Benchmarks overview

Novabench tests five hardware components, each producing a score, with hardware scores combining into an overall Novabench Score that is designed to be comparable across all platforms and configurations.

This page summarizes each benchmark type. Follow the links to individual benchmark pages for details on how each test works.

Hardware benchmarks

CPU

The CPU benchmark measures processor performance across single-core and multi-core workloads. The test runs integer, floating-point, SIMD, hash, and compression operations in both single-threaded and multi-threaded modes.

• Single-core score: per-core speed, important for applications that rely on single-threaded execution
• Multi-core score: parallel throughput across all cores, important for rendering, compilation, video editing, and scientific workloads

GPU

The GPU benchmark evaluates graphics card performance through three test types:

• 3D rendering: real-time scene rendering using the platform's native graphics API (Direct3D 11, Metal, or Vulkan)
• GPGPU compute: general-purpose parallel computation using compute shaders
• VRAM throughput: video memory bandwidth between GPU cores and VRAM, and the host device to the GPU

On Plus, the GPU test includes bottleneck analysis that identifies whether your CPU or GPU is the performance-limiting factor during 3D rendering.

Memory

The memory benchmark measures RAM performance in two ways:

• Transfer bandwidth: how quickly data moves between the CPU and memory (read/write speed in GB/s)
• Latency: how long the CPU waits for data across the cache and memory hierarchy

Storage

The storage benchmark tests your drive's read and write speeds for two access patterns:

• Sequential I/O: large contiguous read/write operations (file copying, media loading)
• Random I/O: small scattered read/write operations (app launching, OS tasks)

NPU

The NPU benchmark measures neural processing unit performance on systems with supported hardware. The test measures:

• Throughput (TOPS): tera operations per second
• Inference latency: time to process a single neural network input

The NPU score appears only when compatible hardware is detected. Systems without an NPU are scored on the four core components without penalty.

Benchmark methodology

All Novabench tests use fixed-duration measurement: the test runs for a set time and measures how much work the hardware completes. This design ensures that test durations stay predictable. Warm-up phases reduce between-test variability, and thermal gaps between test types help prevent thermal throttling from impacting tests that happen later in the test suite.

While one test iteration is designed to measure system capabilities with minimal variance, multiple iteration tests can be enabled for testing that requires more precision, at the expense of longer test times.

Scores are comparable across Windows, macOS, and Linux.

For a deeper look at how cross-platform comparability works, how scoring is calibrated, and how sensor data is collected during tests, see benchmark methodology.

Sensor data

Novabench records temperature, power draw, clock speed and other sensor data during benchmarks. Sensor data helps diagnose performance issues like thermal throttling or power delivery problems. See sensor monitoring for details on real-time and historical sensor tracking.

Choosing the right test

Related pages

• Getting started: run your first benchmark
• Understanding your scores: interpret your results
• Comparing results: benchmark against other systems
• Benchmark methodology: how the tests work under the hood