CPU Public API Benchmark Results

• Suite: cpu/public_api
• Target profile: amd-cpu
• Suite file: benchmarks/cpu/public_api.yaml
• Public API coverage manifest: benchmarks/cpu/public_api_coverage.yaml
• Run metadata: data/results/amd-cpu/cpu/public_api/20261008_060822/run.yaml
• Timestamp: 20261008_060822

Latest run: PUBLICATION_GATE_PROFILE=full ./scripts/run_cpu_public_api.sh 1 4.

• Sampling: full profile, 15 measured runs, 3 warmups per row

This file is generated from sequential CPU public API runs under data/results/amd-cpu/cpu/public_api/20261008_060822.

• tenferro-rs commit: 5cf78c7ec0ad9516dd78bab546d5e7bd42fa3102

CPU Information

• Model: AMD Ryzen 9 9955HX 16-Core Processor
• Vendor: AuthenticAMD
• Logical CPUs: 32
• Sockets: 1
• Cores per socket: 16
• Threads per core: 2
• NUMA nodes: 1
• Python platform: Linux-7.0.0-38-generic-x86_64-with-glibc2.39

Thread Environments

Threads: 1

• Run metadata: data/results/amd-cpu/cpu/public_api/20261008_060822/run_t1.yaml
• OMP_NUM_THREADS: 1
• OMP_THREAD_LIMIT: 1
• OMP_DYNAMIC: FALSE
• RAYON_NUM_THREADS: 1
• OPENBLAS_NUM_THREADS: 1
• GOTO_NUM_THREADS: 1
• MKL_NUM_THREADS: 1
• VECLIB_MAXIMUM_THREADS: 1
• VECLIB_NUM_THREADS: 1
• NUMEXPR_NUM_THREADS: 1
• BLIS_NUM_THREADS: 1
• XLA_FLAGS: --xla_cpu_multi_thread_eigen=false --xla_cpu_experimental_ynn_fusion_type= intra_op_parallelism_threads=1
• JULIA_NUM_THREADS: 1

Threads: 4

• Run metadata: data/results/amd-cpu/cpu/public_api/20261008_060822/run_t4.yaml
• OMP_NUM_THREADS: 4
• OMP_THREAD_LIMIT: 4
• OMP_DYNAMIC: FALSE
• RAYON_NUM_THREADS: 4
• OPENBLAS_NUM_THREADS: 4
• GOTO_NUM_THREADS: 4
• MKL_NUM_THREADS: 4
• VECLIB_MAXIMUM_THREADS: 4
• VECLIB_NUM_THREADS: 4
• NUMEXPR_NUM_THREADS: 4
• BLIS_NUM_THREADS: 4
• XLA_FLAGS: --xla_cpu_multi_thread_eigen=true --xla_cpu_experimental_ynn_fusion_type= intra_op_parallelism_threads=4
• JULIA_NUM_THREADS: 4

