v0.2.0 - Windows preview with automatic PC setup

GPU Prime Generator 0.2.0: Windows preview

A native Windows application and CLI with real CUDA computation, CPU fallback, ordinary text/binary output, and an Optimize for this PC setup flow.

Download and run

• cuda-multi.zip: NVIDIA GPU package with CPU fallback.
• cpu.zip: CPU package, no NVIDIA driver or CUDA toolkit required.
• SHA256SUMS.txt: SHA-256 checksums for both archives.

Extract the whole archive and open bin/primegen-gui.exe. Choose a range and output filename, click Optimize for this PC, and leave the backend on Auto. Text output is the default.

What changed

• Hardware-sized memory reserves and workloads, including smaller-memory PCs.
• Repeated, correctness-checked setup for the selected mode, format and destination drive.
• Measured selection of CPU/GPU settings, one/two compute streams and one/two encoding workers.
• Local tuning shared between GUI and CLI, preserved when setup is cancelled or entry GPU load makes measurements unreliable.
• Native CUDA targets 75, 80, 86, 89, 90, 100 and 120, plus compute_120 PTX. Dedicated sm_86 builds remain available from source.
• Direct .txt and raw little-endian .bin generation, plus compressed native .gprime datasets and checkpoint resume.
• A clear completion banner, frozen job timers, and worker cleanup before completion.
• A rewritten README with an actual application screenshot, first-use instructions, output guide and hardware support table.

Actual validation

On Windows 11, RTX 3090 Ti and i9-12900K: all 13 CUDA Release tests passed. All 12 applicable CPU Release tests passed, with the GPU test explicitly skipped. MSVC AddressSanitizer passed all 10 applicable tests, with one GPU skip.

Freshly extracted CPU and CUDA archives passed native GUI, setup/profile reuse, text/binary output, pause/resume, completion-clock, no-device fallback and process-cleanup checks. Every line of an independently checked 100-million-integer text run matched: 5,761,455 primes, 51,099,000 bytes.

Focused setup tested 11 CPU/CUDA configurations with three samples each in 8.85 seconds. For sampled-exact direct text output through 100 million, CPU with two encoders won at a 101.83 ms median, versus 133.04 ms for the best tested CUDA configuration. These are measurements on this PC, not promises for other hardware.

Clean GitHub-hosted Windows and Ubuntu 24.04 CI also passed for the release commit, including native Qt controller/setup tests and Linux AddressSanitizer/UndefinedBehaviorSanitizer. View the release CI run.

Preview limits

Other physical GPU/CPU models, driverless Windows, CachyOS, and Linux GPU/package execution remain unverified. A compiled target is not physical validation. AMD/Intel GPU acceleration and Apple-platform support are not implemented; compatible Windows PCs can use the CPU engine.

The NVIDIA package uses CUDA 13.2 and requires a compatible NVIDIA driver. The app does not install drivers or change power/watchdog settings. Keep output and checkpoint files together for resume. See the README and included third-party notices.