How fast is a verified web server?

• 100,000 bookmarks, 2,000 tags, 50 users. Simple and synthetic eight-term searches each match 1,000 active bookmarks; the first page returns 20.
• Dual EPYC 7502, RTX 3090. Four concurrent variants, each with 4 server, 4 PostgreSQL and 4 client CPUs. Shared host; clocks and NUMA placement vary.
• Three repetitions, fresh database/server, 10 s warmup, 30 s measurement. Index initialization is in warmup; mixed traffic includes rebuilding after writes. 128,022 measured requests; zero failures. Cost, power and cold start were not measured.

• 100 users, 10,000 articles, 10,000 comments, 1,000 follows, 5,000 favorites. Three repetitions, fresh database/server, 10 s warmup, 30 s measurement; startup excluded.
• Dual EPYC 7502, two RTX 3090s. Two CPU/GPU pairs split the scenarios; each variant has 4 server, 4 PostgreSQL and 4 client CPUs. Shared host; clocks and NUMA placement vary.
• 4,467,960 measured requests; zero failures. GPU execution was logged with no runtime CPU fallback. Paired runs cover 1 and 16 clients; cost and power were not measured.

• EPYC 7502, CPUs 64–79, RTX 3090 device 0, CUDA 12.6. Eight warmup passes; three repetitions, each averaging 100 iterations below 4,096 carts and 10 otherwise; shared host and uncontrolled clocks/NUMA placement.
• GPU snapshots are transposed on upload for contiguous reads; upload-each timing includes that conversion. The same integer pricing source runs on CPU and CUDA using h5i's NVVM launch shell. CPU 16 uses persistent Rayon workers. GPU resident timing includes launch, synchronization, result allocation, download and error validation.
• 4,146 cases agree with the pinned upstream's Decimal methods; 49 exercise its ORM aggregate. Rust, Go, Node.js and CUDA agree with this corpus. CPU/GPU checks also cover malformed snapshots and batch boundaries. This pricing kernel has no Lean proof.
• None, percentage and fixed discounts, final half-up rounding, default zero tax/shipping. Fixed-width snapshots reserve 32 lines; no mutations or live cache invalidation are measured.

Every path retains rows, subtotals and totals between batches. Canonical commands contain explicit u64 query IDs and four u64 fields per replacement. Timings include applying updates, pricing, worker dispatch/join and a complete freshly owned output. GPU paths also include encoding where applicable, upload, synchronization and full download. All returned values are checked exactly outside timing. No SQL/HTTP is included.

One or four persistent computation workers; Python 4 uses fork processes, JavaScript 4 uses worker threads, Go 4 uses goroutines, Rust 4 uses Rayon. NumPy is a separate native vectorized implementation with native thread counts fixed to one. All modes share the same four-core allocation (CPUs 64–67), with GC enabled. CPU timings are collected without concurrent benchmark processes. If GPU references from the prior run are shown, their status is explicit above and their ratios are not contemporary speedups. JavaScript uses exact Float64 integer subtotals and quotient/remainder discount rounding. CPython loops use Python integers with u64 retained storage; NumPy is used for untimed input validation in every Python mode.

Initial placement, loading, worker creation and validation are excluded. Python imports initial subtotal caches from fixtures, computes initial totals itself, and independently rescans subtotals for every upstream gate; other languages construct their initial caches. Initialization timings are not comparable cache-construction costs. Three repetitions, eight changing warmups and 50 changing timed rounds; RTX 3090 device 0. Shared host; clocks, NUMA and cache placement are uncontrolled. This is a resident calculation/update benchmark, not a Django deployment speedup.