← GPU results

gpu/dense_throughput

Legacy report: ingestion provenance is unrecorded; no raw artifact is inferred.

Compute-bound cases (3)

Absolute medians on a log time axis. Bars are shorter for faster calls; the dashed line is the best reference for each case.

Charts below are ordered best first (smallest ratio).

dense_batched_matmul_f64_b16_32_throughput batched_matmul · n=7

elementwise_chain_f64_1k_throughput elementwise_chain · n=7

elementwise_chain_f64_1m_throughput elementwise_chain · n=7

Original report

GPU Benchmark Results

• Target profile: nvidia-gpu
• Suite: gpu/dense_throughput
• Suite file: /workspaces/gpu-throughput/benchmarks/gpu/dense_throughput.yaml
• Timestamp: 2026-10-19T09:28:49.176769+00:00
• tenferro-rs commit: e921b059a5e3c41ea76002553cbb9415d4b5e58d

GPU Information

• Device: cuda:0
• Name: NVIDIA A100 80GB PCIe
• UUID: GPU-530977e1-4968-9283-4129-9fbec3e66542
• Memory: 80 GiB
• Driver version: 580.126.09
• CUDA version: 13.0
• CUDA runtime: 12.9
• cuDNN version: 92700

CPU Information

• Model: AMD EPYC 7713P 64-Core Processor
• Vendor: AuthenticAMD
• Logical CPUs: 64
• Sockets: 1
• Cores per socket: 64
• Threads per core: 1
• NUMA nodes: 1
• Python platform: Linux-6.8.0-101-generic-x86_64-with-glibc2.39

Median time is reported in milliseconds for ok records. Inputs are prepared on the GPU before timed runs; initial host-to-device transfer is outside the timed region. Timed runs include the host API call and backend-native device synchronization. tenferro-rs CUDA uses the explicit tenferro-rs synchronize API without downloading result tensors in the timed region. Dense and einsum inputs use the same deterministic benchmark generator in the Rust and Python runners; host-to-device layout conversion remains outside the timed region. tenferro-rs uses native column-major GPU tensors; PyTorch and vendor-wrapper columns use their native row-major framework tensors unless noted. The cuSOLVER column is torch.linalg with preferred_linalg_library=cusolver; for SVD it pins driver=gesvd as a QR-based cuSOLVER comparison. tenferro-rs CUDA SVD uses its backend default driver policy, currently gesvdj for matrices with both dimensions at most 1024 and gesvd otherwise. Non-ok cells show the structured backend status. Rows whose backend name ends in -deferred-batch are throughput mode: each timed sample enqueues K deferred prepared runs (Runtime::run_prepared_deferred) in one caller-held session and pays one retirement barrier (finish_prepared), so the reported per-operation time is sample_ms / K. They measure batched throughput and are not comparable with, and never averaged together with, per-call latency rows.

gpu/dense_throughput / batched_matmul / allocating output

gpu/dense_throughput / elementwise_chain / allocating output

Sampling

JAX uses compiled JIT execution; compilation, planning and numerical checks are outside timing. PyTorch uses eager execution. Samples retain every output until native synchronization and clock stop. Values are normalized per declared operation.