FP8 is suddenly part of every Pearl (PEARL) mining conversation, because a proposed network upgrade called PIP-3 would move Pearl mining onto it. You do not need to be an AI engineer to follow why. This guide explains what FP8 is, what “E4M3” means, why AI uses it, how it compares with INT8 and why Pearl wants to mine with it.
FP8 in one sentence
FP8 is an 8-bit floating-point number format: every value fits in a single byte. Computers usually store decimals as 32-bit floats (FP32) or, in AI, 16-bit floats (FP16 or BF16). FP8 halves the memory of the 16-bit formats, so a GPU can move and multiply twice as many numbers with the same bandwidth.
How a floating-point number works
A floating-point number is scientific notation in binary. It has three parts:
- Sign: one bit saying positive or negative.
- Exponent: how big the number is, in powers of two. More exponent bits means a wider range.
- Mantissa: the significant digits. More mantissa bits means finer precision.
With only 8 bits to share, an FP8 format has to choose between range and precision. That choice is what the letters in its name describe.
E4M3 and E5M2
The paper “FP8 Formats for Deep Learning” by Micikevicius and others (2022) defined two FP8 formats that the industry adopted:
| Format | Sign bits | Exponent bits | Mantissa bits | Trade-off |
|---|---|---|---|---|
| E4M3 | 1 | 4 | 3 | More precision, largest finite value 448 |
| E5M2 | 1 | 5 | 2 | Wider range, less precision |
E4M3 is the one Pearl’s proposal uses. Its small range is why values are usually scaled before being converted: you divide a row of numbers by a scale factor so the biggest one fits under 448, store the scale separately, and multiply it back in later.
Why AI uses FP8
Neural networks are huge piles of matrix multiplications, and they turn out to be tolerant of low precision. Running those multiplications in 8 bits instead of 16 saves memory and lets more numbers flow through the GPU per second. Precision is protected in two ways:
- FP32 accumulation. Each product is FP8 × FP8, but the running sum is kept in 32-bit precision, so small errors do not pile up.
- Requantization. The FP32 result is scaled and converted back to low precision for the next layer.
This pattern, low-precision multiply, FP32 accumulate, then requantize, is the typical shape of FP8 inference on modern hardware.
Which hardware has FP8
| Nvidia architecture | Examples | FP8 tensor cores |
|---|---|---|
| Ampere | RTX 30 series, A100 | No (INT8 only) |
| Ada Lovelace (2022) | RTX 40 series, L40S, L4 | Yes |
| Hopper (2022) | H100, H200 | Yes |
| Blackwell | RTX 50 series, B200 | Yes |
Having FP8 hardware is necessary for Pearl’s planned FP8 mining but, as the spec is written, not enough. See will my GPU mine Pearl after FP8 for why.
INT8 vs FP8
INT8 is the other popular 8-bit format. It stores whole numbers from −128 to 127, evenly spaced. FP8 spreads its 256 possible values unevenly: dense near zero, sparse at the extremes. Neural network values tend to cluster near zero with a few large outliers, which is the shape FP8 handles well.
| INT8 | FP8 (E4M3) | |
|---|---|---|
| Size | 8 bits | 8 bits |
| Value spacing | Even | Finer near zero, coarser for big values |
| Bit-exact across chips | Yes, integer math is exact | Not guaranteed, results can differ per chip |
| Native format of most AI models | No, models are converted to it | Closer, models are natively floating point |
| Pearl mining today | Yes (as INT7 with one bit of headroom) | Proposed in PIP-3 |
That third row matters for a blockchain. Integer multiplication gives the same answer on any chip. Floating point does not always, because hardware can round and accumulate in slightly different orders.
Why Pearl wants FP8
Pearl is mined with proof of useful work: the mining is matrix multiplication, the core operation of AI. Today’s mining certificates prove integer multiplications. PIP-3 argues that “the overwhelming majority of AI models are natively floating-point”, so integer mining forces miners away from the math real models use. Its plan: take BF16 inputs, quantize them row by row to FP8 E4M3 with a BF16 scale, and multiply with an FP32 accumulator, the same pattern described above.
The rounding problem from the table is handled by declaring hardware. The miner states which GPU family it runs on, and the verifier reproduces the result bit-exactly as that hardware would. In the draft, only Hopper and Blackwell are modeled. The full story is in Pearl’s FP8 fork (PIP-3), explained.
Where to go next
- New to Pearl? Read what is Pearl (PEARL).
- Thinking about mining? Start with how to mine Pearl and the best GPUs for Pearl mining.
- Curious how rule changes get made? See Pearl Improvement Proposals (PIPs).
Frequently asked questions
What does FP8 mean?
FP8 means 8-bit floating point: a number format that stores each value in 8 bits, split into a sign, an exponent and a mantissa. It uses half the memory of 16-bit formats like FP16 and BF16 and is widely used to speed up AI models.
What is the difference between E4M3 and E5M2?
Both are FP8 formats from the 2022 paper 'FP8 Formats for Deep Learning'. E4M3 has 4 exponent bits and 3 mantissa bits, so it is more precise but tops out at 448. E5M2 has 5 exponent bits and 2 mantissa bits, so it covers a wider range with less precision. Pearl's PIP-3 uses E4M3.
Is FP8 better than INT8?
Neither is better in general. Both are 8 bits. INT8 spaces its values evenly, while FP8 trades precision for range, which suits the spread-out values found in neural networks. Most AI models are natively floating point, which is why Pearl's PIP-3 proposes moving mining from integer to FP8 math.
Which GPUs support FP8?
Nvidia added FP8 tensor cores with Hopper and Ada Lovelace (both 2022) and kept them in Blackwell. Ampere cards such as the RTX 30 series and A100 have INT8 tensor cores but no FP8.
