← all results

GPU latency results

Latency and per-call overhead. These cases are dominated by launch, dispatch and synchronization, not by device compute. Read them as API overhead, never as GPU/CPU throughput.

NVIDIA A100

NVIDIA A100nvidia-gpuhardware: NVIDIA A100 80GB PCIedate: 2026-10-27 01:23:06Ztenferro: 06ac2475a286
gpu/dense 1 case · 1 flagged

baseline PyTorch · best first · 1 case · case table, absolute times and raw report →

bar = case median / PyTorch median · dashed line = baseline (1.00x) · log scale · shorter (left) is faster · whisker = estimated 95% CIdense_batched_matmul_f64_b16_32eager 3.49xtrace 7.34x
gpu/elementwise 1 case · 1 flagged

baseline PyTorch · best first · 1 case · case table, absolute times and raw report →

bar = case median / PyTorch median · dashed line = baseline (1.00x) · log scale · shorter (left) is faster · whisker = estimated 95% CIelementwise_chain_f64_1keager 1.63xtrace 0.64xJAX 0.71x
gpu/linalg_ad_latency 30 cases

baseline PyTorch · best first · 30 cases · case table, absolute times and raw report →

bar = case median / PyTorch median · dashed line = baseline (1.00x) · log scale · shorter (left) is faster · whisker = estimated 95% CIlinalg_ad_small_grad_sum_lu_vjp_f64_2trace 0.96xJAX 0.18xlinalg_ad_small_grad_sum_lu_vjp_f64_8trace 0.97xJAX 0.19xlinalg_ad_small_grad_sum_lu_vjp_f64_4trace 0.99xJAX 0.19xlinalg_ad_small_grad_sum_svd_s_jvp_f64_2trace 1.05xJAX 0.76xlinalg_ad_small_grad_sum_svd_s_jvp_f64_4trace 1.09xJAX 0.61xlinalg_ad_small_grad_sum_lu_jvp_f64_2trace 1.11xJAX 0.23xlinalg_ad_small_grad_sum_eigh_vjp_f64_2trace 1.16xJAX 0.50xlinalg_ad_small_grad_sum_svd_s_jvp_f64_8trace 1.19xJAX 0.69xlinalg_ad_small_grad_sum_eigh_jvp_f64_2trace 1.20xJAX 0.53xlinalg_ad_small_grad_sum_eigh_vjp_f64_8trace 1.22xJAX 0.87xlinalg_ad_small_grad_sum_lu_jvp_f64_4trace 1.24xJAX 0.26xlinalg_ad_small_grad_sum_eigh_jvp_f64_4trace 1.27xJAX 0.64xlinalg_ad_small_grad_sum_eigh_vjp_f64_4trace 1.27xJAX 0.59xlinalg_ad_small_grad_sum_svd_s_vjp_f64_2trace 1.34xJAX 0.61xlinalg_ad_small_grad_sum_lu_jvp_f64_8trace 1.35xJAX 0.28xlinalg_ad_small_grad_sum_qr_vjp_f64_2trace 1.38xJAX 0.22xlinalg_ad_small_grad_sum_svd_s_vjp_f64_4trace 1.41xJAX 0.75xlinalg_ad_small_grad_sum_qr_vjp_f64_8trace 1.44xJAX 0.24xlinalg_ad_small_grad_sum_solve_vjp_f64_2trace 1.47xJAX 0.21xlinalg_ad_small_grad_sum_eigh_jvp_f64_8trace 1.47xJAX 0.96xlinalg_ad_small_grad_sum_solve_vjp_f64_8trace 1.49xJAX 0.20xlinalg_ad_small_grad_sum_qr_vjp_f64_4trace 1.50xJAX 0.22xlinalg_ad_small_grad_sum_svd_s_vjp_f64_8trace 1.51xJAX 0.84xlinalg_ad_small_grad_sum_solve_vjp_f64_4trace 1.69xJAX 0.21xlinalg_ad_small_grad_sum_solve_jvp_f64_2trace 1.74xJAX 0.25xlinalg_ad_small_grad_sum_solve_jvp_f64_4trace 1.77xJAX 0.24xlinalg_ad_small_grad_sum_solve_jvp_f64_8trace 1.84xJAX 0.25xlinalg_ad_small_grad_sum_qr_jvp_f64_2trace 1.96xJAX 0.39xlinalg_ad_small_grad_sum_qr_jvp_f64_8trace 2.12xJAX 0.37xlinalg_ad_small_grad_sum_qr_jvp_f64_4trace 2.20xJAX 0.38x
gpu/perf_issues 16 cases

baseline declared reference · best first · 16 cases · case table, absolute times and raw report →

bar = case median / declared reference median · dashed line = baseline (1.00x) · log scale · shorter (left) is faster · whisker = estimated 95% CIqr_readback_latency_f64_n32tenferro-cuda 1.91xsolve_readback_latency_f64_n32tenferro-cuda 1.96xqr_readback_latency_f64_n2tenferro-cuda 2.00xsolve_readback_latency_f64_n8tenferro-cuda 2.29xsolve_readback_latency_f64_n4tenferro-cuda 2.47xqr_readback_latency_f64_n4tenferro-cuda 3.10xqr_readback_latency_f64_n8tenferro-cuda 3.22xsolve_readback_latency_f64_n2tenferro-cuda 3.87x
Cases without a comparable baseline (8)

qr_readback_latency_f64_n2_pytorch, qr_readback_latency_f64_n32_pytorch, qr_readback_latency_f64_n4_pytorch, qr_readback_latency_f64_n8_pytorch, solve_readback_latency_f64_n2_pytorch, solve_readback_latency_f64_n32_pytorch, solve_readback_latency_f64_n4_pytorch, solve_readback_latency_f64_n8_pytorch

Diagnostic reports (24)

nvidia-gpu