Novabench logo

Benchmarks

GPU benchmark

The GPU benchmark evaluates your graphics card's performance across 3D rendering, general-purpose GPU (GPGPU) compute, and VRAM throughput. The combined result produces your GPU score, which reflects how your graphics hardware handles demanding visual and computational workloads.

What the GPU test measures

The GPU benchmark runs three test types that exercise different aspects of graphics hardware.

3D rendering

The 3D rendering test measures your GPU's ability to draw complex scenes in real time. Novabench uses the native graphics API for your platform:

  • Windows: Direct3D 11
  • macOS: Metal
  • Linux: Vulkan

The test renders a 3D scene and measures per-frame render time throughout. The rendering workload includes geometry processing, texture sampling, lighting calculations, and shader execution. These reflect the types of work GPUs perform in games and 3D applications. Novabench derives the score from average frame time. Lower frame times (faster rendering) produce higher scores.

GPGPU compute

The compute test measures your GPU's ability to perform general-purpose parallel calculations. Unlike 3D rendering, compute workloads do not involve drawing pixels on screen. Instead, they use the GPU's parallel architecture for tasks such as scientific simulation, machine learning inference, video encoding, and image processing.

Novabench runs Vulkan compute shaders (MoltenVK on macOS) that execute large chains of fused multiply-add (FMA) operations across thousands of parallel work items. Novabench reports the result in GFLOPS (billions of floating-point operations per second). This reflects raw GPU compute throughput.

VRAM throughput

The VRAM test measures memory bandwidth by transferring 128 MB buffers in repeated passes. Novabench measures two transfer directions: host-to-device (system memory to GPU memory) and device-to-device (within GPU memory only). It reports results in MB/s. VRAM throughput affects how quickly the GPU can access textures, frame buffers, and compute data during rendering and compute workloads.

How the score is calculated

Novabench combines the results from the three workload types (3D rendering, GPGPU compute, and VRAM throughput) into a single GPU score using a geometric mean. This method prevents any one workload type from dominating. A card that excels at one type of operation but lags in another produces a balanced score that reflects overall capability rather than a narrow strength.

Each workload produces its raw result on a different scale. Frame time in milliseconds, compute throughput in GFLOPS, and memory throughput in MB/s are not directly comparable numbers. Before it combines them, Novabench applies a per-workload scaling step calibrated against reference hardware, so the values sit on comparable magnitudes. The geometric mean then combines them. Each workload contributes equally, so no single workload dominates the final score because of raw magnitude differences.

Validation

Before each release, Novabench measures GPU score consistency across supported platforms and a range of integrated, discrete, and external GPUs.

A note on single-score benchmarks

There is no such thing as a fully objective single-number GPU benchmark. Every score reflects choices about which workloads to include, how to weight them, and which graphics APIs to use. Different choices produce different scores, and a card that wins one benchmark can lose another depending on what each one emphasizes.

Novabench's choices make scores generally useful for comparing GPUs. The mix of rendering, compute, and memory throughput is balanced, and each contributes equally in the geometric mean. The 3D test uses the platform's native graphics API, so results reflect what games and applications actually run. Compute and VRAM use Vulkan, so those results compare cleanly across operating systems. Those choices favor general-purpose graphics workloads (most games and creative applications mix all three), and they don't try to capture every architectural strength in the headline number.

If your workload doesn't match those assumptions, the headline score does not tell the full story. Novabench reports the per-workload sub-scores alongside the overall score for exactly this reason. A user who runs heavy 3D applications can look at the rendering result. Someone who runs machine learning inference or video encoding can look at the compute sub-score. Someone who works with very large textures or datasets can look at VRAM throughput. The headline is a summary; the breakdown shows how the GPU handles your work.

Multi-GPU systems

If your system has more than one GPU, Novabench lets you select which GPU to test. This is common on systems with both an integrated GPU (built into the processor) and a discrete GPU (a dedicated graphics card), or on workstations with multiple discrete GPUs.

Select the GPU you want to test from the GPU test configuration widget before you start the benchmark. Novabench identifies each GPU by name and confirms which device ran the test in the results.

Frame timing analysis

During the 3D test, Novabench tracks how long each frame takes to render, and how much time the CPU and GPU each spend waiting for the other. The analysis identifies stutter and potential bottlenecks between CPU and GPU. Stutter is a frame that takes significantly longer than normal to render, which appears as a noticeable gap in animation.

