prototype miners were assembled. However, the wafer yield was lower than expected, and in order to compensate some engineering changes were made to the wafer to control some of the advanced process issues, at the slight expense of power efficiency which we believed we could compensate for in other parts of the miner. The efficiencies gained from using AsicBoost also mitigate some of the losses incurred by the new 10nm process (specifically, electrons are more difficult to control which causes efficiency loss, but AsicBoost allows us to use less power, so it cancels out much of the loss).
Although we still have more work to do with refining the process and increasing the yield (which is expected when working with new processes nodes) as 10nm process matures, in the end, the T1558 chip is now a leader in the market.
Manufacturing the miner has also brought other challenges, which engineers had to spend a lot of time, analysis and iteration in order to find and solve the various bottlenecks. All silicon wafers have some variance, and chips vary across the wafer. We found that once we put 10nm chips on the hash card, they are more prone to variance than expected. Due to the serial design of the PCB, chips are wired together like Christmas lights. It turns out our the chips have a much narrower tolerance to variance so we have to do more specialized sorting so that each board has similar “color”. The chips in the earlier miners were not so well sorted according to this precision and the miner manufacturing process was less mature.
Due to these uncertainties, manufacturing had to proceed cautiously to ensure we ironed out any remaining engineering issues. This is why there is a bit more variance on the early miners, and the hashrate is a little lower than expected. We’re still well within the 8% tolerance with the DragonMint T1 miner, but of course we want to be closer to the advertised 16TH/s.
On the software side, we wrote a completely new firmware based on “Buildroot” and a brand new web interface with rich API. We are also very lucky to have some industry specialists working on the mining like Linux Kernel developer, Con Kolivas, the original author and maintainer of cgminer. New software takes time to mature and the software has been an essential part of tuning performance and efficiency of the miner. Work is still ongoing, but as early adopters of the DragonMint T1 have discovered, the newer firmware released after the first batch of miners shipped in March, gives much better hashrate and performance on the same hardware.
Overall, the manufacturing process is only just becoming mature enough to mass produce with confidence and thus we’re beginning to ramp up production of the T1 seriously as well as re-verifying the first run stock of DragonMint miners.