Antminer S9

The Antminer S9 is Bitmain’s 2016 air-cooled SHA-256 workhorse: roughly 14 TH/s at about 1,350 W from the wall, or close to 96 J/TH. It runs 189 BM1387 16 nm chips across three hashboards. Long discontinued, it survives today as a cheap heater, learning rig, and solo-lottery machine.

Chip and hashboard architecture

The S9 is built around the BM1387, Bitmain’s first-generation TSMC 16 nm SHA-256 ASIC. Each chip carries 114 hashing cores and runs a hardware difficulty of 256, producing roughly 75 GH/s on its own. The machine ships with three identical hashboards, and each board hosts 63 BM1387 chips — 189 chips in total, which is where the ~14 TH/s aggregate figure comes from.

Per hashboard, those 63 chips are organized into 21 voltage domains of 3 chips each. The domains are wired in series to form a single regulated string of roughly 8.4 V, with each domain sitting near 0.40 V. This is the detail most spec sheets get wrong: voltage on the S9 is managed per domain and per board, never per individual chip. A single PIC16F1704 microcontroller per board sets that string voltage through an 8-bit DAC over the FPGA’s I2C bus, addressed at 0x55, 0x56 and 0x57 for the three boards. Its usable window is about 7.94–9.44 V, with the stock firmware defaulting near 8.6 V. The PIC also runs a watchdog (about one minute on stock firmware, tightened to roughly ten seconds under open-source firmware) that cuts power if the control board stops feeding heartbeats.

The brain of the machine is a Xilinx Zynq 7010 (XC7Z010) control board — dual ARM Cortex-A9 cores at 667 MHz alongside a 28K-logic Artix-7 FPGA fabric running the s9io core. The FPGA handles the UART work-distribution FIFOs to the chips and the single I2C bus that reaches the PICs and the TMP75 board-temperature sensors. Those temperature sensors only appear on the bus once the PIC has enabled domain voltage, a quirk worth knowing during diagnostics. Firmware and configuration live on a 256 MB Micron SLC NAND, booted through the Zynq’s first-stage loader and BOOT.bin.

Real-world power and efficiency

The nameplate 1,350 W and ~96 J/TH are wall-plug figures at stock settings; in practice an aging S9 on a worn PSU often measures closer to 98 J/TH once fans, the control board and conversion losses are counted. Efficiency tracks frequency and voltage directly. The BM1387’s PLL exposes 33 discrete clock steps from 100 MHz up to 900 MHz, with 650 MHz as the stock default, so there is real room to tune in either direction.

Two strategies dominate. Underclocking — dropping toward 500–550 MHz and trimming the string voltage — sheds heat and noise and improves J/TH, which is ideal when the S9 is doing double duty as a space heater. Overclocking past 650 MHz buys hashrate but pushes efficiency the wrong way and stresses 2016-era power circuitry, so it is rarely worth it on hardware this old. Because all three boards share their core voltage as a single series string, tuning is coarse and per-board rather than surgical; you adjust the whole string, and the autotuner (when present) calculates a working voltage at runtime rather than reading a preset table. For per-model frequency and voltage targets, see our ASIC power profiles database.

Firmware compatibility

Stock S9 firmware is Bitmain’s cgminer build, which boots the FPGA fans at a fixed ~50% duty and offers only basic frequency control. The S9’s longevity, though, comes from one of the richest third-party firmware ecosystems in mining. The open-source autotuning firmware that pioneered runtime-calculated voltage and frequency — and that remains the only major Antminer firmware to natively speak Stratum V2 — was first proven on this exact BM1387 platform; we credit that work as the foundation everything since has built on. Other community firmwares add direct frequency and voltage entry (the S9 predates named preset profiles, so values are set numerically), letting operators dial in their own efficiency curve.

Common faults and troubleshooting

After most of a decade in service, S9 problems are predictable. The most common is a hashboard not detected or a low ASIC count (for example “0/63” or “57/63” chips found). Because each domain’s three chips sit in a series string, a single failed chip can break continuity for its whole domain, and a dead domain can drop the entire board offline. A hard short inside a domain is worse: it pulls that section’s rail toward zero, and the boost circuit either current-limits or shuts down, taking the whole board with it.

• High hardware-error rate / low hashrate: usually degraded chips, cold or cracked solder joints on the hashboard, or an undervolted string — often improved by a small voltage bump within the safe window.
• Temperature sensor errors: remember the TMP75 sensors only appear once the PIC enables domain voltage, so a “temp sensor lost” fault frequently points back to a PIC or power-rail problem, not the sensor itself.
• Fan and PSU faults: tired APW3-series PSUs and worn fans cause shutdowns and throttling that masquerade as hashboard failures.
• Intermittent chains: dust, heat cycling and aged capacitors on 2016-vintage boards lead to chains that come and go with temperature.

Walk these symptoms through our ASIC fault finder to isolate whether the issue is a board, a domain, the PIC, the PSU or the control board before you reach for a soldering iron.

Repair and longevity

The S9 is one of the most repairable machines ever made, which is exactly why so many are still hashing. The faults above are board-level and fixable: domains can be diagnosed down to the offending chip, dead BM1387s can be reballed or replaced, PICs can be reflashed, and power stages can be rebuilt. D-Central has repaired this platform in-house since 2016 and knows its failure modes intimately — from the I2C quirks to the series-string behaviour that confuses generic testers. If your S9 is showing missing chips, dead boards or temperature faults, our ASIC repair service can usually bring it back for a fraction of replacement cost. Keeping a working S9 alive is also the more sovereign, less wasteful choice than landfilling it.

Who the S9 is for

At ~96 J/TH the S9 is firmly in the legacy efficiency tier, so it is no longer a serious profit machine against today’s network difficulty. Where it still shines is as a heat-recovery space heater (it dumps roughly 4,600 BTU/h that you would otherwise pay a furnace to produce), a solo-mining lottery ticket, and above all a learning platform — cheap, abundant, endlessly documented and forgiving to experiment on. If you want to learn ASIC tuning, firmware flashing and board repair without risking a $5,000 flagship, the S9 is the canonical teaching tool. For an even smaller, fully open-source SHA-256 learning device, our power-profile data and Bitaxe-class builds are the modern entry point; for production hashing, step up to a current-generation machine.

Generational context

The S9 defined the first wave of efficient ASIC mining. Its BM1387 was Bitmain’s 16 nm starting point; the same architecture later moved to 7 nm (the S17/T17 era) and then to today’s 5 nm silicon, with cores-per-chip and domain density climbing at every step. Current 5 nm flagships operate in the mid-teens of joules per terahash, roughly five to six times more efficient than the S9 — but they are also far harder to repair and far less forgiving to tinker with. That trade-off is the S9’s enduring appeal: it is slow and thirsty by modern standards, yet it remains the most accessible, most hackable, most teachable Bitcoin miner ever shipped, and a fixture of the open-source mining movement it helped start.