Pearl (PRL) hashrate by GPU: RTX 30, 40, 50, AMD and datacenter cards

This page lists Pearl (PRL) hashrate, power draw and gross daily revenue for each GPU as of October 6, 2026, from RTX 3060 to RTX 5090, Radeon and datacenter cards. Every figure below names its source and date. Pearl Research Labs publishes no consumer GPU miner and no hashrate figures, so all RTX and Radeon numbers come from third-party miners and benchmark sites.

• Fastest consumer card: RTX 5090, 417.9 TH/s at about 519 W (hashrate.no verified benchmark, October 6, 2026).
• Most work per watt: RTX PRO 6000 Max-Q (0.999 TH/W) and RTX 4090 (0.934 TH/W).
• Yield: 0.0201 PRL per TH/s per day, or $0.0229 at $1.14 per PRL, before fees (October 6, 2026).
• AMD Radeon: works, but the best Radeon (RX 9070 XT, 0.290 TH/W) does about a third of the work per watt of a tuned RTX 4090.

Pearl hashrate by GPU: main table

hashrate.no shows tuned, user-submitted runs it marks as verified. WhatToMine uses conservative, pool-reported figures. Revenue columns convert the hashrate.no speed, except the last one.

Data as of October 6, 2026, ~17:00 UTC. Sources: hashrate.no verified benchmarks (watts = TH/s ÷ TH/W) and the WhatToMine GPU list (* flagged by WhatToMine). PRL/day uses 0.0201 PRL per TH/s per day, derived from difficulty 34.81 million and a 2,277.05 PRL block reward at block 124,157 (official explorer). $/day uses $1.14 per PRL (CoinGecko, 16:52 UTC). Revenue is gross, before pool fee, miner dev fee and electricity; net figures are on the profitability page.

How pearlhash speed is measured

Pearl's proof of useful work is matrix multiplication (see what Pearl is). Miners multiply noised integer matrices tile by tile and hash each tile's result with BLAKE3; a tile wins if its hash falls below a target scaled by the work it contains. So a "hash" in pearlhash, the algorithm name miners use, stands for an amount of matrix work, not one hash-function call.

So TH/s (trillions of these units per second) on Pearl cannot be compared with TH/s on any other coin. Speed tracks tensor-core throughput and core clock, not memory, which is why tuned cards run their memory low.

The network total is implied by difficulty: on October 6, 2026, difficulty of 34.81 million at 165-second average blocks implied about 59.5 EH/s (1 EH/s = 1,000,000 TH/s); hashrate.no estimated 67.75 EH/s. Expected coins follow from your hashrate and the difficulty:

PRL per day = hashrate (H/s) × 86,400 × block reward / (difficulty × 2^48)

Why benchmarks disagree

• Miner version. The RTX 5090 did about 220 TH/s in late May 2026 and about 418 TH/s in October. SRBMiner 3.7.2 (October 6, 2026) adds a new experimental kernel.
• Miner-reported vs pool-reported. A pool estimates your speed from accepted shares, so rejects, stale shares, dev-fee time and luck lower or scatter it. WhatToMine uses pool-reported hashrate.
• Tuning and dates. Tuned user runs can beat a card's headline figure, and some headlines are months old (see the RTX 4090 and 3090 below). Miners older than the salted noise-seed hard fork (block 99,000, August 11, 2026) no longer produce valid shares.

Here is how far apart the published RTX 5090 figures are:

Data as of October 6, 2026. Sources: each site's own page or the miner's README.

As of October 6, 2026, a tuned RTX 5090 on a current miner does about 385–420 TH/s.

RTX 50 series

RTX 50 cards (Blackwell, sm120) are the fastest consumer GPUs for Pearl and the only consumer generation on the FP8 reference miner's allow-list. The RTX 5090 grossed about 8.39 PRL, or $9.56, a day at hashrate.no speed and $1.14 (October 6, 2026).

• RTX 5080 and 5070 Ti: as of October 6, 2026, the AlphaMiner README claims 230.32 and 196.10 TH/s, above hashrate.no's 203.6 and 170.5 TH/s.
• RTX PRO 6000: the Max-Q version has the best TH/W in hashrate.no's Pearl list. WildRig's 0.48.8 notes (June 24, 2026, before the August forks) gave 410 TH/s at 600 W for the full card.
• RTX 5070, 5060 Ti and 5060: no verified benchmark as of October 6, 2026.

