Pearl FP8 Fork: How PRL Mining Moves to Floating Point and Who Stays on the Network
18/11/2026
Updated November 18, 2026. FP8 is live on testnet; the mainnet fork is expected in 4 to 8 weeks.
Kryptex Pool supports every Pearl fork, including future ones, to keep the network secure and stable. You can follow fork news on our site: pool.kryptex.com.

⚡ TL;DR
- Pearl (PRL) is preparing a hard fork: mining moves from integers,
INT7 × INT7 → INT32, to floating point,FP8 × FP8 → FP32. - FP8 is already live on testnet2: the fork activated at block 150283.
- The Pearl team aims to activate it on mainnet in 4 to 8 weeks. There is no exact date or block yet.
- The spec is PIP-3, status: Proposed. The new certificates are version 4.
- Old nodes will not accept the new certificates, so every node has to upgrade.
- Only whitelisted GPUs will be able to mine. PIP-3 names NVIDIA Blackwell, and the whitepaper also covers Hopper. The RTX 40, 30, and 20 series and AMD cards are not in the documents.
🧠 Mining That Computes Neural Networks
Pearl is a blockchain where a GPU does not grind through hashes but performs the same math it uses to run an AI model, such as a chatbot. In ordinary mining, only the network needs the result. In Pearl, the GPU multiplies large tables of numbers (matrices), and that multiplication is the proof of work.
Every AI model contains a huge set of such tables, called weights. To answer your question, the model multiplies your data by them again and again. That is why AI needs powerful GPUs, and why Pearl chose this operation.
Here is how it works. The miner takes the model's weights and the input data, or builds the tables itself when no real AI job is available. To stop anyone from sending an empty shell, the protocol mixes "noise" into the tables. The multiplication result is split into small pieces called tiles. If a tile's hash falls below the target, the miner has found a share. Shares go into a certificate that the network checks.

➗ Integers and Fractions
Today Pearl computes in integers (INT), or whole numbers: 3, 17, −5. The format is called INT7, and the running total is kept in INT32. Whole numbers keep things simple: every GPU gives the same answer down to the last bit, so the network can verify anyone's result.
But AI model weights are stored as fractions, such as 0.0003 or 150.25. This format is known as floating point (FP). To mine in integers, the miner rounds the model's weights to INT7, a step called quantization. FP8 is a compact way to write a fractional number in just 8 bits. After the fork, the miner computes in a format closer to the format that AI models themselves use.

🔬 What PIP-3 Changes
The new scheme in PIP-3 works like this:
- Input tables arrive in BF16 and are converted to FP8 (E4M3) before multiplication.
- The noise is also FP8. The tables' hashes and the chain state fully determine it, so a miner cannot choose the noise.
- The numbers are multiplied in FP8, and the running total is kept in the more precise FP32.
- The miner declares in advance which GPU type it computes on.
With fractions, the answer depends on the GPU: different chips round differently, so results drift slightly apart. The PIP-3 authors reject any tolerance for that drift: it would let a miner nudge an answer to get more chances to find a share without extra computation (a grinding attack). So the network tolerates no difference: it reproduces the computation bit for bit, following the exact rounding rules of a specific GPU type. That is where the whitelist comes from: only GPU types with written rules can mine.
🎯 Why the Fork
The PIP-3 authors give two reasons. First, AI models natively work with fractions, and rounding to integers exists only for mining. Second, mining and useful AI work run on the same operation, so the extra cost is small. By the founders' account at the September 15 AMA, a Q&A session, the extra cost is 5–7% per table multiplication and lower across the whole model run, effectively zero in some tests. These are the team's own estimates, not independent benchmarks.
The price is lock-in to specific GPUs: PIP-3 calls this dependence unavoidable for floating point.
FP4, an even more compact 4-bit format, is not part of this fork. At the AMA, the founders said it will come later, alongside FP8, not instead of it.
🎮 Who Stays on the Network
Neither PIP-3 nor the whitepaper gives a full list of models, so the mainnet node release will settle the exact set. The grouping below draws on Pearl's documents and chip capabilities (NVIDIA Compute Capability list).

🟢 Blackwell and Hopper: in Pearl's documents. PIP-3 names Blackwell. The whitepaper specifies the arithmetic for both Blackwell and Hopper, and H100 verification was added to the code on November 17, 2026 (PR #311). At the AMA, the founders confirmed that the RTX 5090 and RTX PRO 6000 Blackwell will work after the fork.
🟡 Ada and AMD: FP8 in hardware, not in the protocol. Ada Lovelace (RTX 40 series, L4, L40, L40S) and AMD cards with hardware FP8 (MI300, MI350, RDNA4) can compute in FP8, but Pearl's documents do not cover them.
🔴 Ampere, Turing, and older AMD: no hardware FP8. That covers the RTX 30 and RTX 20 series, the A100, the CMP 170HX and CMP 90HX mining cards, and AMD's RX 6000 and RX 7000 series.
Community members propose workarounds, such as an FP16 mode for the A100 (PR #369). These are contributor initiatives, not network decisions.
📅 When Will the Fork Happen
On testnet, the fork has already happened. FP8 activated on testnet2 at block 150283, and the network runs the pearld 1.6.0-fp8-testnet1 node.
On mainnet, the Pearl team is aiming for 4 to 8 weeks from now. There is no exact activation date or block yet. The team will spend that time validating the scheme with the community. It expects the current FP8 protocol version to be the last one before mainnet, unless testing uncovers a major issue.
The testnet miner is a complete reference implementation of the protocol. Pearl's developers keep optimizing mining and inference speed, so expect miner updates both before and after activation.
✅ What Miners Should Do
- Blackwell and Hopper: the fork opens FP8 mining to them. Watch for the mainnet node and miner releases.
- RTX 40, 30, and 20 series, CMP, and AMD cards: mining continues as usual until the fork. After activation, under the current documents, these cards cannot produce valid certificates. Factor that risk into your payback estimates.
- Before the fork, update your miner to a version that supports the new protocol.
- Where to watch: Kryptex Pool, the PIP repository, and Pearl releases.
⛏ How to Mine PRL Now
Until the fork, PRL mines as usual. To start on Kryptex Pool, you need three steps:
- Wallet. Create a PRL address with this guide.
- Miner. Pick a miner in the Pearl miners overview.
- Connection. Enter your wallet address and worker name on the PRL pool page, then start the miner with the server address
prl.kryptex.network:7048.
For a full walkthrough, see How to mine Pearl (PRL).
❓ FAQ
When is the fork? FP8 is already live on testnet. On mainnet, the Pearl team aims to activate it in 4 to 8 weeks; there is no exact date or block yet.
Will the RTX 5090 and RTX PRO 6000 Blackwell mine? Yes. The founders confirmed both cards at the AMA.
Will the RTX 4090 mine after the fork? Not under the current documents: they include no Ada profile. The founders did not comment on the 4090 specifically.
Will the RTX 3090 and other RTX 30 cards mine? No, based on the current documents: Ampere chips have no FP8 hardware, and there is no profile for them.
Does Kryptex Pool support the FP8 fork? Yes. We support every Pearl fork, including future ones, to keep the network secure and stable. Follow fork news at pool.kryptex.com.
Need Help?
Have any questions, something unclear, or trouble connecting?
Contact support — we're happy to assist!
- Support Ticket.
- Telegram Support.
- Discord Support.
- Email support at support@kryptex.com.


