fpga-archive-reports-qwen · EAGLE4 FPGA archive

Source reports/results/QWEN_GPTQ2_20261116/SURVEY_PRECISION_AREA_20261116.md · source md5 5f6bc713fc50113f457daa70e0be31b4 · archived 2026-11-18

Download PDFOpen PDFRaw markdown

SURVEY: bit precision vs area, record 7-card images vs the Qwen GPTQ-2 (Q5120) images, and the room left to enlarge the 2-bit GEMM (2026-11-16)

Program: the 7-card TARGET-side 4C1 drafter (ATT x4 TP4 + MLPV x2, ONE drafter; Qwen3.8-27B target, drafter = Qwen EAGLE-4 GPTQ v2 2-bit, H 5120). Motto: a research track never ends by denying the idea; a wall is a measured constraint to engineer around, toward the fastest flashable PDI. Agent: survey census for box1-qwen-junior. User question (verbatim, 2026-11-16): "others did get larger... so I want to see if we can use the extra space to enlrge the gemm for 2 bit... can you do a survay on the current bit preicions - area trade versus the new image". USER RULING relayed by the coordinator (2026-11-16): activations are A16 on the Qwen GPTQ-2 images, for accuracy. Every headroom row in §1 is priced at A16. Rules kept: read-only on every fork and build; nothing killed; nothing deleted; no build launched; Vivado runs only on my VMs (buildx04 / buildxd00 / buildy00); bulk scratch /mnt/hdd_data/eagle4_scratch/qwen_gptq2_20261116/survey/.

Layout: §0 answers; §1 holds every calculation, projection and hypothesis; §2 to §6 hold only measured values with their sources.

0. Answers

  1. Binding resource per card (measured, placed).
    • Record ATT 0x13BE: SLICE in SLR0, at 98.12 %. SLR0 also carries DSP at 89.0 %.
    • Record MLPV 0x1361 (M1361 final): SLICE in SLR2 (94.55 %) and SLR1 (93.44 %). Next come SLR1 DSP at 84.5 % and SLL SLR1<->SLR0 at 78.2 %.
    • Q5120 ATT 0x1346 / 0x134A: SLICE in SLR1, at 93.3 %. SLR0 is almost empty: SLICE 10 %, LUT 3.8 %, URAM 0, DSP 0.5 %. The device as a whole is at SLICE 55 %, LUT 26 %.
    • Q5120 MLPV 0x1349: SLICE everywhere: device 94.45 %, SLR 92.8 / 93.4 / 97.2 %.
    • 0x135E: close to the same netlist as 0x1349 (synth u_array 312,320 vs 311,050 LUT), placed with AggressiveExplore. It measures SLICE 78.8 % device / 90.8 % SLR1. So SLICE is partly a placement outcome; LUT is 53 % on both.
  2. Where the bits went (§3 to §4).
    • The 2-bit weights stayed in URAM at 16 codes per 32-bit word on every image.
    • The MLPV GEMV changed precision class. The record MLPV computes 2-bit x A16 on DSP58: E3 packed dot, 2 products per DSP58, 5,120 DSP for 256 lanes. The Q5120 MLPV computes 2-bit x A8 in LUTs: LUT-GEMM lane with a per-column zero bank, 311K LUT and 1,889 DSP for 256 lanes.
    • The MLPV URAM grew 512 -> 1,280 because the Q5120 v0 split put G7 on the MLPV, with 2.6x larger weights.
    • G8 went from fp16 rows (788 MB head) to 2-bit rows (340 MB). The synth u_g8 still grew 314K -> 358K LUT, because H 5120 and 248K vocab offset the measured -29K LUT of the 2-bit dot.
    • The fp16 seams grew 2-4x (H 3072 -> 5120 plus the G7 hand-off).
  3. Headroom to enlarge the 2-bit GEMM at A16 (§1.3):
    • ATT card: SLR0 is the space, with about 85K SLICE, 3,965 DSP and 677 URAM free after the base. At A16, neither lane form alone reaches the v0 L512: the LUT lane gives 480 lanes (340 at the 2.1x price) and the DSP58 E3 lane gives 401. A per-die hybrid gives 631 lanes (556 at 2.1x): FITS. The largest products per cycle is LUT K64/K128 at x1.49: 22K to 29K.
    • MLPV card: at A16, only the DSP58 E3 lane fits L256 (382 lanes max). The LUT lane at A16 reaches only 215 (151 at 2.1x). The largest products per cycle is E3 K64, 15,040, DSP-bound.
    • SLL is not the wall for the array on either card (ATT 17.6 / 11.8 % used; MLPV 65.8 % used, about 6.4K free per cut).
  4. The URAM read width is the free K lever.
    • What a URAM gives: a URAM288 is 4,096 x 72 with 2 ports, 144 b per cycle. The shipped bank packs two 32-bit (16-code) words per 72-bit row and reads 32 codes per lane-cycle, about 44 % of what the URAM offers.
    • K64: a 64-bit entry at half the depth needs no extra URAM. It is DSP-free on the LUT lane; on the DSP lane it costs +16 DSP per lane.
    • K128: needs the lane's 2 URAMs read side by side. Both v0 cards have 2 per lane.
    • URAM per card is set by the weight bits, not by K: about 730 minimum for the ATT and about 442 for the MLPV.
  5. Instrument limit: the SMBus licence served to the VMs (2101) currently returns FlexNet -95,378 Connection reset by peer ([Netlist 29-57] on axi_smbus_rpu), so no routed full-chip dcp could be opened today. Two consequences:
    • Per-SLR numbers come from each build's own placed report (and from the M1361 routed -slr report).
    • Per-instance numbers come from report_utilization -hierarchical on each build's drafter IP OOC synth dcp. That dcp holds only our RTL and needs no licence. Its counts are pre-opt: they include logic that opt_design later prunes, such as the legacy fp16 dot in the Q5120 u_g8.
    • No per-instance per-SLR split was measured.