Ignore GitHub repositories offering "5060 Ti" or "5090" miners from unknown accounts; download miners only from their developer's official repository (fake miners and scams).

RTX 40 series

The RTX 4090 (Ada, sm89) is the most efficient consumer card in hashrate.no's list: it does more work per watt than an RTX 5090.

• RTX 4090: 304.6 TH/s at about 326 W on hashrate.no, about $6.97 a day gross at $1.14 (October 6, 2026). Other published figures, as of October 6, 2026: 337.71 TH/s (AlphaMiner README), 323 TH/s at 386 W (hashrate.no user, September 24, 2026), 293 TH/s at 450 W (Kryptex device page), ~292 TH/s (KRig README) and 220 TH/s at 380 W (MiningBoard). Kryptex's RTX 4090 D page gives 275 TH/s at most.
• RTX 4070 Ti, 4070, 4060 Ti and 4060: no verified benchmark. WildRig 0.48.8 (June 24, 2026) says --pearlhash-kernel 2 can improve efficiency on the 4060 and 4070 Super/Ti; SRBMiner's alternative --pearl-k2 kernel lists the 4070 Ti.

RTX 40 (sm89) is not on the FP8 allow-list, and second-hand notes from a September 15, 2026 AMA said RTX 40 support was expected to end. Pearl has not confirmed this.

RTX 30 series and older (1660 Super, RTX 20, V100)

RTX 30 cards (Ampere, sm86) run at about 10–30% of RTX 5090 speed and use more power per TH/s than RTX 40 or 50 cards.

• RTX 3090: hashrate.no's headline (110.0 TH/s) comes from a May 28, 2026 run. User runs on September 29, 2026 reached 132–137 TH/s at 267–295 W, about $3.02–3.14 a day gross at $1.14 (October 6, 2026). Kryptex's device page lists 125 TH/s at 320 W (October 6, 2026).
• RTX 3060: 42.05 TH/s at about 99 W on hashrate.no, about $0.96 a day gross at $1.14 (October 6, 2026).
• RTX 3080 Ti, 3070 Ti and 3060 Ti: no verified benchmark.
• CMP 170HX: 140.21 TH/s per the AlphaMiner README (October 6, 2026). PeakMiner's --cmp-unlock installs a patched driver as root, which is risky.

RTX 20 series (Turing). PeakMiner, SRBMiner, BzMiner and WildRig support these cards; BzMiner's v100.03b8 notes call its Turing kernel about five times faster. We found no verified benchmark as of October 6, 2026.

GTX 16xx. PeakMiner's README says that since its v2.11.0, Maxwell, Pascal, Volta and GTX 16xx cards cannot mine Pearl under the rank-128 rules of the soft fork at block 96,251 (August 6, 2026). WhatToMine's figure is for the GTX 1660; we found no separate GTX 1660 Super benchmark, and the Super is in the same class. WildRig (0.50.3 and later) and BzMiner still run the GTX 1660. At WhatToMine's 0.65–0.8 TH/s it grosses under 2 cents a day, less than its electricity even at $0.10 per kWh.

Tesla V100. SRBMiner added it in 3.5.6 (August 17, 2026) and WildRig 0.50.4 made it up to 130% faster. We found no verified figure.

RTX 30 (sm86) and all older cards are off the FP8 allow-list.

Datacenter cards: H100, H200, A100, B200

Pearl Research Labs' only miner is the official vLLM miner (vllm-miner plus pearl-gateway, in the official repository). Its CUDA kernels compile only for sm90, so it runs on H100 and H200 cards only, mining solo to your own pearld node while serving an AI model through vLLM. No TH/s figure is published for it; the mining guide covers setup.

Third-party miners also run datacenter cards: SRBMiner (A100, V100, H100, H200, B200, B300), PeakMiner (A100, H100/H200, B200/B300) and WildRig.

Data as of October 6, 2026. All figures are developer- or site-reported, not verified benchmarks.

The H100 figure is unreliable. Only MiningBoard lists one, and it shows the SXM version, listed at twice the power, as slower than the PCIe card. Even if 700 TH/s were right, an H100 would gross about 14.05 PRL ($16.02) a day at $1.14, while the cheapest H100 PCIe on Vast.ai cost $1.25 an hour ($30.04 a day) on October 6, 2026.