What bottleneck analysis tells you

  • GPU-limited: the GPU is fully used while the CPU has idle time. A GPU upgrade improves performance.
  • CPU-limited: the CPU cannot feed the GPU fast enough, so the GPU has idle time. This can happen with a powerful GPU paired with an older or slower CPU. A CPU upgrade, or closing CPU-heavy background tasks, helps.
  • Balanced: both the CPU and GPU are well used, with minimal idle time on either side. This is the ideal configuration.

Bottleneck analysis is most useful when you evaluate upgrade decisions. If your system is heavily CPU-limited, a GPU upgrade alone does not necessarily improve real-world graphics performance.

Cross-platform comparability

Novabench runs the same GPU workloads on every supported platform: Windows, macOS, and Linux. The 3D test renders the same scene through each platform's native graphics API (Direct3D 11 on Windows, Metal on macOS, Vulkan on Linux). These are the APIs that games and 3D applications generally use on each platform. The compute and VRAM tests run on Vulkan (via MoltenVK on macOS) so those results compare directly across operating systems.

How the benchmark runs

The GPU benchmark is designed around three priorities: consistent measurement, fair comparison across systems, and a balanced top-level score backed by full per-test detail. Several mechanics support those priorities:

  • Warmup and calibration: each test runs for a set window. Novabench measures how much rendering, compute, and memory throughput the GPU delivers in that time. A fast card and a slow card then both produce a meaningful amount of work to compare.
  • Process isolation: each test runs in its own worker process, separate from the Novabench app. This lets the benchmark control exactly how the workload runs on the GPU. The app's own UI, logging, and sensor sampling do not interfere. It also stops the previous test from influencing the next one.
  • Thermal gaps between tests: a brief cooldown between tests lets the GPU recover, to reduce the thermal impact of prior tests on later ones.
  • Hardware detection and GPU selection: every result captures GPU model and driver version. If your system has more than one GPU, you pick which device to test. When you compare your score to others, you're comparing the same chip in the same configuration.
  • Frame timing and bottleneck analysis: during the 3D test, Novabench tracks per-frame render time alongside CPU and GPU wait time. The result includes both an overall score and a clear picture of where time is spent.

Top-level scores are a summary. Novabench always presents per-test results alongside the overall score, so you can weigh individual workloads against your own use case.

Sensor data during the test

On Plus, Novabench collects sensor data while the GPU benchmark runs:

  • Temperature: GPU core temperature throughout the test, which shows thermal behavior under sustained graphics load
  • Power draw: watts consumed by the graphics card during each test phase
  • Clock speed: GPU core frequency, which shows whether the card maintains its boost clock or throttles under load

A temperature curve that rises sharply during the 3D test and then levels off suggests the GPU reached its thermal limit and throttled. Lower-than-expected scores combined with high temperatures point to a cooling problem. See sensor monitoring for long-term GPU thermal tracking.

Factors affecting GPU scores

Driver versions

GPU drivers have a significant impact on graphics performance. Driver updates frequently include optimizations that improve rendering throughput and compute performance.

  • Windows: update through NVIDIA GeForce Experience, AMD Software, or Intel Arc Control
  • macOS: GPU drivers are included with macOS updates
  • Linux: install the latest proprietary drivers from your GPU vendor for best results

Power and thermal conditions

  • Power delivery: desktop GPUs require adequate power supply wattage. An underpowered PSU can cause the GPU to throttle or crash under load.
  • Thermal throttling: GPUs reduce clock speed when they overheat. Make sure that your case has adequate airflow. On laptops, use a cooling pad. Also make sure that the vents are unobstructed.
  • Power plan: on laptops, power-saving modes can limit GPU performance. Use balanced or high-performance power settings when you run a benchmark.

Display and configuration

  • Resolution: the GPU benchmark uses a fixed resolution, so your display resolution does not affect the score. Other graphics applications can affect the result.
  • Multi-monitor setups: additional displays consume GPU processing cycles for desktop rendering. This typically has no significant impact on test scores. If you're having performance problems, it is still a factor worth isolating.

Hardware factors

  • GPU architecture and generation: newer GPU architectures deliver more performance per watt and per clock cycle.
  • VRAM amount and type: more VRAM lets the GPU store larger textures and data sets without swapping to system memory. VRAM type (GDDR5, GDDR6, GDDR6X, HBM) determines bandwidth.
  • Memory bus width: wider memory buses increase bandwidth. This is a fixed hardware characteristic that varies by GPU model.