1. Calculations and hypothesis (every derived number in this document)

1.1 Per-SLR capacity used below [D on the place reports]

The capacity of each SLR is computed as used / % from Q46_place_report_utilization_0.rpt:393-412, cross-checked against the used + free sums in SYNTH_CENSUS_WEEK_20261105_12.md:490-510.

SLR SLICE LUT DSP58 URAM BRAM tiles
SLR0 112,527 900,528 3,984 677 1,341
SLR1 102,530 820,161 3,456 624 1,200
SLR2 106,719 853,519 3,408 624 1,200

1.2 Per-lane prices (inputs are the measured rows of §5; the scaling is mine)

A8 LUT lane, in the image: 277 SLICE, 1,082 LUT, 6.0 DSP, 1.5 BRAM tiles, 2 URAM per lane.

A16 LUT lane = the A8 price x 1.49, or x 2.1.

DSP58 E3 lane at A16, 2 products per DSP58: about 206 SLICE, 20 DSP58, 1.6 BRAM tiles, 2 URAM per lane.

K scaling.

Non-array bases.

1.3 Headroom table at A16: v0 geometry (ATT LANES 512, MLPV LANES 256), per card and per SLR

Each SLR is listed as SLR0 / SLR1 / SLR2. Lane counts are at 100 % of each die, with the 90 %-SLICE practical bar in brackets. The binding resource for each row is given in words.

Free after the base

card (base) SLICE free DSP free URAM free
ATT (0x1346 + G7 delta) 85.2K / 6.8K / 106.7K 3,965 / 979 / 3,408 677 / 479 / 624
MLPV (0x1349 minus array) 36.3K / 21.5K / 30.9K 3,966 / 1,161 / 3,354 677 / 624 / 624
MLPV (0x1361 minus array) 50.3K / 13.9K / 22.1K 3,933 / 1,237 / 3,294 677 / 624 / 624

ATT card, v0 needs 512 lanes

lane form max lanes per SLR total [90 %] binding v0 fits? products per cycle at max
LUT A8 K32 (shipped, reference) 307 / 24 / 312 643 [579] SLICE in SLR0/SLR1, URAM in SLR2 yes 20,576
LUT A16 K32 x1.49 206 / 16 / 258 480 [411] SLICE no 15,360
LUT A16 K32 x2.1 146 / 11 / 183 340 [292] SLICE no 10,880
LUT A16 K64 x1.49 149 / 12 / 187 348 [298] SLICE no 22,272
LUT A16 K128 x1.49 96 / 7 / 121 224 [191] SLICE no 28,672
DSP E3 A16 K32 198 / 33 / 170 401 [368] DSP in SLR0/SLR2, SLICE in SLR1 no 12,832
DSP E3 A16 K64 110 / 26 / 94 230 [204] DSP no 14,720
HYBRID A16 K32 (x1.49) (127 LUT + 159 E3) / (0 + 33) / (205 + 107) 631 SLICE + DSP together yes 20,192
HYBRID A16 K32 (x2.1) (85 + 172) / (0 + 33) / (138 + 128) 556 SLICE + DSP together yes 17,792

