{ "schema": "aner.mnist-hardware/2", "measurement_dates_local": [ "2026-10-11", "2026-10-12" ], "time_zone": "America/Denver", "hosts": [ { "measured_date_local": "2026-10-11", "host_label": "Atlas-01 / C01 \u00b7 M2 Max", "cpu": "Apple M2 Max", "cpu_logical_cores": 12, "physical_ram_bytes": 68719476736, "gpu": "Apple M2 Max integrated GPU, 38 cores", "os": "26.7.1", "os_build": "25G241", "architecture": "arm64", "compiler": "Apple clang version 21.0.0 (clang-2100.3.34.2)", "build": "CMake Release; -O3 -DNDEBUG; C++20; -fno-fast-math; -ffp-contract=off", "cpu_provider": "Apple Accelerate", "cpu_threading": "VECLIB_MAXIMUM_THREADS=2 requested; exact provider thread count unavailable", "mps_float64": "native unavailable; preserve policy uses CPU Float64", "manifest_sha256": "2c53a427c279b889d953487fac454ee62f2dec1f16598b4ece09f85b0596499b", "measured_binary_sha256": "f041f6f817ebd3bf23750b92ea32562785af473b7cfe573e22d532f33ea37de5", "source_template_sha256": { "simple": "4505fbc2e3ee4887b192e375cbe54a0fe0e194ac3976928f6ca080b591bd9d02", "functional": "20921e2b382f0bf2c7bd12a1e5afd3ba9c8ce7044ecb349cc2c3b94d390777d2", "oop": "35de0266e22da02e730da15a278bced1fdde3784be6e880ebbce6321f96e2d33" }, "host": "atlas-01", "system_memory_scope": "installed unified RAM", "exclusive_hardware": false, "configuration_revision": "C01", "configuration_id": "atlas-01/C01" }, { "host": "atlas-02", "measured_date_local": "2026-10-12", "hardware_snapshot_utc": "2026-10-12T14:08:13.829073+00:00", "cpu": "Intel(R) Xeon(R) CPU E5-2697 v2 @ 2.70GHz", "cpu_vcpus": 2, "virtualization": "KVM", "system_memory_bytes": 16155795456, "system_memory_scope": "guest-visible OS MemTotal", "gpu": "NVIDIA GeForce GTX 1060 3GB", "device_vram_bytes": 3221225472, "cuda_compute_capability": "6.1", "driver": "580.178.04", "os": "Ubuntu 26.04.1 LTS", "architecture": "x86_64", "glibc": "2.43", "compiler": "GCC 13.4.0 (Ubuntu 13.4.0-10ubuntu1), compiled on Atlas-02", "cuda_compiler": "CUDA 12.4.131, native sm_61/sm_70 and compute_70 PTX", "build": "Native Release; -O3 -DNDEBUG; C++20; -fno-fast-math; -ffp-contract=off; CUDA --fmad=false; strict FP32 cuBLAS, no TF32", "cpu_provider": "OpenBLAS 0.3.32", "cpu_provider_core": "Sandybridge", "cpu_threads_queried": 2, "source_packet_sha256": "728ffd9b28f515ce39665dcf97cae226d0c73ec70eb89d8bfc00c9233bb799d8", "measured_binary_sha256": "ea0839a2f567b9a54a56957f68c99347c8b5fdde3541f0b475cf0d0ede153b4d", "manifest_sha256": "2c53a427c279b889d953487fac454ee62f2dec1f16598b4ece09f85b0596499b", "source_template_sha256": { "simple": "4505fbc2e3ee4887b192e375cbe54a0fe0e194ac3976928f6ca080b591bd9d02", "functional": "20921e2b382f0bf2c7bd12a1e5afd3ba9c8ce7044ecb349cc2c3b94d390777d2", "oop": "35de0266e22da02e730da15a278bced1fdde3784be6e880ebbce6321f96e2d33" }, "exclusive_hardware": false, "background_gpu_contexts_before_benchmark": 0, "background_gpu_allocated_bytes_before_benchmark": 0, "background_gpu_note": "Existing services were retained. Resident allocation does not establish active compute contention. Benchmark process memory is attributed to its own PID.", "host_label": "Atlas-02 / C01 \u00b7 GTX 1060 3 GB", "kernel": "Linux 7.0.0-38-generic", "ambient_gpu_monitoring": { "interval_seconds": 1, "start_scope": "Started during the CPU Float64 profile, before CUDA profiles; separate benchmark-PID allocation sampling uses 100 ms for all CUDA trials." }, "configuration_revision": "C01", "configuration_id": "atlas-02/C01" }, { "host": "atlas-03", "measured_date_local": "2026-10-12", "hardware_snapshot_utc": "2026-10-12T14:09:31.764166+00:00", "cpu": "Intel(R) Xeon(R) Gold 6248 CPU @ 2.50GHz", "cpu_vcpus": 24, "virtualization": "KVM", "system_memory_bytes": 64929792000, "system_memory_scope": "guest-visible OS MemTotal", "gpu": "Tesla V100-PCIE-32GB", "device_vram_bytes": 34359738368, "cuda_compute_capability": "7.0", "driver": "580.178.04", "os": "Ubuntu 26.04.1 LTS", "architecture": "x86_64", "glibc": "2.43", "compiler": "GCC 13.4.0 (Ubuntu 13.4.0-10ubuntu1), compiled on Atlas-02", "cuda_compiler": "CUDA 12.4.131, native sm_61/sm_70 and compute_70 PTX", "build": "Native Release; -O3 -DNDEBUG; C++20; -fno-fast-math; -ffp-contract=off; CUDA --fmad=false; strict FP32 cuBLAS, no