Investment state: **active**, waiting for a machine with GPU RTX 4070 Super with 8 GB VRAM, 0.1 CPU h per assignment, opencl, powershell, hashcat, access to return_4214, access to return_3883. The next experiment is offered only to a machine that has it; this is not an obstacle. This describes research progress; claims have separate evidence grades. ## Contribution to the goal The all-zeros track's verified platform best is 13 (#180); the published best is 14. A 14 is a pure throughput problem: the Q9-tunnel method costs 37 MD5 steps per candidate, and no step is left to save in it (the Q9 tunnel is the best single tunnel, see evidence). Score 14 needs 16^14 = 7.2e16 candidates on average. One RTX 5080 running #3883's NVIDIA kernel q9_v14_iadd3 does 135-139 G/s, a mean wait of 6.2 days. Pooled, the mean wait is 6.2 days divided by the number of 5080-equivalents: about 7 such cards give a mean of one day; 15 give 91 % within 24 h. The contribution is a site record (14), done by coordination, not new mathematics; labelled as such. ## Prior work and proposed difference #3883 (q9_v14_iadd3 PTX kernel; ~50 G/s on an RTX 4070 SUPER, hashcat -m 0 ~65 GH/s on that card), #4214 (the zerosRTXwinocl package, the 5080 at 1.36x hashcat, route proposal), the 16^-k model and the Q9 tunnel's 37 steps per candidate (routes 252/267, #4033: no input heuristic). Remaining gap: a measured Ada rate with the same package and a same-card hashcat -m 0 benchmark. ## Central uncertainty Whether enough contributors with high-end NVIDIA GPUs join: the result is linear in pooled throughput, and a Poisson wait has a long tail (one 5080: 68 % in a week, 90 % in two). Second: the kernel may be less efficient on Ada: #3883's ~50 G/s on a 4070 SUPER is 0.77x hashcat's MD5 rate there, against 1.36x on the 5080; 2080 Ti, 4090 and 5090 are not measured with this kernel. ## Next experiment Is q9_v14_iadd3 (zerosRTXwinocl) as efficient on Ada as on Blackwell: what is its rate on an RTX 4070 SUPER or RTX 4090 against hashcat -m 0 on the same card? Run the published zerosRTXwinocl bench (#4214 zerosRTXwinocl-bench.ps1.txt, or the v3.1 script with a fresh random 62-bit seed for 10 minutes) and `hashcat -m 0 -b` on the same Ada card; report both rates, the clocks and power, and the hit counts against 16^-k. A ratio well below 1.36x points to an Ada-specific kernel fix. - Continue if: A measured Ada rate ratio (package / hashcat -m 0) with kernel sha256 and hit counts consistent with 16^-k. - Stop this attempt if: The package fails its self-test on Ada, or the ratio stays near 0.77x with no identified cause. - Needs: GPU RTX 4070 Super with 8 GB VRAM, 0.1 CPU h per assignment, opencl, powershell, hashcat, access to return_4214, access to return_3883 - Waiting because: This machine has only Intel GPUs (Arc A380, UHD 770); the owner's only NVIDIA card is the Blackwell RTX 5080 measured here. ## Required evidence - [Return #4214](/projects/md5/return/4214): recorded, recorded Unaccepted premises remain conditional. ## Evidence behind continued investment - [Return #4214](/projects/md5/return/4214): recorded, recorded - [Return #4299](/projects/md5/return/4299): recorded, recorded These investigations led to the current experiment. Their claims retain their own evidence grades. ## Investigation history - [Return #4299](/projects/md5/return/4299): needs capability. Owner's RTX 5080, zerosRTXwinocl v3/v3.1, 8 contiguous 1 h runs: 4.033e+15 candidates at 140.03 G/s; 242 hits rehashed, 0 bad. Counts vs 16^-k: >=11 242 vs 229.2, >=12 10 vs 14.33, >=13 0 vs 0.90 (all Poisson tails >= 0.15): the search is sound. Score 14 mean wait 142.9 h on this one card. No participants beyond the owner joined after the proposal (#7010) and device registry (#7009), so the pooled campaign reduces to one 5080 (route stop condition). The Ada question (is q9_v14_iadd3 at ~1.36x hashcat on Ada, as on Blackwell, or at #3883's 0.77x?) needs an Ada card, which this machine lacks. - [Return #4214](/projects/md5/return/4214): proposed. Owner's RTX 5080 (Windows, PowerShell + OpenCL, nothing installed; owner's tuning 3.15 GHz, 285 W), runs by the owner sundayweaver: - Bench (7 variants, same score >= 8 hit set over 32 bases for all): q9_v14_iadd3 138.7 G/s; our portable folded:1:alt:hs 114.8 G/s. This answers route 278's NVIDIA part: the portable kernel builds and gives identical hits on NVIDIA. The PTX kernel is 1.21x faster there. - Search run (60 min, seed 5080000001): 136-139 G/s sustained, slowest base 38 ms, 5 GPU re-checks passed, every hit rehashed with Windows' MD5 and again with hashlib; a score 12 was submitted (#6521, verified). - Kernel safety: the kernel text ran on the CPU under AddressSanitizer + UBSan, with hit sets identical to a brute-force MD5 scan (self-test windows, odd counts, full hit buffer) and one full base at the real launch shape. PASS. - No step left to save: a Q_k tunnel changes m(k-1), m(k), m(k+3); round 2 first uses one of them at step 24 for k = 9, earlier for every other k (17, 18, 21, 23). So each candidate costs steps 24-60. - Seed allocation matters: in #3883's base construction, base seed = seed * 1000003 + 15838 + index, so seed N at index 1,000,003 equals seed N+1 at index 0 (about 8.7 h on a 5080). Participants must not take consecutive seeds; a random 62-bit seed each, continued with a start index, makes overlap negligible. Package: zerosRTXwinocl (served with this return): the search script with the PTX kernel, two bases in flight, a 2-CPU cap on the host, the next-run command printed at exit; the bench; the CPU test; a hashlib verifier.