MLPV card, v0 needs 256 lanes

lane form max lanes per SLR total [90 %] binding v0 fits? products per cycle at max
LUT A8 K32 (shipped, reference) 131 / 77 / 111 319 [203] SLICE at 100 % only 10,208
LUT A16 K32 x1.49 88 / 52 / 75 215 [136] SLICE no 6,880
LUT A16 K32 x2.1 62 / 36 / 53 151 [96] SLICE no 4,832
LUT A16 K128 x1.49 41 / 24 / 35 100 [62] SLICE no 12,800
DSP E3 A16 K32 175 / 58 / 149 382 [273] SLICE in SLR0/SLR2, DSP in SLR1 yes 12,224
DSP E3 A16 K64 110 / 32 / 93 235 [207] DSP no (at 256 lanes) 15,040
HYBRID A16 K32 (x1.49) (0 + 175) / (28 + 48) / (0 + 149) 400 SLICE + DSP together yes 12,800
record base, DSP E3 A16 K32 196 / 61 / 107 364 [261] DSP in SLR0/SLR1, SLICE in SLR2 yes 11,648

What the table says:

1.4 URAM and K (the read-width fact) [D on R + M]

1.5 Precision ledger bytes [D on R]

Weight bytes = N·K / 4 for 2-bit codes, N·K / 64 for int16 scales per 128-group, N·K / 512 for 2-bit zeros. Stage shapes come from ARRAY.md §4 (record) and QWEN5120_P1_20261108.md §1.2 (Q5120).

card codes (URAM) scales int16 (BRAM) zeros 2-bit (BRAM18)
record ATT 0x13BE (G4 3,072x6,144 + G6 3,072x768 + G7g/u 2,048x3,072) 8.45 MB (= 4,128 words x 512 lanes x 4 B, ARRAY.md:117) 0.53 MB 66 KB
record MLPV 0x1361 (G9 6,144x4,096 + G12 64,128x256) 10.4 MB (10,160 words x 256 x 4 B) 0.65 MB none (zero broadcast per group)
Q5120 ATT 0x1346 (G4 3,840x10,240 + G6 5,120x1,280) 11.5 MB (11,200 entries per lane x 256) 0.72 MB 90 KB
Q5120 MLPV 0x1349 (G7g + G7u 8,192x5,120 + G9 5,120x8,192 + G12 124,160x128) 35.4 MB 2.2 MB 0.28 MB
v0 ATT (TP4 v1 split) 23.9 MB 1.5 MB 0.19 MB
v0 MLPV 14.5 MB 0.90 MB 0.11 MB

1.6 Hypotheses


2. Device and per-SLR utilisation per image (measured, placed / routed)

Each per-SLR column reads SLR0 / SLR1 / SLR2. SLL figures are used / 18,870 per cut. The V80 totals are LUT 2,574,208, FF 5,148,416, DSP58 10,848, BRAM36 tiles 3,741, URAM 1,925, SLICE 321,776.

