The design that works on talon today has no slack left. Adding the QSFP ring pushes it over the edge in two separate places: the GT transmit clock and the 225 MHz engine clock.
Result, 2026-10-17 12:47 UTC: the recovery image closed timing. v8 (0xE4B71262: C_COLS 2, 225 MHz, GT island, crumbc fanout fix) finished post-route optimisation with WNS +0.003 / WHS +0.010, 0 failing endpoints [M]. The worst clock is hbms_clk +0.003, which is +0.093 true after adding back the 0.090 ns setup margin (0x1254 was released at +0.002 / +0.092 the same way). The GT clocks txoutclk_3 +0.047 and txoutclk_2 +0.073 were −1.813 and −0.487. Finding F1299.
Silicon, 2026-10-17 ~18:45 UTC [M]: flashed to partition 1 on all 8 talon cards (FLASH_EXIT=0 ×8, device_boot OK ×8). QM-38 PASS. HBMS ladder 52/52 PASS on all 8 cards. Ring: 16/16 links up; discovery finds one 8-card cycle; 30 s BIST PASS on all 16 directions with 0 word errors, about 4.19 GB/s each. Finding F1300.
Diagnosis. 0x1254 s2, the image flashed on all 8 talon cards, met 225 MHz with +0.002 ns of slack and contains no ring hardware. 0x1256 adds the talon ring core, 8 GT links and the HBMS all-reduce bridge. That logic lands mostly in SLR0, which goes from 89.8 % to 94.2 % SLICE, and the result fails by −1.813 ns [M].
The two failures need different fixes. The GT failure is a floorplan error and has a clean fix. The engine-clock failure is caused by how full the device is; fanout fixes help, and lowering the clock to 200 MHz is the sure fix.
Final routed timing per clock. Setup slack must be at least +0.03 ns, and hold slack at least +0.010 ns, to release a PDI.
| Clock | Domain | 0x1254 s2 (on talon) | 0x1256 s1r | R8_BEST s1c |
|---|---|---|---|---|
| hbms_clk | engine, 225 MHz | +0.002 | −0.492 · 61,366 ep | −0.631 · 68,940 ep |
| hbms_mclk | HBM NMU side, 400 MHz | not present | not present | −1.210 · 97,123 ep |
| ch0_txoutclk_3 | QSFP GT transmit | not present | −1.813 · 3,475 ep | −1.766 · 2,689 ep |
| ch0_txoutclk_2 | QSFP GT transmit | not present | −0.487 · 2,069 ep | −0.536 · 871 ep |
| ch0_rxoutclk_2 / _3 | QSFP GT receive | not present | +0.108 / +0.058 | +0.158 / +0.154 |
| talon_clk_wiz | ring core, 150 MHz | not present | +0.060 | +0.035 |
| clk_pl_2 | shell | +0.120 | +0.023 | +0.003 |
| Overall | WNS +0.002 · WHS +0.010 | WNS −1.813 | WNS −1.766 |
SLICE utilisation per SLR after placement. Above about 85 %, routing detours grow quickly. The known-good image was already at 96.9 % in SLR1, which is why it had only 0.002 ns to spare.
| What the ring added (SLR0 unless noted) | 0x1254 s2 | 0x1256 s1r | Change |
|---|---|---|---|
| SLICE used | 101,097 | 105,987 | +4,890 |
| CLB LUTs | 587,666 | 615,056 | +27,390 |
| LUT as memory (ring FIFOs) | 8,268 | 12,288 | +4,020 |
| Block RAM tiles | 120 | 175 | +55 |
| SLR1 ↔ SLR0 crossing wires used | 6,482 (34%) | 9,442 (50%) | +2,960 |
| Ring IPs in synthesis (talon core; 2 Aurora links) | none | 30,323 + 2 × 10,355 LUT | +51k LUT |
The QSFP GT quads sit in SLR1. The hook put the whole talon core in SLR0, including each port's transmit FIFO and the first synchroniser stages, which run on the GT clock. A GT-clock path then crosses the SLR boundary twice inside a 2.48 ns period (402.8 MHz). That cannot close.
Worst path −1.813 ns on ch0_txoutclk_3; the same class in R8_BEST s1c is −1.766 [M].
The FIFO's read side runs on the GT clock, yet its LUTRAM and read pointer sit in the far SLR. 4.6 ns of wire inside a 2.48 ns period.
A new pblock in SLR1, next to the quads, holds every GT-clocked register, both async FIFOs, every synchroniser and the first capture flop. The FIFO becomes the only SLR crossing, on its 150 MHz side (6.67 ns).
checked on the netlist 12,428 of 12,428 island cells placed in SLR1; 630 TX and 1,258 RX FIFO RAM cells [M]. Not yet timed in a routed build.
The hbms_clk paths fail by about 0.49 ns. Almost all of the delay is routing, not logic. Many are single registers driving thousands of loads across a crowded region.
