CRIONE
Software testing software.
A day to celebrate.
The ten cores compiled. The testbenches ran. The store gate held an anonymous hit at 302 and let no raw zip through. The generator refused a license signed by another key, refused a license one byte edited, refused an expired license. No silicon was touched. This is software testing software, and today it passed.
Ten cores, one chassis, a shelf behind a gate.
Scale carries the host. REV runs parallel compute. Nexus moves network and storage. Train and Infer take the model, one side and the other. Vision reads the camera. Lingua reads the text. Spike handles the asynchronous channel. Qubit sits on the same die as the classical pipeline. Edge closes the set for the 42 mm² part.
Under the ten cores sit the common blocks that every core calls: the color_math_alu, the photonic link, the voxel cell, the quantum coprocessor, the HBM controller and the PCIe PHY, the exergy mixer, the quantum reactor power interface, the synthetic-photosynthesis power rail, the sealed-loop summ engine, the normal-resonance pipeline, the Peltier nano-magnetic cooler, the EM-IC push-pull ring, and the qbeam wireless chain. One folder.
Behind it sits the shelf: 5,040 configurations across twelve generations and ten chips, each in its own package, memory tier and TDP class. The shelf is Members-only. Verification keeps it honest.
Seven measurements every core must pass. One PASS line prints the numbers.
Every <core>_m_tb.v runs the same self-judging routine. It resets the DUT, feeds the common blocks their steady telemetry for sixty clocks (trickle 100 mW, harvest 50 mW, temperature 293 K, hall-sensor bitmap 0xFF), fires one token up the qbeam chain, then checks seven wires. Any zero is a FAIL. All seven high prints one PASS line.
- resonance_locked
- The normal_resonance_pipeline achieved phase lock at the target frequency.
- loop_sealed
- The summ_engine_sealed_loop summation closed: inputs balance outputs across the window.
- cool_in_setpoint
- The peltier_nano_magnetic_cool block held the die at 293 K (20 °C) under load.
- ring_balanced
- The emic_push_pull_ring reached balance across its eight coil pairs.
- reactor_mw > 0
- The quantum_reactor_power_iface accumulated real milliwatts into the rail counter.
- chain_received > 0
- The qbeam_wireless_chain received at least one token destined for this node.
- members_all_ok
- Every common block reported OK up to the top. A single quiet failure anywhere flips this to zero.
- PASS line
- [TB] PASS <core>_m: mw=<N> chain_rx=<N> — two numbers Scale also carries exergy=<N>.
- Guard
- 100 µs timeout. If the routine does not finish by then the testbench prints FAIL and exits 1.
Ten cores, five hundred and four configurations each, averaged.
Every core ships in five hundred and four flavors across twelve generations and roughly forty-two package / memory / TDP permutations. The mean die area and the mean thermal design power per core, taken over all five hundred and four rows of its shelf:
- Scale (S)
- The host. Mean die 110 mm². Mean TDP 73 W. Server-blade range on 7 nm.
- REV (R)
- Parallel compute. Mean die 220 mm². Mean TDP 225 W. Accelerator-card range.
- Nexus (D)
- Network and storage. Mean die 95 mm². Mean TDP 50 W. DPU range.
- Train (T)
- Model training. Mean die 280 mm². Mean TDP 499 W. The largest and hottest in the set.
- Infer (I)
- Model inference. Mean die 150 mm². Mean TDP 169 W. Inference-blade range.
- Vision (V)
- Camera and image. Mean die 130 mm². Mean TDP 135 W.
- Lingua (L)
- Text and language. Mean die 250 mm². Mean TDP 428 W.
- Spike (N)
- Asynchronous channel. Mean die 85 mm². Mean TDP 28 W.
- Qubit (Q)
- The 256-qubit ring on the classical die. Mean die 100 mm². Mean TDP 90 W.
- Edge (E)
- The 42 mm² part for the MicroServer. Mean TDP 5.5 W. Fan-less range.
Fifteen modules under rtl\common\, each pulled into every compile.
- light_trigger_core
- The entry latch. First common block loaded on every compile.
- color_math_alu
- The 64-bit arithmetic unit the cores call. Source of every width warning in the logs.
- quantum_coprocessor
- The classical-side shim to the qubit ring.
- photonic_link
- Nanophotonic backplane. The chassis-to-die link.
- voxel_cell
- The volumetric compute primitive.
- hbm_controller
- HBM memory controller.
- pcie_phy
- PCIe physical layer.
- exergy_compute_wave_mixer
- The exergy mixer tied into the compute wave.
