Bitcoin ASIC Chip Evolution: From BM1385 to BM1373

Hold a Bitaxe Gamma in one hand and an old Antminer S7 in the other. Same algorithm. Same SHA-256. Same Bitcoin protocol from the same Satoshi whitepaper. But the silicon at the heart of each device tells a story of 11 years, a 20× efficiency improvement, and a fundamental redesign of how mining hardware is built. The S7’s BM1385 chip was state-of-the-art in 2015 at 200 J/TH. The Gamma’s BM1370 sits at 15 J/TH. And the BM1373 — Bitmain’s first 3nm SHA-256 chip, shipping in the Antminer S23 series — pushes that down to 10 J/TH per chip.

Key takeaways

• 11 years of silicon = a 20× efficiency gain (200 → 10 J/TH) and an 83× per-chip hashrate gain (30 GH/s → 2.5 TH/s) — same algorithm, same network.
• The BM1373 (3nm) is the current frontier: ~2.5 TH/s and ~10 J/TH per chip, powering the Antminer S23 series.
• The biggest single jump was architectural, not lithographic: the BM1368 (S21) redesigned voltage domains and dropped the PIC controller, roughly 6–7× per-chip hashrate over the prior generation.
• Architecture beats node: Auradine ships 3nm too, yet lands near ~16–17 J/TH — behind Bitmain’s 3nm BM1373, proving process node alone doesn’t win.
• For solo miners, the chip doesn’t change your odds per hash. A home Bitaxe has the same per-hash probability as an industrial BM1373 — only the hash count differs.

Each generation of ASIC silicon is a story: a process-node shrink, a voltage-domain redesign, a transistor count that doubles or triples, a thermal envelope reshaped to extract every joule of useful work. Most miners never look inside their hardware — the chips are black squares under heatsinks. But understand the silicon, and you understand the economics of solo mining: why some chains favor old chips, why a Bitaxe is genuinely different from an industrial farm, why the next halving will hurt some operators and not others. This is a complete chip atlas — every major Bitmain mining ASIC from 2015 to 2026, compared honestly with MicroBT and Auradine, ending with educated speculation about what comes after BM1373.

What an ASIC chip actually is (briefly)

A Bitcoin mining ASIC — Application-Specific Integrated Circuit — is a chip designed for exactly one purpose: computing the SHA-256 hash function as fast as possible while burning as little power as possible. Unlike a CPU or GPU, which is a generalist, an ASIC is a savant: it can do nothing else, but the one thing it does, it does roughly 100,000× faster per watt than a high-end GPU.

Inside the chip are many small SHA-256 compute cores running in parallel, each computing hashes every clock cycle. Modern Bitmain chips contain hundreds of thousands of these cores on a single die. Per-chip throughput is measured in terahashes per second (TH/s), and efficiency in joules per terahash (J/TH). Lower J/TH = more useful work per watt = lower power bill = competitive miner.

Two physical levers control everything:

1. Process node — how small the transistors are. Smaller = more transistors per square millimeter, lower switching voltage, less heat. The industry has gone from 28nm (BM1385) to 3nm (BM1373) in a decade.
2. Architecture — how the cores are arranged, how they communicate, how power is delivered. Smart architecture extracts more useful work from the same silicon area.

Both improve every generation. Bitmain has shipped nine-plus generations of mining chips since 2013, each making the previous obsolete in 18–24 months. This is why mining is hard: the hardware competes against itself.

The Bitmain chip family tree

Here’s every major Bitmain SHA-256 mining ASIC of the past decade, oldest to newest, with the device that made each chip famous:

Read that last row twice. From 30 GH/s in 2015 to 2,500 GH/s in 2026 — an 83× improvement in per-chip hashrate. From 200 J/TH to 10 J/TH — a 20× improvement in efficiency. The same algorithm. The same network. The same SHA-256 puzzle. Just better silicon, year after year.

Process nodes — what those numbers actually mean

”Process node” is shorthand for the manufacturing technology used to fabricate the chip. The number — 28nm, 7nm, 3nm — historically referred to the smallest feature size, though modern naming is more marketing than measurement. What matters: smaller numbers mean more transistors fit in the same area, and each switches at lower voltage with less leakage.

Each node shrink delivers roughly:

• 2× transistor density — twice as many compute cores in the same chip area
• ~30% lower power per operation — less heat for the same work
• ~15-25% higher clock speeds — more hashes per second per core

BM1385 (28nm, 2015): 30 GH/s, 200 J/TH
BM1373 (3nm, 2026): 2,500 GH/s, 10 J/TH
Improvement: 83× hashrate per chip, 20× efficiency, 11 years

For context: an Antminer S9 from 2017 needed 189 chips to deliver 14 TH/s. The Antminer S23 needs ~127 BM1373 chips to deliver 318 TH/s — 23× the hashrate from 67% the chip count, on a single device in a similar form factor. That’s what a decade of silicon evolution looks like in practice.