Timing Discipline

• Input fixture tensors are created during warmup and outside the measured region for tenferro-rs, PyTorch, and JAX.
• The tenferro-rs direct column measures immediate public operations (normally concrete Tensor + CpuBackend; lstsq/svd_full use EagerTensor because no concrete spelling exists). It is not labeled as the EagerTensor call layer.
• tenferro-rs trace graphs are constructed and compiled outside the measured region; each compiled graph is reused for every warmup and timed run.
• JAX functions are compiled with jax.jit during warmup, outside the measured region; timed calls include dispatch through jax.block_until_ready.
• Allocation-returning API rows create their output tensor inside each timed call.
• cpu/output_reuse rows allocate the destination during warmup and reuse it; tenferro-rs *_into is compared with PyTorch out=/copy_.
• Trace mode is unsupported for caller-output rows because compiled tenferro-rs graphs own their output tensors; JAX is shown as missing because it has no equivalent mutable out= API.
• PyTorch view-producing indexing operations are cloned inside the timed region to match tenferro-rs owned, materialized outputs.
• reshape compares materialized outputs: PyTorch clones its reshape view inside timing to match tenferro-rs' output-sized write; JAX has value semantics and no public strided-view contract.
• cpu/view_metadata separately compares concrete tenferro-rs, PyTorch, and Julia (julia-base) view creation without an output-sized copy; trace mode is unsupported and JAX is missing because neither exposes the same concrete strided-view contract.
• Julia cpu/view_metadata rows are reshape/transpose/@view lazy wrappers with no output-sized copy; broadcast_in_dim_view has no natural Base spelling and stays missing for Julia.
• Julia cpu/output_reuse rows use broadcast-into (.=), mul!, and copyto! with destinations allocated during warmup and reused, the same reuse discipline as the tenferro-rs _into/PyTorch out= rows above.
• Rust, PyTorch, and JAX fixtures contain identical logical values. Python/JAX reconstruct the Rust column-major fixture in each framework's native layout before timing.
• PyTorch complex conjugation uses torch.conj_physical to match tenferro-rs physical output rather than the lazy conjugate view from torch.conj.
• dot_general_with_conj instead uses PyTorch's lazy conjugate view so conjugation can be handled by the contraction, matching tenferro-rs' conjugation flags; trace is unsupported because there is no equivalent public traced API. Julia's dot_general_with_conj row materializes the conjugate before the GEMM instead, since Julia has no lazy conj-without-transpose spelling that BLAS can fuse.
• pad has no natural Base spelling and stays missing for Julia.
• dynamic_update_slice reports trace mode as unsupported because tenferro-rs does not currently expose a corresponding TracedTensor API.
• full_piv_lu and full_piv_lu_solve are excluded because PyTorch has no direct public full-pivot equivalent; substituting torch.linalg.solve would compare different algorithms.
• cpu/linalg_batched rows use a batch of 1024 matrices with one rhs column. batched_lu_factor times the packed LU factorization alone. batched_lu_solve reuses LU factors prepared during warmup, outside the measured region, and times only the solve (tenferro-rs LinalgBackend::lu_solve_prepared, PyTorch torch.linalg.lu_solve, JAX jax.scipy.linalg.lu_solve, Julia LAPACK.getrs!). batched_triangular_solve times a lower-triangular solve.
• Trace mode is unsupported for batched_lu_factor and batched_lu_solve because LinalgOp is not public, so a trace cannot hold a bare LU factor or prepared solve. The traced solve (LU factor plus prepared solve) and its backward are measured by the cpu/cpu_ops batched_solve and grad_sum_batched_solve_backward rows.
• Julia has no batched LAPACK entry point, so the Julia cpu/linalg_batched rows loop the per-matrix LAPACK call over the batch axis and allocate their outputs inside the timed call.
• The tenferro-rs runner builds its direct CpuBackend, the trace runtime, and the eager runtime with exactly --num-threads workers, and it fails at startup when a thread environment variable disagrees with that count or when the backend execution scope reports a different Rayon thread count.
• svd_full remains in the table even when the selected tenferro-rs provider reports it as unsupported.
• Julia is column-major, like tenferro-rs, so the julia-base/strided-jl columns need no PyTorch/JAX-style layout reconstruction to keep the same logical fixture values.
• Julia warmup runs move JIT compilation outside the measured region, the same way PyTorch/JAX warmups do.
• Julia factorization rows materialize their factors inside the timed call (cholesky returns the factor matrix, lu returns P/L/U, qr returns Q/R). A Julia Factorization keeps its factors packed in LAPACK's working storage, so timing the compact object would compare strictly less work than the tenferro-rs and PyTorch columns, which return separate materialized tensors.
• Julia dense linalg rows run through Julia's own BLAS/LAPACK (libblastrampoline, by default OpenBLAS), recorded as julia.blas_provider in the run metadata. When that differs from the provider tenferro-rs and PyTorch link against (Accelerate on macOS), those rows partly compare BLAS implementations rather than framework overhead; read them together with the recorded providers.
• The Julia lstsq row uses qr(a) \ rhs rather than a \ rhs: the bare backslash runs a column-pivoted, rank-revealing QR (the LAPACK gelsy algorithm), while the PyTorch row selects the gels driver, so the unpivoted spelling is the like-for-like comparison.
• julia-base uses the natural Base/LinearAlgebra spelling and strided-jl the natural Strided.jl (@strided) spelling; each is populated only for rows where that spelling naturally applies (reductions and dense linalg have no natural Strided.jl spelling, so strided-jl covers elementwise/chain/transpose rows, the elementwise cpu/output_reuse _into rows, and the elementwise cpu/complex rows conj/mul/div/exp/log).
• Strided.jl (https://github.com/Jutho/Strided.jl) is prior art for tenferro-rs' strided-rs kernel layer; the strided-jl column credits that lineage directly in the report.

Threads: 1 4

• CSV: data/results/amd-cpu/cpu/public_api/20261008_060822/cpu_public_api_t1_20261008_060822.csv
• CSV: data/results/amd-cpu/cpu/public_api/20261008_060822/cpu_public_api_t4_20261008_060822.csv
• Source table: data/results/amd-cpu/cpu/public_api/20261008_060822/cpu_public_api_20261008_060822.md

CPU Benchmark Items

Median ± IQR (ms). Missing backends are shown as -.