Skip to content

W974b: the silicon answered — Done=1 and ok=1 read off the die, after twenty-three waves - #682

Merged
gHashTag merged 3 commits into
mainfrom
silicon-answered-w974
Aug 21, 2026
Merged

gHashTag merged 3 commits into
mainfrom
silicon-answered-w974

Conversation

@gHashTag

Copy link
Copy Markdown
Collaborator

The silicon answered

On xc7a200tfbg676-1, IDCODE 0x3636093, board 1:5:

--- hardware, board 1:5 ---
OK   B1 our bitstream   Done Some(1)  (must be 1)
     reading USER2, derived from the FASM
OK   B2 read            0xa5a5a5a7   magic, ok=1 beat=1   on index [0]

PASS -- the silicon answered, and its answer is ok=1.

The design is mvp_ternary_classifier, carried from its .t27 spec: 123 LUT, 52 CARRY4, 0 DSP48E1, Fmax 80.35 MHz (PASS at 70.77), bitstream 9 730 834 B. The word read back through USER2 is derived from the FASM, not from the source, and carries the magic 0xA5A5A5A with ok = 1, beat = 1.

The axis this project had no number on now has one, and it is a pass.

The first attempt failed on the address, not the artefact

Ten minutes earlier, the same bitstream on the same bench:

FAIL B2 read   no magic on any cable
  index 0: ALL ZERO -- dead chain or no BSCANE2 in this bitstream
  index 1: usb_open(index=1): rc=-3 device not found
  index 2: usb_open(index=2): rc=-3 device not found

The service defaults to the three-cable bench — 1:4 / 1:6 / 1:8 — recorded in the SSOT. This bench now has exactly one cable, at 1:5, which #? W948 measured and wrote down. Passing --busdev-num 1:5 turned the identical artefact into a pass.

Two things worth keeping:

  • "ALL ZERO — dead chain or no BSCANE2 in this bitstream" names the wrong suspect. A read of all zeros must implicate the address before the artefact: wrong cable and no BSCANE2 are indistinguishable at the wire.
  • A default that encodes a physical layout expires silently — correct when written, a bug once the bench changed, with no error at build time and a runtime message pointing elsewhere.

What this establishes, and what it does not

Established: the whole chain works end to end — .t27 spec → generated Verilog → yosys → nextpnr → FASM → frames → bitstream → JTAG load → Done = 1 → a value read off the die matching what the FASM predicts. Every stage demonstrated, not assumed, in about a minute per run.

Not established: anything about the φ-format's merit. What answered is a classifier, and its verdict is a liveness-and-integrity check — not a measurement of the number system. The format's own operators (gft_sadd, gft_signed_mac, gft_signed_dot4) have synthesis and timing (#681) and are unread.

The honest statement of the hardware axis today: the path is proven and the classifier passes; the operators are built, timed, and unread. The next run closes that gap with the same command and a different spec.

python3 research/arxiv_tnf/verify_numbers.py

270 derived checks, all passing. Related: #680, #681.

🤖 Generated with Claude Code

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant