fpga-archive-schematics-system · EAGLE4 FPGA archive
talon2 QSFP connection map (2026-11-10 07:3xZ, box1-qwen-junior)
What is plugged where on talon2's 7 V80 cards, for designing a NIC that connects to them. Sources:
- slots:
dmidecodeslot-to-BDF map (11-10); - cables:
reports/results/SEVENCARD_BOARD_20261026/talon2_cabling_20261110.html:55-67(live pagetalon2-cabling-map); - QSFP presence:
qsfp_id2.shread 07:0xZ (module id byte 0x11 = QSFP28-class at every probed cage); - link layer:
reports/QSFP_LINK_CAMPAIGN_INDEX.md:75-90.
1. Slots (left to right; the H200 is the leftmost slot)
| PCIe slot | BDF | card | role (7-card plan) |
|---|---|---|---|
| PCIE8 | cf:00.0 | H200 | GPU verifier (no QSFP) |
| PCIE7 | e1:00.0 | V80 "e1" | attention rank 3 |
| PCIE6 | 99:00.0 | V80 "99" | 7th card (planned prefill / 3rd MLP) |
| PCIE5 | bd:00.0 | V80 "bd" | attention rank 2 |
| PCIE4 | 17:00.0 | V80 "17" | attention rank 0 |
| PCIE3 | 3d:00.0 | V80 "3d" | attention rank 1 |
| PCIE2 | 50:00.0 | V80 "50" | MLP card 1 (MLPV0) |
| PCIE1 | 63:00.0 | V80 "63" | MLP card 2 (MLPV1) |
Each V80 has 4 QSFP cages, numbered c0..c3 as the card's AMC/ami_tool numbers them; all 4 cages sit in the GTM quads of SLR1.
2. Cables (13 point-to-point QSFP cables, card.cage <-> card.cage)
| # | end A | end B | use |
|---|---|---|---|
| 1 | 17.c0 | 3d.c3 | attention ring |
| 2 | 17.c3 | bd.c0 | attention ring |
| 3 | bd.c2 | e1.c2 | attention ring |
| 4 | 3d.c1 | e1.c1 | attention hypercube (NEW 11-10, user-plugged) |
| 5 | 17.c2 | 50.c2 | rank 0 -> MLP |
| 6 | 3d.c2 | 63.c1 | rank 1 -> MLP |
| 7 | 63.c0 | e1.c3 | rank 3 -> MLP |
| 8 | bd.c1 | 50.c1 | rank 2 -> MLP (NEW 11-10, user-plugged) |
| 9 | 50.c0 | 63.c3 | MLP <-> MLP |
| 10 | 17.c1 | 63.c2 | 4-card link |
| 11 | 3d.c0 | 50.c3 | 4-card link |
| 12 | 99.c0 | bd.c3 | 99 <-> bd |
| 13 | 99.c3 | e1.c0 | 99 <-> e1 |
3. Cage occupancy per card
| card | c0 | c1 | c2 | c3 | free cages |
|---|---|---|---|---|---|
| 17 | 3d.c3 | 63.c2 | 50.c2 | bd.c0 | none |
| 3d | 50.c3 | e1.c1 | 63.c1 | 17.c0 | none |
| bd | 17.c3 | 50.c1 | e1.c2 | 99.c0 | none |
| e1 | 99.c3 | 3d.c1 | bd.c2 | 63.c0 | none |
| 50 | 63.c3 | bd.c1 | 17.c2 | 3d.c0 | none |
| 63 | e1.c3 | 3d.c2 | 17.c1 | 50.c0 | none |
| 99 | bd.c3 | — | — | e1.c0 | c1, c2 |
For a NIC: the ONLY free QSFP cages on talon2 are 99.c1 and 99.c2 (card 99, slot PCIE6). Every cage on the six drafter cards is in use. Adding a NIC on any drafter card means unplugging one of the cables above. Rank links (1-8) and MLP links (5-9) carry the drafter; the 4-card links (10, 11) are spare once the 7-card image runs.
4. Link layer (what a NIC must speak to talk to a V80 cage)
- PHY: 4 lanes x 25.78125 Gb/s per cage (100G-class, NRZ), GTM transceivers, refclk 322.265625 MHz (QSFP_LINK_CAMPAIGN_INDEX.md:89-90; pin file link_mcio0.xdc).
- Protocol: Xilinx Aurora 64B/66B (link BD
src/bd/create_bd_design_link.tcl). This is NOT Ethernet: a standard Ethernet NIC will not link to these cages unless that NIC (or an FPGA NIC) runs Aurora 64B/66B at the same rate. - On-wire framing used by the drafter: tp_link4 engines e0/e1/e2 bind cages by beacon. Registers: CTRL 0x008 = card_id[7:0], beacon_en[8], role[23:16]; PEER[q] at 0x010+4q = card[7:0], cage[11:8], role[23:16], seen[28] (
scripts/4c1/c4_tp4.py:10). A NIC wanting to be "seen" must send that beacon frame; otherwise the engines refuse to bind (c4_tp4.py:148-155). - Module: QSFP id byte 0x11 present at every probed cage. The vendor-name page read errors on all of them ("could not read data"), so the cable/DAC part number is NOT recorded.
5. What is NOT measured yet
- Link up / error counters / per-cage latency: NOT measured this session (B4 flashing on 17/3d/bd/e1 since 07:07Z). The beacon check of cables 4 and 8 and the PEER-register dump run right after the next cold AC; this file gets a "live link" table then.
- Physical left-right order of the cages on the bracket: NOT recorded; read the bracket labels.