slime/research/BENCHMARK.md

# Fixed-position benchmark

Date: 2026-10-03

Hardware used for this run:

- GPU: NVIDIA GeForce RTX 3060 Laptop GPU, compute capability 8.6
- CPU: the current WSL host, using 16 worker threads
- CUDA toolkit: 12.0

## Workload

Find every lower-48 seed producing a solid 5x5 slime pattern whose top-left
chunk is `(51069, 419124)`.  The known answer contains 24 seeds.

The new solver performs the complete workload and checks its final candidates
with Java-compatible `nextInt(10)`.  Running a complete old-style `2^48` scan is
unnecessarily expensive, so the brute-force baselines scan large contiguous
ranges, report throughput, and linearly project the full-space time.

## Results

| Implementation | Measured work | Measured time | Projected/complete time | Speedup |
|---|---:|---:|---:|---:|
| Old brute-force CPU | `2^32` seeds | 1.851 s | 121,290 s (33.7 h) | 1x |
| New solver CPU | complete, 24/24 results | 0.61 s wall | 0.61 s | ~199,000x |
| Old brute-force CUDA | `2^36` seeds | 1.256 s | 5,145 s (85.7 min) | 1x |
| New solver CUDA | complete, 24/24 results | 59.6 ms kernel | 59.6 ms | ~86,300x |
| New solver CUDA, end-to-end | complete, 24/24 results | 0.36 s wall | 0.36 s | ~14,300x |

The end-to-end CUDA row includes process startup, CUDA context creation, CPU
enumeration of the 84 lower-18 roots, allocation, copies, kernel execution, and
readback.  The kernel-only comparison isolates the algorithms on the GPU.

The new CUDA kernel is about 10.2x faster than the new 16-thread CPU solver for
this particular 5x5 position.  The old brute-force CUDA baseline is about 23.6x
faster than its CPU counterpart.

## Commands

```bash
make -C research

# New complete CPU solver
/usr/bin/time -f 'wall_seconds=%e' \
  research/slime_pattern_cpp search \
  --x 51069 --z 419124 -n 5 -p 16 --max-results 0

# Old-style CPU throughput
research/slime_bruteforce_cpp_bench \
  --x 51069 --z 419124 -n 5 --count 4294967296 -p 16

# New complete CUDA solver
research/slime_pattern_cuda \
  --x 51069 --z 419124 -n 5 --threads 256 --max-results 0

# Old-style CUDA throughput
research/slime_bruteforce_cuda_bench \
  --x 51069 --z 419124 -n 5 --count 68719476736
```

## Cross-check against the original root CUDA executable

The repository's original `slime-cuda` was also run directly over one billion
seeds with `-n 5 -d 25 -r 1`.  It took 0.439 seconds.  That program checks nine
top-left positions per seed, so it is not the same fixed-position workload.
Normalized as position/seed pairs it processed about 20.5 billion pairs/s,
whereas the dedicated old-style benchmark processed about 54.7 billion
seeds/s.  Using the faster dedicated baseline therefore makes the reported
new-vs-old acceleration conservative.

## Interpretation

Most of the acceleration is algorithmic, not CUDA-specific:

- brute force visits all `2^48 = 281,474,976,710,656` seed values;
- the known position has only 84 viable lower-18 roots;
- those expand to 11,010,048 lower-35-bit values;
- each low-bit value's remaining 13 bits are solved by threshold/residue intersection,
  rather than enumerating another 8192 values.

Results depend strongly on pattern selectivity.  Sparse patterns and patterns
with few required-slime cells retain more roots and can produce many more upper
solutions, reducing the speedup.  These numbers should not be extrapolated to
world-border position enumeration, which remains a separate problem.

## Expanded 64-position run

The initial result above is useful for correctness, but one known hit is not a
performance distribution.  A second, deterministic stratified run therefore
used 64 positions:

- 32 positions at and around the known difficult coordinate, deliberately
  retaining a range of nonzero lower-18 root counts;
- 32 positions covering the origin, four world-border corners, large signed
  coordinates, and pseudorandom world-border positions (seed `20261004`).

For every position the new CPU and CUDA solvers completed the whole lower-48
search.  The old CPU baseline actually checked `2^28` seeds per position and
the old CUDA baseline checked `2^34` per position.  Aggregate measured old-style
work was therefore:

```text
CPU:  64 * 2^28 =  17,179,869,184 seeds
CUDA: 64 * 2^34 = 1,099,511,627,776 seeds (2^40)
```

CPU and CUDA agreed on the result count for all 64 positions.  Only the known
coordinate had matches (24); every other sampled position had zero.

### Baseline stability

Across the 64 coordinates, brute-force throughput was:

| Baseline | Mean | Median | Range | 95% CI of mean |
|---|---:|---:|---:|---:|
| CPU, 16 threads | 2.187B seeds/s | 2.171B | 1.665B–2.686B | 2.139B–2.235B |
| RTX 3060 CUDA | 53.81B seeds/s | 53.76B | 49.36B–58.13B | 53.46B–54.15B |

The CUDA coefficient of variation was only 2.6%, which supports linear
projection of its complete `2^48` time to roughly 87 minutes.

### Root and runtime distribution

The full set contained 42 positions with zero lower-18 roots and 22 with
nonzero roots.  Zero-root positions are a real algorithmic result—the initial
`2^18` parity pass proves that their entire lower-48 space is impossible—but
they are excluded from the speedup table below to avoid inflating it.

For the 22 nonzero-root positions:

| Metric | Minimum | Median | Mean | P95 | Maximum |
|---|---:|---:|---:|---:|---:|
| lower-18 roots | 1 | 20 | 28.6 | 84 | 152 |
| lower-35-bit values | 131,072 | 2,621,440 | 3,753,425 | 11,010,048 | 19,922,944 |
| New CPU search time | 6.50 ms | 166.6 ms | 225.3 ms | 606.5 ms | 1.153 s |
| New CUDA kernel time | 1.64 ms | 14.61 ms | 20.85 ms | 60.81 ms | 107.6 ms |
| CPU algorithm speedup | 107,104x | 774,825x | 4,444,739x | 18,288,517x | 19,220,240x |
| CUDA algorithm speedup | 48,055x | 369,354x | 969,525x | 3,120,648x | 3,252,226x |

The deliberately hardest sampled position was `(51069, 419125)`: 152 roots,
19,922,944 lower-35-bit values, 1.153 seconds on CPU, and 107.6 ms on CUDA.  Even
there, the conservative same-platform acceleration was 107,104x on CPU and
48,055x on CUDA.

The expanded run is reproducible with:

```bash
python3 research/benchmark_suite.py \
  --samples 64 \
  --cpu-threads 16 \
  --cpu-old-count 268435456 \
  --cuda-old-count 17179869184 \
  --cuda-threads 256 \
  --output /tmp/slime_benchmark_64.json
```

Times in this section intentionally compare the measured search/kernel regions.
CUDA process startup and context creation are not included; the earlier
single-position end-to-end row shows their effect.  The upper solver still has
the documented eight-value Java rejection-tail completeness caveat.

## Coordinate-side outer-sum CUDA experiment

The follow-up coordinate scanner fixes world seed `61692151016558` and checks
an `8193 x 8193` rectangle containing the known 5x5 top-left position:

```bash
./research/slime_position_cuda --repeats 5
```

Each invocation runs two kernels and requires identical sorted hit lists:

- direct: one CUDA thread per position, evaluating the full coordinate
  polynomial as cells are needed;
- outer-sum: one 16x16 tile per CUDA block, sharing expanded `F(x+i)` and
  `G(z+j)` terms across the tile.

On the same RTX 3060 Laptop GPU, median kernel times were:

| implementation | positions | kernel time | positions/s |
|---|---:|---:|---:|
| direct | 67,125,249 | 1.608 ms | 41.75 billion |
| tiled outer-sum | 67,125,249 | 1.332 ms | 50.39 billion |

That is a 1.21x outer-sum speedup.  Both paths found exactly one hit,
`(51069, 419124)`.  Additional 4,004,001-position cross-checks at negative and
near-world-border coordinates agreed exactly and showed 1.30--1.31x speedups.

The direct rate is high because a random position usually fails after one or
two slime tests; it does not normally pay for all 25 cells.  Consequently this
experiment measures the arithmetic/reuse benefit of outer-sum, not the hoped-for
algorithmic benefit of signature grouping.  It still scans the Cartesian area
and therefore is not a feasible full-world-border method.

## Full-world fixed-seed residue filter

The next prototype uses the fact that the necessary first-output parity is
periodic in both coordinates modulo `2^18`.  It checks all `2^36` residue pairs
on CUDA, expands survivors over the world border, and exactly verifies only
those expanded coordinates:

```bash
./research/slime_position_residue_cuda --seed 61692151016558
```

Results for four complete residue-space runs were:

| seed | residue kernel | survivors | expanded exact candidates | exact time | hits |
|---:|---:|---:|---:|---:|---:|
| 61692151016558 | 1.931 s | 180,476 | 36,937,048 | 284 ms | 1 |
| 0 | 1.920 s | 195,672 | 40,046,949 | 352 ms | 0 |
| 1 | 1.935 s | 175,992 | 36,010,787 | 283 ms | 0 |
| 281474976710655 | 1.944 s | 181,276 | 37,098,020 | 313 ms | 0 |

The searched world rectangle contains 14,062,500,000,000 positions.  For the
known seed, the necessary filter reduces the exact workload by about 381,000x.
Axis-table preparation took 224--289 ms, making total measured runtime roughly
2.5 seconds.

For a larger direct baseline, both rectangular kernels were run five times over
`2^34 = 17,179,869,184` positions.  Median direct time was 524.1 ms (32.78
billion positions/s), while tiled outer-sum took 315.9 ms (54.38 billion/s).
Linear projection over the full rectangle is about 429 seconds direct or 259
seconds tiled, versus about 2.5 seconds for the residue pipeline.  Thus the new
fixed-seed method is approximately 170x faster than direct and 100x faster than
the tiled Cartesian scan.  The residue run itself actually processes all
`2^36` pairs; only the Cartesian full-world numbers are projections.

Like the seed-side prototype, the necessary parity filter omits the extremely
rare branch where Java rejects its first `next(31)` output.  Exact verification
eliminates false positives but cannot restore a candidate already removed by
that filter.

## Retired joint-search experiments

X/Z independent filtering and its lower35 generation experiments were removed on 2026-10-11.
Their historical measurements and limitations are consolidated in
[X/Z 筛轴试错报告](XZ_FILTER_TRIAL_REPORT.md). No full-world joint speedup was established.