The chips, one by one

BM1385 (2015) — the patriarch

The first Bitmain chip widely deployed at scale, built on TSMC’s 28nm process. The Antminer S7 used 162 of these chips to deliver 4.7 TH/s at 1,293W — around 275 J/TH at the wall, ~200 J/TH at the chip level. By 2026 standards, the S7 produces less hashrate than a single Bitaxe Gamma chip. By 2015 standards, it was state of the art.

BM1387 (2017) — the legend

The chip that won Bitcoin mining for half a decade. The Antminer S9 used 189 BM1387 chips to deliver 14 TH/s at 1,372W (~98 J/TH). For years the S9 was the most-deployed Bitcoin miner on the planet — millions of units shipped. Even in 2026, some S9s still earn in regions with sub-$0.04/kWh power (and make excellent space heaters, as covered in our old-Antminer guide). No other Bitmain chip has matched that longevity.

BM1397 (2019) — the 7nm pivot

Bitmain’s first mainstream 7nm chip, used in the Antminer S17 series. It roughly halved J/TH from the BM1387 era and became the basis for the original Bitaxe MAX — the first DIY single-chip solo miner. The BM1397 used pre-calculated midstates rather than receiving full block headers, an architectural detail that distinguished it from later generations.

BM1366 (2022) — the 5nm jump

The first 5nm chip in Bitmain’s mining lineup, a massive efficiency leap to ~21 J/TH. Used in the Antminer S19 XP (140 TH/s, 21.5 J/TH) and the Bitaxe Ultra. The Bitaxe Ultra holds a special place in solo mining history — in March 2025, a single Bitaxe Ultra at ~0.48 TH/s solved Bitcoin block #887,212 for 3.125 BTC plus fees, the most-cited example of “lottery mining actually paying off” in the modern era.

BM1368 (2024) — the architecture redesign

This is where it gets interesting. The BM1368 was the first chip in a generation to make deep architectural changes — not just a process shrink. Two key moves:

• Voltage-domain redesign: the BM1368 moved from the traditional ~0.4V domain toward ~1.0-1.2V. That sounds backwards — higher voltage usually means more power — but coupled with the new architecture it allowed simpler power delivery, fewer regulators, and substantially higher per-chip hashrate.
• Eliminated the PIC controller: previous Bitmain chips relied on a separate PIC microcontroller to manage voltage scaling and chip communication. The BM1368 integrated these functions directly — simpler hashboards, fewer points of failure, easier firmware development.

The Antminer S21 used 108 BM1368 chips to deliver 200 TH/s at 17.5 J/TH; the Bitaxe Supra used a single BM1368 for 600-750 GH/s at ~22 J/TH on the desktop. The redesign delivered roughly 6-7× per-chip hashrate over the BM1366 generation — the biggest single-generation jump in Bitmain history.

BM1370 (2024-2025) — the refinement

The BM1370 took the BM1368 architecture and pushed it harder. Same 5nm process, refined for higher per-chip hashrate (~1.2 TH/s vs 0.7) and better efficiency (~15 J/TH vs 17.5). Used in:

• Antminer S21 Pro — 195 chips × 1.2 TH/s = 234 TH/s at 15 J/TH (~3,510W)
• Antminer S21 XP Hyd — 324 chips × 1.46 TH/s = 473 TH/s at 12 J/TH hydro (~5,676W)
• Bitaxe Gamma — 1 chip, 1.0-1.2 TH/s stock, up to ~1.84 TH/s overclocked
• NerdQAxe++ / Zyber 8G — 4 chips, 4.8+ TH/s
• NerdOCTAxe — 8 chips, 10-12 TH/s

The BM1370’s wide voltage window (0.65V to 1.30V) and frequency headroom (525 MHz stock, up to 900-1000 MHz overclocked on good silicon) made it a community favorite. Bitaxe overclocking guides sprouted everywhere; AxeOS firmware added tuning UIs. The chip became the bridge between industrial silicon and DIY desktop mining culture.

BM1373 (2026) — the 3nm future

Bitmain’s first 3nm SHA-256 chip, sold by TinyChipHub as the core of the Antminer S23 mainboard. Per-chip specs:

• ~2.5 TH/s per chip — roughly double the BM1370
• ~25W per chip — slightly higher than BM1370 (~17W stock)
• 10 J/TH efficiency — about 33% better than BM1370
• 3nm process — Bitmain’s first node shrink in four years

Deployed across the Antminer S23 series (announced at the 2025 World Digital Mining Summit, shipping into 2026):

The S23 Hyd 3U is genuinely remarkable: 1.16 PH/s in a single rack-mount unit, drawing 11kW on 380-415V three-phase power — more hashrate than four S21+ machines combined. Bitmain quotes a 7-year warranty, signaling confidence in the silicon’s longevity. (For the full S23 buyer’s breakdown — prices, ROI, solo odds — see our Antminer S23 guide.)

The Bitaxe and NerdQAxe community is already adapting boards for the BM1373: TinyChipHub ships sealed reels of the chip, and 4-chip NerdQAxe++ builds are projected at 10-12 TH/s — matching the Zyber 8G Solo Miner’s hashrate at significantly lower J/TH. The desktop solo-mining ceiling just moved up another order of magnitude.

The competition: MicroBT (Whatsminer)

MicroBT is Bitmain’s most serious SHA-256 competitor. It designs its own ASIC chips (not licensed from Bitmain) and has built a parallel evolution path:

MicroBT’s strategy has been steady refinement rather than dramatic architecture jumps. Its M60 series uses 5nm chips and competes directly with Bitmain’s S21 lineup — per-watt efficiency roughly 10-15% behind the BM1370 generation, close enough that Whatsminer stays popular where Bitmain availability is constrained (parts of Asia, Russia, certain African operations). MicroBT hasn’t yet announced a 3nm equivalent to the BM1373; analysts expect a Whatsminer M70-series in late 2026 or 2027 to close the gap. Until then, the BM1373 / S23 gives Bitmain a real efficiency lead at the top end.

The wildcard: Auradine

Auradine is a US-based (Silicon Valley) ASIC startup, founded in 2022 and backed by $80M+ in funding including from Marathon Digital. It announced the first Western-designed 3nm Bitcoin mining line — the Teraflux AT2880 — in November 2023, with shipments beginning in 2024, per TheEnergyMag. Notably, that put a 3nm Bitcoin ASIC in the field before Bitmain’s BM1373. Verified specs:

• Process: 3nm — the same node as BM1373
• AT2880 (air): up to 260 TH/s at roughly 16-17 J/TH in real-world testing (TheMinerMag measured ~17, above Auradine’s 15 J/TH marketing claim)
• Immersion (AI3680) and hydro (AH3880) variants: ~375-600 TH/s at ~14-14.5 J/TH
• Made-in-USA narrative: appealing to North American institutional buyers wary of supply-chain politics

Here’s the instructive part: despite sharing the 3nm node, Auradine’s air model lands near 16-17 J/TH — comparable to Bitmain’s 5nm S21/BM1370 generation, and well behind the 3nm BM1373’s 10 J/TH. Same process node, very different result — proof that architecture and power design matter as much as lithography. Auradine isn’t a high-volume player yet, but in November 2025 it announced a next-generation Teraflux targeting 9.8 J/TH; if delivered, that would be among the most efficient Bitcoin hardware ever shipped. If geopolitical pressure on Chinese chip exports intensifies, Auradine could grow fast — the open question is manufacturing scale.

Architecture deep dive: BM1368 → BM1370

For miners who actually open their hardware, the BM1368→BM1370 transition is the most interesting recent engineering change. Both use the same 5nm node, same logical architecture, same SHA-256 cores — yet the BM1370 delivers ~70% more hashrate per chip at similar power. How? Three things:

1. More cores per die — a refined 5nm cell library allowed denser core placement, roughly 1.5× the core count on similar die area.
2. Optimized power delivery — the BM1370’s wider voltage window (0.65V to 1.30V) lets it scale dynamically between low-power steady-state and high-power burst modes; the BM1368’s window was narrower.
3. Better thermal coupling — package improvements (solder-ball pitch, thermal interface) allowed sustained higher clocks without throttling.

For Bitaxe overclockers, this is why BM1370 chips can be pushed to 900+ MHz on stock voltage — frequencies that would cook a BM1368. It’s not magic; it’s metallurgy and packaging. Same silicon, smarter delivery.

What BM1373’s 3nm jump actually delivers

The BM1373 is the first Bitmain chip to leave the 5nm node. The leap to 3nm produces:

• ~33% efficiency improvement at the chip level (15 → 10 J/TH)
• ~2× per-chip hashrate (1.2 → 2.5 TH/s)
• ~50% reduction in chip count for equivalent device hashrate
• Lower thermal density — even at higher per-chip power, a smaller die makes heat easier to extract

The S23 Hyd at 580 TH/s, 9.5 J/TH represents what 3nm SHA-256 silicon can currently do. For comparison, the previous-gen S21 XP Hyd needed 12 J/TH for 473 TH/s in the same hydro envelope. The 3nm node delivered 22% more hashrate at 21% better efficiency — both axes at once. A genuine generational leap, not a marketing refresh. What it means for solo miners:

1. Older hardware (S19, M30 series) is approaching obsolescence for anyone paying retail power — the efficiency gap is now too wide. Expect significant fleet retirements through late 2026.
2. The desktop solo-mining ceiling moves up. Single-chip BM1373 builds will deliver 2.5 TH/s on a desk; 4-chip NerdQAxe-class builds 10+ TH/s. That’s the new floor for “consumer-scale” SHA-256 silicon.

Chip generation and ROI in mid-2026

For a miner running a fleet, the question isn’t “is this chip cool?” — it’s “does it pay for itself before the next generation makes it obsolete?” Rough margins at $0.07/kWh, a BTC price near $61k, and a hashprice around $29/PH/day (mid-2026), per Hashrate Index data:

Numbers are illustrative — real margins move with hashprice, difficulty, BTC price, and uptime — but the directional message is stark. At mid-2026’s brutal hashprice, S19 hardware is underwater and S21-class gear is thin-to-breakeven at seven-cent power; only the newest efficient silicon clears a healthy margin. The S23 resets the table, and operators who wait too long to upgrade get squeezed by every difficulty increase. (Why margins got this thin is the subject of our halving aftermath analysis.)

Implications for solo mining

1. Bitaxe-class hardware is the most viable it has ever been

A BM1373-based single-chip miner at 2.5 TH/s changes the math. On BC2 / BCH2 chains where a single BM1370 Bitaxe finds blocks every day or two, a BM1373 unit finds them in hours; on XEC, expected time on a single chip roughly halves. Consumer-scale solo mining is genuinely returning to viability, not just lottery mode.

2. Small S21+ setups remain the BCH sweet spot — for now

A four-machine S21+ setup (~940 TH/s) averages a BCH block on the order of once every few weeks at mid-2026 difficulty. That math holds across chip generations until BCH network hashrate rises significantly — which would require S21+/S23 deployments at scale on BCH specifically, not currently happening. For 2026-2027, a small S21+ cluster remains a cost-effective entry point for individual BCH solo miners. (Run your own numbers in the profitability tool.)

3. The competitive pressure is real

Bitcoin’s network hashrate rises as S23 deployments scale, forcing operators on older hardware to upgrade, find subsidized power, or shut down. Solo miners on smaller chains (BCH, BC2, BCH2, XEC) are insulated, because those networks aren’t seeing aggressive S23 deployment. The smaller chains remain solo mining’s structurally protected niche.

What comes after BM1373

What’s likely for 2027-2028:

• 2nm silicon — Bitmain’s next node shrink. Expect roughly 5-7 J/TH at the chip level, deployable around late 2027.
• Vertical / 3D stacking — borrowed from memory, stacked dies could deliver 2-3× per-chip hashrate without further node shrinks.
• Power-delivery innovations — direct-on-chip DC-DC conversion, integrated cooling channels, more aggressive voltage scaling. Architecture wins now matter as much as process wins.
• End of “trivial” gains — shrinks below 2nm get prohibitively expensive, so future efficiency increasingly comes from architecture rather than lithography. Innovation slows but doesn’t stop — Auradine’s 9.8 J/TH target on 3nm is exactly this kind of architecture-led gain.

The bottom line

A decade of Bitcoin mining silicon has produced a 20× efficiency improvement and an 83× per-chip hashrate increase — smaller, faster, cheaper per unit of work. The same algorithm, the same network, the same Satoshi whitepaper; just better silicon, year after year.

For solo miners, that’s good news and bad news. Bad: the network gets harder every year, and small operators must upgrade or accept smaller relative shares. Good: your Bitaxe Gamma at home has the same per-hash probability as a BM1373 in an industrial farm. The chip doesn’t know it’s small; the network doesn’t care. Probability is uniform across hash count, regardless of who computed it.

The BM1373 is where Bitcoin mining sits in 2026: 3nm silicon, 10 J/TH, 2.5 TH/s per chip, in machines ranging from desktop NerdQAxe builds to 11kW rack-mount 3U behemoths. Eleven years from BM1385’s 200 J/TH to BM1373’s 10 J/TH — and the next eleven will probably bring another 5-10× gain. The chips keep shrinking, the network keeps adjusting, and the math keeps working. Pick your silicon, pick your chain, plug in, and wait — the dice are still rolling.