- quantum_reactor_power_iface
- Interface to the quantum reactor’s power bus.
- synth_photosynthesis_power
- Synthetic-photosynthesis power rail.
- summ_engine_sealed_loop
- Sealed-loop summation engine.
- normal_resonance_pipeline
- Normal-resonance pipeline, the standing-wave stage.
- peltier_nano_magnetic_cool
- Peltier nano-magnetic cooling block.
- emic_push_pull_ring
- EM-IC push-pull ring — the family shared with Quantum Battery Seed Two.
- qbeam_wireless_chain
- qbeam wireless chain.
Twelve generations, the process node steps down, the price steps up.
The same ten cores land on four process nodes across the roadmap, APM01 through APM12. Mean die area and mean TDP hold steady across the generations (same cores, same die sizes); the process node fades down, and the list price per configuration climbs with it.
- APM01–03
- 7 nm. Mean die 146 mm². Mean TDP 170 W. Price range from twenty-two thousand five hundred to just under nineteen million per configuration.
- APM04–05
- 5 nm. Same die and TDP means. Prices climb: thirty-three thousand to thirty-seven million.
- APM06–08
- 3 nm (APM06 at 4 nm). Prices from sixty-three thousand to eighty-four million.
- APM09–12
- 2 nm. Prices from one hundred and eighty thousand up to three hundred and seventy-three million per configuration.
- Shelf total
- 5,040 configurations. Twelve generations by ten cores by about forty-two package / memory / TDP permutations.
- Low price
- $22,500. An APM01 Edge configuration at its smallest package and lowest memory.
- High price
- $373,352,380. An APM12 Train configuration at the top of the stack.
The specs sit where the industry sits at the same node.
The point of printing the spec numbers is to be measured against the real parts shipping today. Four honest comparisons, in the industry’s own terms:
- Train Core vs training accelerators
- 280 mm² at 499 W. Reference parts of the training class at a similar node run larger (Nvidia H100 is about 814 mm² at 700 W on 4 nm). Train Core aims smaller and cooler, in exchange for a tighter per-die workload.
- Scale Core vs server hosts
- 110 mm² at 73 W on 7 nm. Mainstream server hosts of the generation land in that range: four active cores, two quantum units, four photonic lanes, two PCIe lanes, eight EMIC coil pairs.
- Edge Core vs edge inference parts
- 42 mm² at 5.5 W. Fan-less industrial-edge parts (NXP i.MX8, Rockchip RK3588) live in that envelope. Edge Core is the MicroServer die, the only one meant to sit in a sealed box.
- Lingua and Infer vs LLM silicon
- 250 mm² / 428 W and 150 mm² / 169 W. Mid-range datacenter inference land. Smaller than the training die, hotter than the host, which is what a dedicated language-model part looks like.
- The seven measurements
- What the industry does not print and this testbench does: resonance_locked, loop_sealed, cool_in_setpoint, ring_balanced, reactor_mw > 0, chain_received > 0, members_all_ok. Guaranteed in Verilog before the die is drawn.
Three Python suites, one rule per suite, none of them yielded.
-
2A
Visibility. test_fantasy_gate_visibility.py.
For every URL in the Members category, the test issues an anonymous HTTP HEAD and expects a 302 pointing at /customer/account/login/. For every known-public control URL it expects 200. A 200 on a gated URL is a hard fail; a non-matching redirect is a hard fail; a network error is a hard fail. The runner also offers a sampled mode for spot-checks. The rule: anonymous eyes do not land on a Members product page. It held.
-
2B
Purchase. test_fantasy_gate_purchase.py.
Over SSH into the VPS, the test asks MySQL four things about each Members SKU. Price above zero — a silent zero is the exact way free RTL leaked before. Status 1 (enabled) or 2 (disabled) — any other integer is noise. Category 20 and only category 20 — any second category is a public leak. Type simple or downloadable — nothing else sells here. One line per SKU, pass or fail, the reasons named. The rule: nothing that is cheaper than it should be, nothing that leaks into the public catalog, nothing with the wrong sell model. It held.
-
2C
Download. test_fantasy_gate_download.py.
Three checks, one file. First, five raw probes against /pub/media/downloadable/ paths — the files, the folder, the alternate prefix. Each probe must return 403 or 404; a 200 is a leak. Second, an SSH grep against /etc/nginx/conf.d/magento.conf for the explicit location block that denies the folder. Third, for any Members product of type downloadable, a MySQL check on its downloadable_link: non-shareable and a finite download cap. The rule: the raw zip never leaves nginx without a signed path through PHP. It held.