image (role, build host, report, md5) SLICE LUT FF DSP58 BRAM tiles URAM SLL 1<->0 / 2<->1 binding
0x13BE record ATT HCUT, Llama H3072, L512 LUTG A8; buildx04 s3LAtp4BE placed, BE_place_report_utilization_0.rpt 03fdd895 263,389 (81.85 %); 110,410 / 70,094 / 82,885 = 98.12 / 68.37 / 77.67 % 1,023,252 (39.75 %); 45.69 / 31.00 / 41.90 % 932,324 (18.11 %) 5,267 (48.55 %); 3,546 / 694 / 1,027 = 89.0 / 20.1 / 30.1 % 1,783 (47.66 %) 744 (38.65 %); 232 / 171 / 341 9,501 (50.4 %) / 1,880 (10.0 %) SLICE in SLR0
0x1361 record MLPV G8_OWN, L256 E3 A16 (M1361 routed -slr, M1361_final_slr_util.rpt edfa7302; device totals from build placed M61build_… bf16ecff) 288,227 per-SLR sum (89.6 %); 91,520 / 95,807 / 100,900 = 81.33 / 93.44 / 94.55 % 1,185,094; 35.70 / 50.31 / 52.83 % 1,018,982 (19.8 %) 7,504 (69.17 %); 2,891 / 2,919 / 1,694 = 72.6 / 84.5 / 49.7 % 1,035.5 (27.7 %) 512 (26.6 %); 284 / 70 / 158 14,764 (78.2 %) / 12,533 (66.4 %) SLICE in SLR2/SLR1, then SLR1 DSP, then SLL 1<->0
0x1346 Q5120 ATT TP4, L256 LUTG A8, 10x128 heads HP5/HPS2; buildx04 ~/build4c1/s3LAq5120t4c placed, Q46_place_… 0670e6d2; route WNS -2.565 (failed setup: area only) 177,862 (55.28 %); 11,306 / 95,700 / 70,856 = 10.05 / 93.34 / 66.39 % 661,503 (25.70 %); 3.79 / 42.73 / 32.44 % 693,438 (13.47 %) 4,036 (37.21 %); 19 / 2,477 / 1,540 1,367.5 (36.55 %); 24 / 959.5 / 384 657 (34.13 %); 0 / 145 / 512 3,328 (17.6 %) / 2,233 (11.8 %) SLICE in SLR1 (SLR0 empty)
0x134A = 0x1346 + G6_EARLY_ISSUE + KVF_TFIX, stopped copy on buildx00 ~/build4c1/s3LAq5120t4k placed, Q4A_place_… e40e6d2e (closed copy +0.047/+0.011 is on buildxd05, not opened) 177,186 (55.07 %); 9.81 / 93.06 / 66.28 % 658,621 (25.59 %) 693,717 (13.47 %) 4,036 (37.21 %) 1,367.5 657; 0 / 145 / 512 3,243 / 2,277 SLICE in SLR1
0x1349 Q5120 MLPV L256 LUTG A8, TREE_P 32, G8_W2; buildx00 s3LCq5120m49 placed, Q49_place_… 5a21d27e 303,917 (94.45 %); 104,451 / 95,738 / 103,728 = 92.82 / 93.38 / 97.20 % 1,376,072 (53.46 %); 46.65 / 59.00 / 55.31 % 1,285,443 (24.97 %) 4,256 (39.23 %); 771 / 2,686 / 799 2,271.5 (60.72 %); 708 / 953.5 / 610 1,280 (66.49 %); 510 / 265 / 505 12,413 (65.8 %) / 12,420 (65.8 %) SLICE, every SLR
0x135E Q5120 MLPV (+A8Q_INREG; place AggressiveExplore); buildy03 s3LCq5120m5E placed, Q5E_place_… 2f61df52; route +0.013 (CONFIG_MANIFEST_135E.md) 253,589 (78.81 %); 72.37 / 90.78 / 74.10 % 1,367,658 (53.13 %) 1,286,295 (24.98 %) 4,256 (39.23 %); 738 / 2,875 / 643 2,271.5 1,280; 500 / 350 / 430 12,559 (66.6 %) / 11,011 (58.4 %) SLICE in SLR1
0x135D Q5120 MLPV L256/TP64/THARD + A8Q_INREG; buildxd05 (not opened: box1-two-v2 partition) 96.30 % / worst SLR 99.58 % 64.39 % 39.23 % 60.72 % 66.49 % 13,269 / 12,011 SLICE; route -2.862

The 0x135D row is the census line SYNTH_CENSUS_WEEK_20261105_12.md:269.

3. Per-instance hierarchy (measured, drafter IP OOC synth dcp, report_utilization -hierarchical, pre-opt)

Each row reads LUT / FF / RAMB36 / RAMB18 / URAM / DSP. The line number is that of the instance row in the image's *_synth_hier.rpt, and the md5 of each report is in the source line at the end of this section. Each dcp was copied from .../prj.runs/top_eagle4_drafter_lb_0_0_synth_1/; the copy md5s are in survey_evidence/*_ooc_dcp.md5.txt.

