Building a Home GPU Rig for Crypto Password Recovery

Every day someone with a locked wallet.dat or MetaMask vault asks the same question: should I just buy a couple of 3090s and crack it myself? The answer is "sometimes, but probably not", and the difference comes down to the exact KDF your wallet uses, the entropy you suspect your password carries, and the opportunity cost of the hardware. This page lays out the arithmetic without the usual reseller hype.

Step one: understand your target mode

Different wallets use radically different key derivation functions, and the speed you can expect from a GPU varies by roughly four orders of magnitude across them. Before you buy anything, identify your wallet's hashcat mode and its realistic rate on a single RTX 3090 at stock settings.

These numbers are with default workload profile -w 3 on Linux with driver 550.120 and hashcat 6.2.6. Note the extreme spread: Ethereum Scrypt at 900 H/s is about five thousand times slower than Electrum's mode 16600. If your wallet sits on the slow end of that table, a home rig with one GPU will take months where the same password on Electrum would finish in hours.

Step two: pick the right GPU

RTX 3090 (used, $700–$850)

The enthusiast's default. 24 GB of GDDR6X at 936 GB/s, 10,496 CUDA cores, 350 W TDP. For memory-hard wallet KDFs it is roughly 60 percent of the 4090's rate at 35 percent of the price. Two used 3090s on a mining frame cost about $1,600 all-in and deliver 200 KH/s on Bitcoin Core, enough to chew through a serious hint-guided attack in a week or two.

RTX 4090 (new, $1,800–$2,200)

Best raw performance per card. Also 24 GB, but with 1,008 GB/s bandwidth and a sharply improved compute-to-bandwidth ratio. If you have a budget and want a single-card rig that does not embarrass itself, the 4090 is it. The 16-pin connector does require care — use the cable that came with the card, not an adapter, and never bend within 35 mm of the socket.

RTX 3060 12 GB (new, $290)

The budget recovery card. About 30 percent of a 3090's rate, but four can fit in a single motherboard for the price of one 3090. For Scrypt-heavy modes where VRAM matters, the 12 GB is enough; for SHA-heavy modes (which most wallet modes are not) the 3060 is merely okay. If you are optimising for dollars-per-hash and have time to tune, a 4x3060 rig is genuinely competitive.

AMD RX 7900 XTX (new, $900)

The heretic's choice. Under ROCm OpenCL on Linux it hits about 85 percent of a 4090's rate on mode 11300. Caveats: driver stability on suspend/resume is poor, some new hashcat modes ship with broken AMD kernels for months, and you will spend a weekend getting ROCm 6.x talking to your kernel. Worth it only if you already know your way around AMD Linux.

Step three: the rest of the platform

Motherboard and CPU

For a 1–2 GPU rig, any modern consumer board (B650, Z790) works. For 4+ GPUs go straight to an Asus B250 Mining Expert or an MSI Z390-A Pro with PCIe bifurcation — they were built for this. CPU is irrelevant to hashcat performance; a cheap i3-10100 or Ryzen 5 5500 is ample. 16 GB of DDR4 is enough; hashcat is entirely GPU-resident.

Power supply

Budget 400 W per 3090, 500 W per 4090, 200 W per 3060. Add 150 W for the rest of the system. Buy Platinum rated 80+ and never run above 80 percent of nameplate — you will save the difference in electricity within a year. Corsair AX1600i and Super Flower Leadex Titanium 1600 are the two models that reliably survive multi-GPU rigs for years.

Risers

PCIe 4.0 x1 risers from reputable brands (MintCell, Veddha, 6pack) are fine. USB 3.0 mining risers work but their voltage regulation is often marginal — use molex-to-molex powered risers, never SATA-to-molex chains. For hashcat, unlike for mining, riser bandwidth is a non-issue.

Chassis and cooling

Closed case with three or more cards is a bad idea; thermal throttling will cost you more hashrate than you saved in looks. Build on an open mining frame ($40–$80) with GPUs vertically mounted and at least 40 mm of air between cards. Keep ambient under 25 °C and GPU junction temps under 85 °C; the 3090's GDDR6X memory pads are the usual weak point and benefit from a one-time replacement with Gelid GP-Extreme.