TF32", "cpu_provider": "OpenBLAS 0.3.32", "cpu_provider_core": "SkylakeX", "cpu_threads_queried": 2, "source_packet_sha256": "728ffd9b28f515ce39665dcf97cae226d0c73ec70eb89d8bfc00c9233bb799d8", "measured_binary_sha256": "ea0839a2f567b9a54a56957f68c99347c8b5fdde3541f0b475cf0d0ede153b4d", "manifest_sha256": "2c53a427c279b889d953487fac454ee62f2dec1f16598b4ece09f85b0596499b", "source_template_sha256": { "simple": "4505fbc2e3ee4887b192e375cbe54a0fe0e194ac3976928f6ca080b591bd9d02", "functional": "20921e2b382f0bf2c7bd12a1e5afd3ba9c8ce7044ecb349cc2c3b94d390777d2", "oop": "35de0266e22da02e730da15a278bced1fdde3784be6e880ebbce6321f96e2d33" }, "exclusive_hardware": false, "background_gpu_contexts_before_benchmark": 1, "background_gpu_allocated_bytes_before_benchmark": 2392850432, "background_gpu_note": "Existing services were retained. Resident allocation does not establish active compute contention. Benchmark process memory is attributed to its own PID.", "host_label": "Atlas-03 / C01 \u00b7 V100 32 GB", "kernel": "Linux 7.0.0-38-generic", "ambient_gpu_monitoring": { "interval_seconds": 1, "start_scope": "Started during the CPU Float64 profile, before CUDA profiles; separate benchmark-PID allocation sampling uses 100 ms for all CUDA trials." }, "configuration_revision": "C01", "configuration_id": "atlas-03/C01" }, { "host": "atlas-04", "measured_date_local": "2026-10-12", "hardware_snapshot_utc": "2026-10-12T16:20:33.395447+00:00", "cpu": "Intel(R) Xeon(R) CPU E5-2630 v3 @ 2.40GHz", "cpu_vcpus": 32, "virtualization": "QEMU", "system_memory_bytes": 135058767872, "system_memory_scope": "guest-visible OS MemTotal", "gpu": "NVIDIA L40", "device_vram_bytes": 48305799168, "device_vram_scope": "driver-reported total memory", "cuda_compute_capability": "8.9", "driver": "615.71.09", "os": "Ubuntu 22.04.5 LTS", "architecture": "x86_64", "glibc": "2.35", "compiler": "GCC 11.4.0 (Ubuntu 11.4.0-1ubuntu1~22.04.3), compiled on Atlas-04", "cuda_compiler": "CUDA 12.4.131, native sm_89", "cuda_runtime": "12.4.127", "cublas": "12.4.5.8", "build": "Native Release; -O3 -DNDEBUG; C++20; -fno-fast-math; -ffp-contract=off; CUDA --fmad=false; strict FP32 cuBLAS, no TF32", "cpu_provider": "OpenBLAS 0.3.32", "cpu_provider_core": "Haswell", "cpu_provider_config": "OpenBLAS 0.3.32 NO_LAPACK NO_LAPACKE DYNAMIC_ARCH NO_AFFINITY Haswell MAX_THREADS=64", "cpu_threads_queried": 2, "cpu_provider_build_difference": "Private source build without Fortran/LAPACK; earlier Ubuntu guests use distribution pthread builds of the same version. This experiment uses SGEMM/DGEMM only.", "source_packet_sha256": "9e0d97c5718e2cd7661e893a623170b85aa18cbb868da416d113739ab3d68b30", "source_core_sha256": "bcb84a354a252ee1926cf1105faa197848fbd7dc3fd7431858eb1142e0cb90dd", "measured_binary_sha256": "7ba67a07afb7444ad8c88fb5d5cb65715132df91c73613accf636e2a7903be52", "manifest_sha256": "2c53a427c279b889d953487fac454ee62f2dec1f16598b4ece09f85b0596499b", "source_template_sha256": { "simple": "4505fbc2e3ee4887b192e375cbe54a0fe0e194ac3976928f6ca080b591bd9d02", "functional": "20921e2b382f0bf2c7bd12a1e5afd3ba9c8ce7044ecb349cc2c3b94d390777d2", "oop": "35de0266e22da02e730da15a278bced1fdde3784be6e880ebbce6321f96e2d33" }, "exclusive_hardware": false, "background_gpu_contexts_before_benchmark": 0, "background_gpu_allocated_bytes_before_benchmark": 0, "background_gpu_note": "No other compute context was present in the preparation snapshot. The VM was not reserved exclusively; resident allocation and utilization snapshots do not establish absence of contention throughout the experiment. Benchmark GPU process memory is attributed to its own PID.", "host_label": "Atlas-04 / C01 \u00b7 L40", "kernel": "Linux 5.15.0-198-generic", "ambient_gpu_monitoring": { "interval_seconds": 1, "start_scope": "Started before the first native profile and retained through the subsequent private comparison; separate own-PID allocation sampling uses 100 ms for CUDA trials." }, "configuration_revision": "C01", "configuration_id": "atlas-04/C01" } ], "comparison_scope": "Matched model, data selections, seeds, batch schedule and CPU-provider thread request; different hardware/providers and normal system activity. Not an exclusive hardware laboratory or maximum-core throughput study." }