What is posit arithmetic?
An alternative number format to IEEE floating point that uses a variable-length regime field, allocating the highest precision near magnitude one and degrading gracefully toward the extremes of the range.
PRODUCT · FAMILY 01
Posit-based Processor
A processor using posit arithmetic, an alternative number format to IEEE floating point that allocates precision differently across the dynamic range. The interesting property is accuracy per bit near unity magnitude, which is where a great deal of numerical work actually lives.
WHAT IT IS
The short version.
IEEE floating point allocates its bits uniformly: the same exponent and mantissa split across the whole range, whether the value is 10 to the minus 30 or 10 to the plus 30. Most real computation does not use that range uniformly. It clusters near unity.
Posits allocate precision differently, using a variable-length regime field so that accuracy is highest near magnitude one and degrades gracefully toward the extremes. For workloads that live in that middle range — which includes a great deal of numerical and machine learning computation — that can mean comparable accuracy in fewer bits, or better accuracy in the same bits.
Fewer bits means less memory bandwidth and less energy per operation, which is the argument that makes this worth investigating for AI and scientific compute rather than a purely academic curiosity.
DETAIL
Characteristics
| Parameter | Detail |
|---|---|
| Function | Processor implementing posit arithmetic |
| Format property | Variable-length regime giving highest precision near unity magnitude |
| Argument | Comparable accuracy in fewer bits for workloads clustered near unity |
| Relevance | Memory bandwidth and energy per operation in AI and scientific compute |
| Typical markets | AI, scientific computing, research |
Maturity — silicon-proven, FPGA-validated or RTL stage — is confirmed at enquiry for the specific configuration you need, rather than claimed generically here.
APPLICATIONS
Where it is used.
- Machine learning inference — where reduced precision is already standard and the trade is well understood.
- Scientific computing — numerical workloads where accuracy per bit is the figure of merit.
- Energy-constrained compute — where memory bandwidth dominates the energy budget.
- Research platforms — evaluating alternative numeric formats against real workloads.
- Accelerator datapaths — as an arithmetic option in custom compute.
- Embedded numerics — where floating point hardware would be too costly.
WHAT YOU RECEIVE
Deliverables and support.
- Processor IP and supporting documentation.
- Guidance on evaluating format suitability against a specific workload.
- Engineering support through integration and evaluation.
Next step
Send the target node, the interface requirements and the integration context. If this is not the right fit, that will be said early rather than discovered at integration.
COMMON QUESTIONS
Questions asked before an evaluation.
Why would posits be preferable to floating point?
For workloads whose values cluster near unity — which includes much numerical and machine learning computation — posits can deliver comparable accuracy in fewer bits, which reduces memory bandwidth and energy per operation.
Is this a product or a research programme?
That is exactly the right question to ask, and it is answered directly at enquiry for the specific configuration and target you have in mind rather than claimed generically here.
KEEP READING
Related work.
BUILD WITH FASTSTREAM
Bring us the difficult part.
Tell us the specification, the constraint and the deadline. Programmes that cross silicon, radio, embedded and AI are where Faststream is strongest.