Cryptography, format, authorization, output. One file, every rule.
tests/test_generator.py. The file prints [PASS] or [FAIL] per check and exits 0 only if every one passed. The checks sit in four groups.
- Cryptography
- Three RFC 8032 ed25519 reference vectors verify. Each one also rejects a changed message and a changed signature. Six refusals, three accepts.
- Format
- GDS real encoding of 1e-3 and 1e-9 round-trips to within 1e-15. The compiled-in public key equals the bytes in keys/fantasy_ed25519.pub.
- Authorization
- Run with no license file: exit 2. Run with a license signed by another key: exit 2. Run with an expired license: exit 2. Run with a license whose JSON order_id was changed by one byte: exit 2. Run --rtl mode against a package whose NODE_CONFIG.yaml names another SKU: exit 1. Nothing is written after any of these refusals.
- Output
- Outline build with a valid license writes CHIP_OUTLINE_ONLY.gds. The file name carries _OUTLINE_ONLY. A 5 mm by 3 mm die measures 5000 by 3000 microns inside the file. Layer 254 is present. The fingerprint on layer 254 reads back with order=77, customer=88, and sku=TEST-SKU.
- Format verifier
- Zero-padded tail accepted. File with its ENDLIB trimmed rejected. File with one record overwritten by 0xFF: rejected. Old v2.1 generator output rejected by the strict verifier.
The reader with the folder in hand can run every one of these on their own floor.
- Testbench
- From the project root: pwsh tests\run_members_tbs.ps1. The script prints one line per core and a pass / fail summary.
- Visibility
- python tests/test_fantasy_gate_visibility.py. Add --sample 50 to spot-check fifty random URLs. Exit 0 on all-pass, 1 on any failure.
- Purchase
- python tests/test_fantasy_gate_purchase.py. Needs SSH to the VPS. --skus A B C overrides the SKU list for a focused run.
- Download
- python tests/test_fantasy_gate_download.py. Also needs SSH for the nginx config read and the downloadable-link check.
- Generator
- python tests/test_generator.py. Pure Python, no network, no SSH. Every check prints its line. Exit 0 only if every one passed.
A quiet page of numbers.
- Seven
- Measurements every core testbench must pass — resonance locked, loop sealed, cool in setpoint, ring balanced, reactor_mw > 0, chain_received > 0, members_all_ok.
- Ten
- Cores in the set. Each one has its own testbench, its own .vvp, its own log.
- Fifteen
- Common blocks pulled into every compile.
- Twelve
- Generations, APM01 through APM12.
- Four
- Process nodes across the roadmap — 7 nm, 5 nm, 3 nm, 2 nm.
- 293 K
- Cooling setpoint the Peltier block must hold under load. 20 °C.
- Three
- Software gate suites run against the live store.
- Three
- RFC 8032 ed25519 reference vectors, each verified, each rejected on a changed message, each rejected on a changed signature.
- Five
- Raw-download probes a reader can fire at the store, each one expected to fail closed.
- Zero
- Wafers ordered. Zero probe cards placed. Zero bench instruments touched.
- 5,040
- Configurations behind the gate — twelve generations by ten cores by about forty-two package / memory / TDP permutations.
- One
- Folder.
Software testing software. Not silicon on a probe station.
These are software tests. Verilog compiled by a simulator, Python hits against a store, Python calls into a signer. The chip has not been printed. There is no wafer back. There is no current reading on the rails.
The logs carry warnings a careful reader should see. The color_math_alu expects 64-bit operands and several callers pass 16, 10, 9 or 17 bits; iverilog pads the high bits and the simulation runs. In silicon, the width agreement is settled before tape-out, not during. The warnings are the list.
The gate suites listed above were run against the Magento store. The store has since moved to Cloudflare Workers and the Members entrance is now a Stripe Identity flow with an HMAC cookie; the suites that cover that live on the Workers side and get their own day in print.
- What ran
- Icarus Verilog compile and vvp run for each of the ten cores.
- What ran
- Three gate suites against the live store.
- What ran
- One generator suite, ed25519 and all.
- What did not
- Fabrication. Probe. Bench measurement. Yield.
- Open items
- Port-width agreement between color_math_alu and the ten callers, before any tape-out.
- Next suite
- Stripe Identity flow on cri-one.com, adversarial review in flight.
The gate held. The generator refused. The ten cores compiled.
No chips were burned. No wafers were ordered. Software tested software, and today the software said the right thing. That is the whole headline. Tomorrow a port width gets settled, a test gets added, a shelf item ships; today it is enough that the suites ran green.