Step four: operating system and drivers

# Ubuntu 22.04 LTS minimal install, then:
sudo apt update && sudo apt install -y build-essential git curl

# NVIDIA driver (550+ recommended for Ada + Ampere)
sudo add-apt-repository ppa:graphics-drivers/ppa
sudo apt install nvidia-driver-550 nvidia-utils-550
sudo reboot

# Hashcat 6.2.6 binary (do NOT use distro package — it's usually ancient)
wget https://hashcat.net/files/hashcat-6.2.6.tar.gz
tar xf hashcat-6.2.6.tar.gz
cd hashcat-6.2.6
./hashcat.bin -b -m 11300           # benchmark Bitcoin Core

# Power cap to 75% — keeps 95%+ of hashrate, drops temp 8C, saves electricity
sudo nvidia-smi -pm 1
sudo nvidia-smi -pl 270              # 3090 default 350, cap to 270
sudo nvidia-smi -lgc 1400,1900       # lock clocks to stable range

The ./hashcat.bin -b benchmark is the first thing you run. If your numbers come in 20 percent below the table above, your cooling or power is the bottleneck — fix that before investing more time. nvtop shows real-time utilisation; a healthy hashcat run pins all GPUs at 99 percent with junction temp around 70 °C.

Step five: candidate generation strategy

Raw brute force over random strings is almost always a dead end. Your success hinges on targeted candidate generation. The professional stack looks like this:

# 1. Personal base wordlist — seeded by hints
# Names, pets, addresses, band names, favourite movies, dates
cat hints.txt

# 2. Dictionary + rules — huge payoff per GPU cycle
hashcat -m 11300 wallet.hash hints.txt -r rules/OneRuleToRuleThemAll.rule

# 3. Rockyou + dive rules if hints exhausted
hashcat -m 11300 wallet.hash rockyou.txt -r rules/dive.rule

# 4. PRINCE — combines dictionary words into candidate phrases
hashcat --princeling wallet.hash hints.txt

# 5. Mask attacks — for known structures like Name1234!
hashcat -m 11300 -a 3 wallet.hash "?u?l?l?l?l?d?d?d?d!"
hashcat -m 11300 -a 6 wallet.hash hints.txt ?d?d?d?d

# 6. Incremental masks last — physically limited
hashcat -m 11300 -a 3 wallet.hash --increment --increment-min 6 \
    --increment-max 10 ?1?1?1?1?1?1?1?1?1?1 -1 ?l?d

Step six: the money math

A home rig wins when you expect to run more than 5,000 hours of GPU time, which corresponds to roughly five wallets per year or a single wallet with a very large keyspace worth committing weeks to. For a single locked wallet with a modest hint set, vast.ai for $100 total or our success-fee service is almost always cheaper.

When NOT to build a rig

• Your password is weak (under 8 chars, lowercase). Your CPU can do it in an afternoon. Use john the ripper, not hashcat.
• You cannot name a plausible candidate. No hardware solves "I have absolutely no memory of what I chose".
• The wallet is worth under $3,000. After hardware, electricity, and your time, you will lose money.
• Your KDF is Scrypt-heavy and wallet is small. 900 H/s on mode 15700 means even a 100 million candidate attack takes 30 hours per GPU — a service does this in minutes across 20 GPUs.
• You have the seed phrase already. Restore; move on with your life.

When a rig genuinely shines

• Multiple wallets to try over months — the per-attempt cost crashes.
• Exotic modes not supported by commercial services (rare token contracts, custom brainwallets).
• You want the hardware anyway for AI/ML, 3D, or gaming — hashcat becomes a free side-benefit.
• You want full control over the candidate pipeline and are willing to write Python.

The honest bottom line

Building a home cracking rig is a rewarding project if you already enjoy hardware, Linux, and the Shannon-entropy arithmetic of password attacks. It is almost never the economically correct path for a single stuck wallet. If this is your first rodeo and your goal is just to get your coins back, send the file to a professional service, pay the success fee, and use the money you would have spent on GPUs for something else. If instead you are intrigued by the craft — welcome, you will learn a tremendous amount, and the 3090 rig will still be there when the next wallet goes quiet.

Last updated 2026-04-21. Benchmark numbers verified on Ubuntu 22.04 + driver 550.120 + hashcat 6.2.6, with GPUs power-capped to 75 % TDP for stability.