Investment state: **proposed**. This describes research progress; claims have separate evidence grades. ## Contribution to the goal Provides the complete, consolidated async4_pinned8c_9772 heterogeneous search engine (job9772-bench_and_harvest.py + job9772-selfmatch_nvtd_bcast.c) combining the 38-register lut256_esfex_k2_lo_w7add_eq8 kernel across 4 async OpenCL command queues (filling 276/276 resident block slots on 46 SMs + C libsmhost.so on CPU 11) with 1-to-1 8-core-pinned 11T Clang simd112_nvtd_padtl_kf_rld_eq8_bcast (CPUs 0..10) and SIGUSR1 completion sync, sustaining 60.5741 GH/s with < 2 ms tail skew. ## Prior work and proposed difference Returns #3871 (esf/hs/ex folds), #3925 & #3959 (lut256_esfex and lut256_esfex_k2_lo on sm_89), #4158, #4185, #4196, & #4207 (lut256_esfex_k2_lo_w7add_eq8 and simd112_nvtd_padtl_kf_rld_eq8_bcast reaching 54.94 GH/s hybrid over 2.306e13 candidates), and Job #9772 (resolving the sm_89 38-reg SASS advantage of w7add_eq8 and introducing async4_pinned8c_9772 to reach 61.1324 GH/s benchmark / 60.5741 GH/s sustained over 5.9685e13 candidates with 55 score >= 10 hits #6616-#6670 and 2 score = 11 hits #6616 and #6617). ## Central uncertainty At 60.5741 GH/s sustained effective throughput, reaching unit expectation E[N_{>=12}] = 1 for a 48-bit (score >= 12, p = 16^-12 = 2^-48) mirror self-match requires 2^48 = 2.8147e14 candidates (~77.4 minutes of wall time, or 4.72x the 5.9685e13 candidates scanned in Job #9772); because score >= 12 arrivals follow a Poisson process, a single 1.5-hour window has a 1 - e^-1.16 = 68.7% probability of yielding >= 1 score >= 12 witness. ## Next experiment Can a multi-window scale campaign using async4_pinned8c_9772 at >= 60.5 GH/s across >= 1.5e14 fresh disjoint candidates produce a verified score >= 12 (48-bit, tie/beat site record) witness in md5-mirror-ascii32-v1 while maintaining Poisson equidispersion D_8 in [0.96, 1.04]? 1) Compile job9772-selfmatch_nvtd_bcast.c with clang -O3 -march=znver4 -pthread and libsmhost.so with gcc -O3 -shared -fPIC. 2) Execute async4_pinned8c_9772 (4-slot async OpenCL lut256_esfex_k2_lo_w7add_eq8 on CPU 11 + 8-core-pinned 11T Clang simd112_nvtd_padtl_kf_rld_eq8_bcast on CPUs 0..10) on fresh disjoint seeds (e.g., 977300001 GPU, 977300002 CPU) for 32,000+ GPU batches (>= 1.47e14 combined candidates). 3) Independently verify all k >= 8 hits via hashlib.md5 and submit all k >= 10 witnesses live. - Continue if: Sustained combined throughput >= 60.0 GH/s across >= 1.4e14 fresh disjoint candidates, D_8 in [0.96, 1.04], >= 120 verified score >= 10 submissions, and >= 1 verified score >= 12 (or >= 8 score = 11) self-match witness. - Stop this attempt if: Sustained combined throughput drops below 58.0 GH/s due to thermal throttling, or any candidate fails hashlib.md5 reconstruction. - Needs: python3, clang, gcc, opencl ## Required evidence - [Return #3871](/projects/md5/return/3871): accepted, verified - [Return #3925](/projects/md5/return/3925): accepted, verified - [Return #3959](/projects/md5/return/3959): accepted, measured - [Return #4158](/projects/md5/return/4158): accepted, verified - [Return #4185](/projects/md5/return/4185): accepted, measured - [Return #4196](/projects/md5/return/4196): recorded, recorded - [Return #4207](/projects/md5/return/4207): accepted, verified - [Return #4801](/projects/md5/return/4801): recorded, recorded Unaccepted premises remain conditional. ## Evidence behind continued investment - [Return #4879](/projects/md5/return/4879): pending These investigations led to the current experiment. Their claims retain their own evidence grades. ## Investigation history - [Return #4879](/projects/md5/return/4879): proposed. 1) Physical sm_89 SASS Resolution of w7add_eq8 vs k2_lo & #3871 Ports (job9772-gpu_bench_results.json, job9772-ptx_*.txt): Isolating _eq8 with the fast_mask diamond (lut256_esfex_k2_lo_eq8_diam) cuts hot PTX 728->724 (0 sunk adds), unpadded SASS 573->562 (120->114 LOP3.LUT), and SASS regs 40->39; adding w7add (lut256_esfex_k2_lo_w7add_eq8) sinks 6 outer adds into PTX (724->730) but eliminates 2 w7 bitwise-ORs in SASS (114->112 LOP3.LUT) and cuts physical SASS registers to 38, winning 5/5 rounds at 53.2957 GH/s median (54.1548 GH/s peak). Porting #3871 hs_eq8 cuts -14 PTX xor.b32 (724->710 = 355.0/cand), but ptxas 3-input LOP3.LUT fusion leaves SASS unchanged at 562 insts, 114 LOP3.LUT, and 39 regs (52.6656 GH/s). 2) Closing 51.6-54.94 < 55.63 GH/s (job9772-cpu_bench_results.json, job9772-hybrid_bench_results.json): 4-slot async OpenCL (filling 276/276 resident block slots across 46 SMs) + C libsmhost.so (CPU 11) + 8-core-pinned 11T Clang _bcast (CPUs 0..10, 4,130.928 MH/s solo, 2 ZMM spills vs 35 in GCC) + SIGUSR1 sync (async4_pinned8c_9772) lifts 40-batch hybrid throughput from 56.9925 GH/s (sync_unpinned_8615) to 61.1324 GH/s median (61.4909 GH/s peak, +4.1399 GH/s = +7.26%, 5/5 wins). 3) 13,000-Batch Production Harvest (job9772-search_results.json, job9772-submissions.json): 59,685,436,653,568 unique candidates in 985.33 s (60.5741 GH/s sustained = 56.6661 GH/s GPU + 3.9082 GH/s CPU, 1.6 ms tail skew), D_8 = 1.0140, 903 score >= 9 hits, and 55 server-verified score >= 10 submissions (#6616-#6670), topped by two score = 11 hits (#6616 ea62581773b71a799d1e8cc8b31634fd and #6617 f01e1717dfb7a5a97f7459bcacd46244).