All Core Services & GPU Verification Clusters Operational
90-Day Global Uptime: $99.984\%$ • Zero active incidents
Distribution & Licensing Endpoints
90-day SLA metricsmanylinux_2_31_x86_64 & aarch64 binary wheels
ORAS / Docker v2 multi-arch images ($\text{SM\_80}, \text{SM\_89}, \text{SM\_90a}, \text{gfx942}$)
Ed25519 lease signing for commercial multi-node clusters
CGNS / ExodusII validation grids (SHA256 verified)
Hardware-in-the-Loop & GPU CI Runners
Automated nightly Method of Manufactured Solutions ($\text{MMS}$) order-of-accuracy $p_{\text{obs}} = \frac{\ln(E_{h_1}/E_{h_2})}{\ln(h_1/h_2)}$ & bit-exactness verification
| Runner Node | Hardware Target | Observed MMS Order $p_{\text{obs}}$ | $L_2$ Residual Drift | Result |
|---|---|---|---|---|
| zrh-gpu-node04 | $4\times\text{NVIDIA H100 SXM5}$ (NVLink 4) | $4.98 \pm 0.01$ ($\text{WENO5-Z}$) | $0.00 \times 10^0$ ($\text{ULP}=0$) | PASS (14m 12s) |
| zrh-wkst-ada02 | EPYC 9354 + RTX 6000 Ada | $5.02 \pm 0.01$ ($p=4\text{ SEM}$) | $0.00 \times 10^0$ ($\text{ULP}=0$) | PASS (22m 48s) |
| bos-rocm-node01 | $2\times\text{AMD Instinct MI250X}$ | $4.97 \pm 0.02$ ($\text{WENO5-Z}$) | $\le 1.11 \times 10^{-16}$ ($1\text{ ULP}$) | PASS (19m 05s) |
| sgp-hil-rt03 | x86_64 $\text{PREEMPT\_RT}$ (Kernel 6.6-rt) | $|\Delta\mathcal{H}| \le 3.2 \times 10^{-14}$ | WCET Jitter: $6.4\ \mu\text{s}$ | PASS (60m Soak) |
Past Maintenance & Incident Log
Migrated backing ZFS object pools for registry.dophys.top to NVMe RAID-Z2. Container pulls experienced brief read-only failovers to the Frankfurt mirror for 11 minutes between 02:00 and 02:11 UTC.
Addressed a symbol visibility conflict between bundled libnvJitLink.so.12 and PyTorch 2.4 on Ubuntu 22.04 systems. All wheels on pkg.dophys.top were patched and re-signed.