01
judge overview
cpu.mode
live benchmark judge
Fastest code on the internet.
Compete on tiny systems-code challenges with single-file solutions measured on real CPU targets. Climb the leaderboards with runtime, cycles, IPC, branch misses, cache misses, records, and profiles.
global leaderboard
Σ 100×(Tbest/Tyou)2 | TOP 5
02
challenges
challenge catalog
Rust / C++ / Asm
01
Counting Bytes
Input is a 250 MiB byte array. Return the number of bytes equal to 127 by printing the count to stdout.
#1
josusanmartin
10.727ms
02
Summing Integers
Input is 10 million ASCII-encoded non-negative 32-bit integers in [0, 2^31 - 1], separated by single newlines (no trailing newline) on stdin. Print their sum to stdout.
#1
josusanmartin
6.763ms
03
Alyze UAX #29 Word Break
Implement turbopuffer/alyze's UAX #29 word-break tokenizer. The judge builds your source as libsolution.so and calls the exported function from a verifier executable.
#1
s7nfo
117.745ms
04
Convolution F_2^64
Input contains two length-6400 sequences of unsigned 64-bit integers representing elements of F_{2^64}. Print their convolution over F_{2^64}.
#1
josusanmartin
2.948ms
05
Tree Hash Traversal
This is the hot kernel from Anthropic's original performance take-home, without the Python VM or simulated VLIW machine.
#1
josusanmartin
22.905ms
06
Llama Q4_K Dot Product
Input is binary data containing packed llama.cpp-style Q4_K weight rows and Q8_K activation vectors.
#1
josusanmartin
4.934ms
07
Llama Q4_Kx8 GEMV
Same dot-product totals as llama_q4k_dot, but the Q4_K rows are delivered pre-packed in llama.cpp's block_q4_Kx8 layout (eight interleaved rows per packed block), so a GEMV-shaped kernel can be benchmarked without paying the repack cost.
#1
josusanmartin
4.292ms
08
NumKong Sparse Dot Product
Implement NumKong's u32-indexed sparse f32 dot product. The judge builds your source as libsolution.so and calls the exported function from a verifier executable.
#1
josusanmartin
6.497ms
09
Llama.cpp's ggml_gemv_q4_K_8x8_q8_K
Implement llama.cpp's ggml_gemv_q4_K_8x8_q8_K function. The judge builds your source as libsolution.so and calls the exported function from a verifier executable.
#1
s7nfo
4.802ms
10
Llama.cpp's ggml_gemv_q4_K_8x8_q8_K v2
Implement the effective llama.cpp ggml_gemv_q4_K_8x8_q8_K wrapper behavior with the unused bs argument removed. The judge builds your source as libsolution.so and calls the exported function from a verifier executable.
#1
josusanmartin
15.114ms
11
Llama.cpp's ggml_vec_dot_q6_K_q8_K
Implement llama.cpp's ggml_vec_dot_q6_K_q8_K function. The judge builds your source as libsolution.so and calls the exported function from a verifier executable.
#1
josusanmartin
9.536ms
03
recent important events
recent important events
10 tracked
first solution
NekrozQliphort submitted their first successful C++ solution to Counting Bytes; runtimes: Raptor Cove P-core 14.488 ms, Gracemont E-core 11.935 ms, AMD Zen 4 8.024 ms.
record
NekrozQliphort improved the C++ record on Counting Bytes / Gracemont E-core: 15.067 ms -> 11.935 ms (-20.8%).
record
s7nfo improved records on Tree Hash Traversal:
- C++ on Raptor Cove P-core: 25.698 ms -> 25.196 ms (-2.0%)
- C++ on AMD Zen 4: 11.790 ms -> 11.271 ms (-4.4%)
record
record
record
s7nfo improved records on Summing Integers:
- C++ on Raptor Cove P-core: 8.375 ms -> 7.387 ms (-11.8%)
- C++ on Gracemont E-core: 18.478 ms -> 16.106 ms (-12.8%)
- C++ on AMD Zen 4: 4.385 ms -> 3.761 ms (-14.2%)
record
s7nfo improved the Asm record on NumKong Sparse Dot Product / AMD Zen 4: 3.411 ms -> 3.379 ms (-0.9%).
record
record
record