Blog
Equal stored width removed an accuracy lead
A corrected equal-stored-width remeasurement withdrew an earlier 2.1×/2.6× lead over takum and records the oracle and budget defects that changed the reading.
- THE OLD READINGA lead was read from unequal footing.
- EQUAL WIDTHThe same stored width for both sides.
- THE WITHDRAWALThe earlier lead was withdrawn, not defended.
[measured — equal stored width] The site’s earlier 2.1×/2.6× accuracy lead over takum was withdrawn after a remeasurement exposed two confounds: an oracle sign error on negative codes and a nominal rather than equal-stored-width budget.
At 16 stored bits, TNF(4,8) records 5.323e-3 against takum16 at 5.697e-3. At 32 stored bits, TNF(4,24) records 1.203e-7 against takum32 at 1.264e-7. Under the equal-stored-width comparison, these values are treated as a tie rather than an accuracy lead; the earlier 2.1×/2.6× statement remains visible as retracted.
Why the result changed
The earlier number combined a faulty sign path with a budget that did not count stored width on the same basis. The correction keeps the old claim visible, records why it was withdrawn, and moves the comparison to the measured equal-width rows.
| Stored width | TNF | takum | Reading |
|---|---|---|---|
| 16 bits | 5.323e-3 | 5.697e-3 | tie at equal stored width |
| 32 bits | 1.203e-7 | 1.264e-7 | tie at equal stored width |
A corrected number is more useful when the retracted number remains visible.
What this does not establish
- This is an equal-stored-width software/oracle remeasurement, not a measurement on the binary ALINX AX7203 FPGA and not a custom-silicon result.
- It does not provide a head-to-head hardware comparison against takum.
- It does not establish downstream model accuracy, throughput, energy, or area.
- It does not establish a universal accuracy ranking for numerical formats; the corrected rows only replace one withdrawn claim.
The useful engineering result is procedural: preserve the withdrawn figure, name the oracle defect and budget mismatch, and keep the equal-width measurement beside the correction. That makes a change in belief auditable rather than silent.
What this does not settle
- This is an equal-stored-width software/oracle remeasurement, not a measurement on the binary ALINX AX7203 FPGA and not a custom-silicon result.
- It does not provide a head-to-head hardware comparison against takum.
- It does not establish downstream model accuracy, throughput, energy, or area.
- It does not establish a universal accuracy ranking for numerical formats; the corrected rows only replace one withdrawn claim.
Receipts
Work with me
Want this kind of check on your own design?
I audit RTL and build independent, bit-exact models, then take the result through synthesis and, when useful, onto an Artix-7 board. The first conformance module is free.