In e8_crumbc_pd, the registers v1, f1 and l2 each drove all 128 rows × 4 columns of 27-bit accumulators (about 14k loads). Each is now replicated once per 8-row group, with the same latency and the same values.
xsim equivalence 200,000 random cycles, 0 mismatches; a deliberately broken copy was caught [M].
limit Only the u_ceng class is addressed. The tile broadcast and VPU paths are still open.
A 5.000 ns period instead of 4.444 ns adds 0.556 ns, which turns the measured −0.492 into about +0.06 ns [D]. That clears the +0.03 gate on its own.
trade-off Engine and HBM-side throughput drop by about 11 % [D]. For a fallback ring image, a working PDI is worth more than the extra 25 MHz. Your call which one to flash.
Full netlist dump of 0x1256 s1r (one row per failing endpoint). Source: council H2, council H1.
| Engine clock class (61,366 ep) | Endpoints | Fix |
|---|---|---|
| crumbc f1/l2/v1 fanout | 23,416 | v5 replication ✓ |
| crumbc product broadcast | 6,844 | v7 replication |
| ATTN-W (SLR1) | 14,146 | v7 registers |
| VPU (SLR1) | 5,190 | v7 registers + move to SLR0 |
| other (attn, leng, t1x, cab, ring, axi) | ≈ 11,770 | density (C_COLS 2) |
| crossing an SLR boundary | 512 | not the problem |
| GT clock class (5,544 ep) | Share | Fix |
|---|---|---|
| link_rst_n as clock-enable into PRBS test registers | 45 % | v9b: register + replicate the reset (BIST kept) |
| TX/RX CDC FIFOs | 29 % | island + FIFOs in BRAM |
| Aurora IP internals | 14 % | density in SLR1 |
Failing paths are 87–98 % routing delay on 0–1 logic levels and almost all stay inside one SLR: SLR0 28.5k, SLR1 24.2k, SLR2 8.1k. This is a density signature. The known-good 0x1254 needed a placement estimate of +0.144 ns to route at +0.002, so each new build must place at about +0.1 or better.
| Build | Contents | Host | Expected result | PDI ETA (UTC) |
|---|---|---|---|---|
| 0x1256 v5 s1/s2/s3 | GT island + crumbc fanout fix; three directive sets | olegvm3, olegvm2, midyard | near 0, unproven [P] | 17:30–19:00 |
| 0x1261 v6 | v5 at 200 MHz | olegvm3 (queued) | ≈ +0.06 on hbms_clk [D] | ≈ 22:00 |
| 0x1256 v4 s1 | GT island only | olegvm3 | ≈ −0.49 [P] | ≈ 18:00 |
| R8_BEST BD5 | A + B RTL, crumbc fix, floorplan D island | olegvm2 (queued) | unknown | ≈ 22:00 |
| R8_BEST kit_B / kit_BM | older fix sets | olegvm3 / olegvm2 | −0.76 / −1.87 [M, mid-route] | diagnostic only |
✓ means the fix is in that build. Options 4–12 are being tried now by parallel agents (P = pipeline registers, R = no-cost trims, T = trade-offs). Agent results open in the council report browser.
Decision (user, 2026-10-17): C_COLS 2 at 225 MHz. Agent T found every talon serving result is W8, which does not use the W2 crumbc columns. C_COLS 4 → 2 frees about 150k LUT in each of SLR0 and SLR2 at 0 W8 tok/s cost [D]. 200 MHz would cost about 11 % on every model. The recovery image is now v8 (BUILD_ID 0xE4B71262): GT island, crumbc fanout fix, C_COLS 2, 225 MHz. v7 adds the pipeline registers and no-cost trims on top. v6 (200 MHz) stays as the safety net.
| # | Option | Cost | Agent | v4 | v5 | v6 | BD5 |
|---|---|---|---|---|---|---|---|
| 1 | Pipeline registers on the remaining engine-clock paths (tile broadcast, VPU tk2, ATTN-W, kvnet) | +1–2 cycles latency | P | – | – | – | ✓ |
| 2 | crumbc fanout replication | none, bit-exact | done | – | ✓ | ✓ | ✓ |
| 3 | SerDes island next to the GT quads | none | done | ✓ | ✓ | ✓ | ✓ |
| 4 | Ring FIFOs from LUTRAM to block RAM | none | R | – | – | – | – |
| 5 | Compile out ring BIST and debug | no on-card self-test unless strapped | R | – | – | – | – |
| 6 | Retiming and fanout options in synthesis and phys_opt | runtime | R | – | – | – | – |
| 7 | Rebalance SLRs for the talon core and hub blocks | none | R | part | part | part | part |
| 8 | 400 MHz HBM-side registers (reset tree, read data, CE) | none | done | – | – | – | ✓ |
| 9 | Engine clock 200 MHz | ≈ −11 % engine throughput | T | – | – | ✓ | – |
| 10 | crumbc C_COLS 4 → 3 or 2 (≈ 0.6M LUT at C=4) | slower W2 / spec verify | T | – | – | – | – |
| 11 | Narrower accumulators | must stay bit-exact | T | – | – | – | – |
| 12 | Fewer engine ports | loses HBM bandwidth | T | – | – | – | – |