block record ATT 0x13BE Q5120 ATT 0x1346 record MLPV 0x1361 Q5120 MLPV 0x1349
drafter total 1,046,766 / 940,325 / 1,362 / 610 / 744 / 5,459 (:25) 645,796 / 653,474 / 1,091 / 321 / 657 / 4,100 (:25) 1,574,493 / 935,306 / 931 / 26 / 512 / 7,602 (:25) 1,465,370 / 1,230,084 / 2,187 / 26 / 1,280 / 4,419 (:25)
GEMV array u_array 597,623 / 466,864 / 512 / 512 / 512 / 1,635 (:32) 311,016 / 241,632 / 256 / 256 / 512 / 1,633 (:30) 147,305 / 137,895 / 280 / 0 / 512 / 5,120 (:31) 311,050 / 236,414 / 1,536 / 0 / 1,280 / 1,889 (:30)
array lanes 512 x lutgemm_lane_w2 988 / 628 / DSP 1 (:39) 256 x 986 / 627 / DSP 1 (:37) 256 x E3 lane 348 / 366 / DSP 1 (:37) + 256 x pair dot 125 / 53 / DSP 17 (:33,:35) 256 x 988 / 628 / DSP 1 (:37); u_bank 54 LUT / 5 URAM / 2 DSP per lane (:41)
A8 quantizer u_a8q 5,356 / 707 / DSP 96 (:34) 5,282 / 715 / DSP 96 (:32) none (A16) 5,260 / 704 / DSP 96 (:32)
attention u_attn_full 164,151 / 169,464 / 672 / 64 / 232 / 3,416 (:2181); 8 engines x 10,974 LUT / 395 DSP 127,275 / 123,022 / 673 / 47 / 145 / 2,221 (:1092); 5 engines x 11,470 / 393 DSP none none
KV u_kv (K URAM, V BRAM, 16-bit) 10,392 LUT / 448 RAMB36 / 64 RAMB18 / 232 URAM (:2319) 6,326 / 560 / 40 / 145 URAM (:1179) none none
RoPE u_rope 5,818 / 216 RAMB36 / 256 DSP (:2321) 2,177 / 108 RAMB36 / 256 DSP (:1181) none none
G6 TP exchange u_tp (48-b partials) 104,881 / 45,059 / 11 RAMB36 / 32 DSP, LUTRAM 14,056 (:29) 50,679 / 21,879 / 11, LUTRAM 7,036 (:29) none none
G8 exact rescore u_g8 none none 314,123 / 71,687 / DSP 84 (:821); 4 fp16 dots x 22,751 LUT / 16 DSP (:825) 358,081 / 77,246 / DSP 100, LUTRAM 10,688 (:1337); 4 W2 dots x 23,634 LUT / 4 DSP (:1341), plus 4 legacy fp16 dots x 23,429 / 16 DSP still present pre-opt (:1342)
GLM g_glm.u_lmdown (bf16 down-proj x 16-bit act) none none 30,240 / 79,439 / DSP 2,048 (:29) 30,543 / 79,436 / DSP 2,048 (:29)
TREE u_tree / u_cheap_ser / u_tree_tiler none none 75,099 / 46,280 (LUTRAM 15,360) / 13,743 (:854,:852,:853) 75,099 / 46,292 / 13,800 (:1427,:1425,:1426)
tree DMA / out-walk / bridge none none 45,675 / 13,523 / 21,445 (:808,:810,:807) same values (:1324,:1326,:1323)
fp16/int16 activation seams (FF) r4g / r4u 3,697 / 3,868 (SDP BRAM 4 each); g5g6 657 / 8 RAMB36 r1_e 796 / 8 RAMB36; g5g6 7,439 / 8 RAMB36 r1_e / r1_h 17,460 / 17,769, 50,182 FF each; r5 23,075 / 66,824; r8 17,308 / 49,427; r6 16,397 / 49,161 r1_e / r1_h 56,615 / 56,320, 83,512 FF each; r5 89,822 / 133,436; r8 28,700 / 82,197; r6 27,406 / 81,930
RMSNorm (each of g1, g1_e, premlp, g11, g1_e_d) ~2,100-2,450 / ~975 / DSP 48 ~2,060-2,340 / ~980 / DSP 48 ~2,050-2,250 / ~970 / DSP 48 ~2,040-2,360 / ~975 / DSP 48
SiLU u_silu 3,010 / 545 / DSP 32 none (G7 not on this ATT) none 3,010 / 545 / DSP 32
u_ctrl 28,926 n/a (< 10K row) 48,100 50,840
drafter own logic (inst) 68,776 / 202,087 102,170 / 227,659 618,175 / 163,190 126,842 / 171,957

Sources:

Notes:

4. Precision ledger (measured placement of each precision class; bytes in §1.5)

4.1 Storage

precision class record ATT 0x13BE record MLPV 0x1361 Q5120 ATT 0x1346 Q5120 MLPV 0x1349
2-bit weight codes, 16 per 32-b word URAM 512, 1 per lane, DEPTH 8,192 (ARRAY.md:117; u_array :32) URAM 512, 2 per lane, DEPTH 16,384 (:117; u_array :31) URAM 512, 2 per lane (u_array :30; all in SLR2, placed :411) URAM 1,280, 5 per lane (u_bank :41)
GPTQ scales (Q3.13, int16; Qwen F 16, GEMV_SCALE_FRAC 16) RAMB36 512, 1 per lane (GEMV_SCALE_FRAC 13, verilog/clean7c/ref/att_13be/CONFIG_MANIFEST.md:111) RAMB36 256 (+ 24 LCACHE activation BRAM = 280) RAMB36 256 RAMB36 1,536 (6 per lane: scale + zero banks for 5 URAM of weights)
GPTQ-v2 2-bit zero RAMB18 512, per (lane, group) none: one zero per group, broadcast (qz_raw, ARRAY.md:119) RAMB18 256 within the 1,536 RAMB36
activations inside the GEMV int8 (A8Q), FF per 8-lane cluster int16 (A16), BRAM per-die cache, then FF int8 (A8Q) int8 (A8Q)
fp16/int16 activation seams SDP BRAM (SEAM_BRAM_W 1) and FF: small (§3 seam row) FF: r1 / r5 / r6 / r8 ~92K LUT, ~266K FF SDP BRAM, small FF: ~259K LUT, ~465K FF
KV cache fp16 K URAM 232 (8 heads x 96), V 448 RAMB36 + 64 RAMB18 (u_kv :2319) none K URAM 145 (10 heads x 128, HPS 2), V 560 RAMB36 + 40 RAMB18 (:1179) none
RoPE tables 216 RAMB36 none 108 RAMB36 none
norm gamma inside rmsnorm / drafter own logic: no separate RAM row same same same
cheap lm-head (G12) 2-bit none in the array URAM (4,016 words per lane of 10,160) none in the array URAM
exact lm-head rows (G8) none fp16 in HBM: 128,256 x 3,072 x 2 B = 788 MB (DESIGN_DUAL_MODEL_4C1.md:144) none 2-bit in HBM: codes 317.8 + scales 19.9 + zeros 2.5 = 340.2 MB, 1,408 B per row vs 6,144 B fp16 at H3072 / 10,240 B at H5120 (DESIGN_FABLE_FULL2BIT_LUT_ARRAY_20261102.md:89; P2_G8W2_20261108.md:7)
G8 hidden-state store (FF row buffers in u_g8 own logic) LUTRAM 10,688 in the W2 dots (:1341)

4.2 Compute

precision class record ATT 0x13BE record MLPV 0x1361 Q5120 ATT 0x1346 Q5120 MLPV 0x1349
2-bit x activation GEMV 2b x A8 in LUT: 512 lanes x 988 LUT; u_array 597,623 LUT / 1,635 DSP 2b x A16 on DSP58: 256 lanes x (348 + 125) LUT / 18 DSP plus bank / scale DSP; u_array 147,305 LUT / 5,120 DSP 2b x A8 in LUT: 256 x 986; u_array 311,016 / 1,633 DSP 2b x A8 in LUT: 256 x 988; u_array 311,050 / 1,889 DSP
lane accumulators 48-b (result_q, ARRAY.md:135), inside the lane rows same same same
attention QK / softmax / PV (16-bit) 8 engines, 87,792 LUT / 3,160 DSP none 5 engines, 57,350 LUT / 1,965 DSP none
RMSNorm (fp/int 16-b, 48 DSP each) 4 instances 4 instances (incl. g11) 4 4
GLM low-rank (bf16 x 16-b) none 2,048 DSP none 2,048 DSP
G8 rescore dot none fp16 x fp16 wide dot, 4 x (22,751 LUT, 16 DSP) none exact 2-bit x int dot, 4 x (23,634 LUT, 4 DSP); P2_G8W2_20261108.md:108-113: u_g8 at H5120 -29,291 LUT, -3,463 FF, -48 DSP vs fp16
TREE top-K (16-bit scores) none 75,099 LUT u_tree none 75,099 LUT

5. Record -> Q5120 delta per card (measured, same instrument)

5.1 ATT (0x13BE -> 0x1346; different lane counts by design: L512 1 URAM per lane vs L256 2 URAM per lane; 0x1346 has no G7)

5.2 MLPV (0x1361 -> 0x1349)

What grew:

What shrank:

Net: device SLICE 89.6 -> 94.45 %, LUT 46 -> 53 %, DSP 69 -> 39 %.

6. Run log and instrument record