SRBMiner's notes suggest a 900 to 1,200 MHz core lock for H100 efficiency (3.5.4 and 3.6.2). The "Official PearlHash miner" on GitHub belongs to the PearlHash pool, not Pearl Research Labs. Under the FP8 proposal, H100/H200 and B200/B300 stay in; the A100 drops out, though PR #369 proposes an FP16 path for it.

AMD Radeon

AMD cards mine Pearl with community miners; the official and FP8 reference miners are NVIDIA-only. Guides saying "NVIDIA only" or "RDNA 2 only" are out of date.

Data as of October 6, 2026. Sources: each miner's README and release notes; KRig from its README and Kryptex's miner.download page.

As of October 6, 2026, the fastest consumer Radeon is the RX 9070 XT: 92.36 TH/s on hashrate.no, 98.5 TH/s on Kryptex's device page, ~106 TH/s in the KRig README. The KRig README also lists the RX 7900 XT at ~41.4 TH/s and RX 6750 XT at ~17.8 TH/s.

Intel Arc and Mac

For Intel Arc, Blockz-miner documents Pearl support on Battlemage (B-series) cards only. BzMiner's README also lists Intel, but other Intel claims are unconfirmed, and we found no Intel Arc benchmark on hashrate.no or WhatToMine as of October 6, 2026. On macOS only BzMiner runs, and we found no published Mac benchmark.

Efficiency per watt

TH/W (TH/s per watt) decides whether a card covers its power bill. Break-even is the electricity price at which a day's power costs as much as the day's gross revenue.

Data as of October 6, 2026, ~17:00 UTC. Sources: hashrate.no verified benchmarks; break-even uses 0.0201 PRL per TH/s per day and $1.14 per PRL, before fees.

At $0.10 per kWh, a day of power costs $1.25 for an RTX 5090 and $0.78 for an RTX 4090; at $0.30, $3.74 and $2.35. A 2% miner dev fee and a typical 1–2% pool fee lower each break-even by 3–4%. Tuning moves efficiency: the 323 TH/s RTX 4090 run works out to 0.84 TH/W, and the September RTX 3090 runs to about 0.45–0.51 TH/W instead of 0.344.

Overclock and power-limit tips

Benchmarks and miner developers agree: lock the core clock with a positive offset, lock memory low (5,001 MHz on RTX 30 and 40 cards, as WildRig's developer also advises) and cap power.

Data as of October 6, 2026. Sources: hashrate.no benchmark pages for each card; WhatToMine's RTX 5090 page.

In SRBMiner, the RTX 4090 settings for GPU 0 look like this:

--gpu-coffset0 250 --gpu-cclock0 2400 --gpu-mclock0 5001 --gpu-plimit0 400

WildRig uses --gpu-core-offset, --gpu-core-clock, --gpu-memory-clock and --gpu-powerlimit. BzMiner's --pearl_opt takes auto, eff or hr.

• Experimental kernel: SRBMiner 3.7.2's --pearl-k1 (RTX 30, 40 and 50) needs memory at 810 MHz and fresh tuning of power limit and core offset. Errors E2001 or E1027 may mean the core offset is too high.
• Drivers: NVIDIA 580 or newer (SRBMiner; Fl4shMiner on Linux), and newer than 610 for RTX 50 on Windows (PeakMiner).
• Method: change one setting at a time and compare your miner's speed with the pool's reported hashrate and rejected shares.

How we collect numbers

• Benchmarks: hashrate.no's verified figure per card (its headline, not the highest submission) and WhatToMine's GPU list. Watts are TH/s divided by hashrate.no's TH/W. They cover the card only, not the rest of the rig or power-supply losses, so a wall meter will read higher.
• Other figures: pool websites and the READMEs and release notes of miner developers and pools, with their date. We take nothing from anonymous GitHub repositories, and we flag unreliable data such as the H100 numbers.
• Revenue: the formula above. hashrate.no's yield is lower (0.0172 PRL per TH/s per day) because it uses a higher network hashrate estimate. Subtract the miner dev fee (0% WildRig, 2% SRBMiner, BzMiner and PeakMiner), the pool fee and electricity.
• Freshness: blocks averaged about 165 seconds against the 194-second target (3 min 14 s), so difficulty rose about 2.6% in the 24 hours before our snapshot. Until block times return to target, revenue per TH/s keeps falling even at flat price and hashrate.

Frequently asked questions