Pearl’s proposed FP8 fork (PIP-3) changes mining from integer math to 8-bit floating point and, as written, only accepts Nvidia Hopper and Blackwell hardware. If you mine Pearl (PEARL) on an RTX 4090, an RTX 3090 or a rented A100, you want to know where you stand. This page is a GPU-by-GPU checker. The background is in Pearl’s FP8 fork, explained.
The rule in one paragraph
Floating-point results differ bit by bit between chips, so PIP-3 makes miners declare their hardware and has verifiers reproduce the result exactly as that hardware would. That only works for modeled hardware, so the draft whitelists two device families: Hopper (device 0) and Blackwell (device 1). The per-model lists below follow from each card’s architecture. They match a breakdown published by PearlStation, a third party, on October 22, 2026, based on the spec as written.
GPU checker
Hashrates are for today’s integer protocol, in TH/s (tera matrix multiplications per second), from AlphaMine (May 2026) unless marked as Kryptex. They are not FP8 hashrates, which nobody has published yet.
| GPU | Architecture | FP8 hardware | After PIP-3 as written | Hashrate today |
|---|---|---|---|---|
| RTX 5090 | Blackwell | Yes | Keeps mining | 220 TH/s (Kryptex) |
| RTX 5080 | Blackwell | Yes | Keeps mining | About 105 TH/s |
| RTX 5070 Ti, 5070, 5060 Ti, 5060, 5050 | Blackwell | Yes | Keeps mining | Not listed |
| RTX PRO (Blackwell), DGX Spark | Blackwell | Yes | Keeps mining | Not listed |
| B200, B300, GB200, GB300 | Blackwell | Yes | Keeps mining | Not listed |
| H100, H200, H800, H20, GH200 | Hopper | Yes | Keeps mining | Not listed |
| RTX 4090 | Ada | Yes | Not listed (PR #16 would add it) | 150 to 160 TH/s |
| RTX 4080, 4070, 4060 | Ada | Yes | Not listed (PR #16 would add it) | Not listed |
| RTX 6000 Ada, L40S, L40, L4 | Ada | Yes | Not listed (PR #16 covers RTX 6000 Ada, L4, L40S) | Not listed |
| RTX 3090 | Ampere | No | Stops mining | About 90 TH/s |
| RTX 3070 | Ampere | No | Stops mining | 50 to 60 TH/s |
| RTX 3060 Ti | Ampere | No | Stops mining | 40 to 50 TH/s |
| CMP 100-210 | Older Nvidia (no FP8) | No | Stops mining | About 70 TH/s |
| A100 | Ampere | No | Stops mining | Not listed |
| AMD MI300X, MI325X, MI350X | AMD CDNA | Yes | Not listed | Not listed |
| AMD RX 9070 XT, 9070, 9060 XT | AMD RDNA4 | Yes | Not listed (PR #18 proposes RDNA4) | Not listed |
| AMD RDNA3 and older, Intel Arc | Various | No | Not supported | Not listed |
“Not listed” in the hashrate column just means we have no sourced figure. AlphaMine notes each miner instance needs about 2 GB of VRAM, so memory is rarely the limit.
The three groups
Keeps mining: Hopper and Blackwell
This is every RTX 50 series card, Blackwell workstation and datacenter parts, and Hopper datacenter cards like the H100 and H200. According to Kryptex Pool, the founders confirmed the RTX 5090 and RTX PRO 6000 Blackwell will keep mining. Note that the official Pearl Research pool only accepts H100 and H200 and charges a 20% fee; community pools support consumer cards. See Pearl mining pools compared.
Has FP8 but is not listed: Ada and AMD
The RTX 40 series, the L40S and L4, and AMD’s recent chips all have FP8 hardware. They are left out because the protocol has not modeled their exact floating-point behavior, not because they lack the feature. Two community proposals aim to fix that:
- PR #16 (Ada, sm_89), opened October 27, 2026, covers RTX 40 series, RTX 6000 Ada, L4 and L40S and asks to activate together with PIP-3. Its author says the Rust reference matches RTX 4070 Ti SUPER silicon on 1,662,976 measurements.
- PR #18 (AMD RDNA4, gfx12), opened October 31, 2026, covers RX 9060, RX 9070 and Radeon AI PRO R9700.
Neither is merged as of November 5, 2026. If PR #16 lands before activation, the RTX 4090 and friends move into the first group. Track it in the PIPs repository; our guide to Pearl Improvement Proposals explains how to read it.
No FP8 hardware: Ampere and older
The RTX 30 series, A100 and CMP cards have INT8 tensor cores, which is why they mine well on today’s integer protocol, but no FP8 units. No device entry can fix that, so as the fork is planned these cards stop producing valid Pearl work at activation. The same goes for AMD RDNA3 and older and Intel Arc. Why the format matters is covered in what is FP8.
What to do with your card
- Blackwell or Hopper: keep mining and keep your node and miner updated from the official releases page.
- Ada: keep mining, and watch PR #16. Do not assume either outcome.
- Ampere and older: mine while you can, but price in a likely end date before buying more of these cards. Our mining profitability guide helps with payback maths, and the best GPUs for Pearl mining compares upgrades.
- Renting: prefer short rentals on cards outside Hopper and Blackwell, since a long commitment carries fork risk.
Whatever card you mine with, the payout side stays the same. Oyster’s miner mode gives you one fixed prl1p address for your pool. See the mining payout wallet guide.
Frequently asked questions
Will the RTX 4090 mine Pearl after the FP8 fork?
Not as PIP-3 is written today. The RTX 4090 is an Ada card: it has FP8 hardware, but the draft only whitelists Hopper and Blackwell. A community proposal (PR #16) asks to add Ada devices, including the RTX 40 series and L40S, at the same time as PIP-3. As of November 5, 2026 it is not merged. Until the fork activates, the RTX 4090 keeps mining.
Will the RTX 5090 keep mining Pearl?
Yes. The RTX 5090 is a Blackwell card, which the spec whitelists. According to Kryptex Pool, the founders confirmed the RTX 5090 and RTX PRO 6000 Blackwell keep mining after the fork.
Can RTX 3090 or A100 cards mine after the fork?
No, as far as the spec goes. Ampere cards such as the RTX 30 series, A100 and CMP mining cards have INT8 tensor cores but no FP8 hardware, so they cannot produce version 3 certificates. They mine normally until activation.
When do I need to stop mining on an unsupported card?
Only when the fork activates. As of November 5, 2026 there is no activation height. Watch the Pearl node releases page and your pool's announcements for a date.
