How AI Upscaling Complicates Computer Hardware Benchmarking

Artificial intelligence upscaling technologies, such as Nvidia's DLSS, AMD's FSR, and Intel's XeSS, have revolutionized gaming performance, but they have also thrown the traditional hardware benchmarking process into chaos. By rendering games at lower resolutions and using algorithms to reconstruct images, coupled with frame generation techniques, these tools make direct, apples-to-apples comparisons between graphics cards nearly impossible. This article explores how variable internal resolutions, proprietary hardware requirements, subjective image quality, and artificial frames have fundamentally complicated the way reviewers and consumers measure computer hardware performance.

The End of Apples-to-Apples Comparisons

Historically, benchmarking a graphics card was straightforward: reviewers ran a game at a native resolution like 1080p, 1440p, or 4K and recorded the frames per second (FPS). Upscaling breaks this standard.

When DLSS (Deep Learning Super Sampling) or FSR (FidelityFX Super Resolution) is enabled, the GPU renders the game at a lower internal resolution and scales it up to fit the monitor. Even though these technologies use standardized preset names like "Quality" or "Performance," the underlying algorithms work differently and utilize different hardware components. Consequently, comparing a GPU running DLSS against another running FSR is no longer a strict test of raw hardware power; it is a test of competing software algorithms.

The Subjectivity of Image Quality

Before upscaling, a frame rendered on an AMD card looked virtually identical to a frame rendered on an Nvidia or Intel card. Today, higher frame rates often come at the cost of visual fidelity.

Upscalers can introduce visual artifacts such as ghosting, shimmering, and blurring. Because Nvidia's AI-driven DLSS typically retains better image quality than AMD's spatial/temporal FSR at lower resolutions, two cards producing the exact same FPS with upscaling enabled are not delivering the same user experience. Benchmark bar charts only show numerical performance, failing to capture the degradation in visual quality required to achieve those numbers. Reviewers must now perform subjective visual analyses alongside objective frame-rate testing.

Frame Generation Distorts Performance Metrics

The introduction of frame generation in DLSS 3 and FSR 3 has further muddied the waters. These technologies use AI and motion data to insert entirely artificial frames between traditionally rendered ones.

While this makes the game look smoother to the eye and drastically increases the reported FPS on a benchmark chart, these generated frames do not respond to user input. In fact, frame generation adds processing overhead that can increase input lag. A game running at 120 FPS with frame generation might have the slower input responsiveness of a game running at 50 FPS. Therefore, traditional FPS counters are no longer a reliable indicator of how playable or responsive a game actually feels.

Vendor Exclusivity Creates Testing Dilemmas

Hardware reviewers face a difficult philosophical choice when setting up benchmarks. Nvidia's DLSS requires proprietary Tensor cores, meaning it only works on Nvidia GPUs. AMD’s FSR and Intel's XeSS are open-source and work on almost all modern graphics cards.

If a reviewer tests all cards solely at native resolution to keep the test fair, they ignore a massive, proprietary performance feature that buyers of Nvidia cards will heavily rely on. Conversely, if they enable DLSS for Nvidia and FSR for AMD, the test becomes scientifically flawed because the workloads and resulting image qualities are no longer identical.

Shifting the Bottleneck to the CPU

Upscaling also alters how different components within a computer interact during a benchmark. When testing a powerful GPU on a 4K monitor using an upscaler's "Performance" mode, the graphics card is actually rendering the game internally at 1080p.

At 1080p, modern high-end GPUs process frames so quickly that the computer's central processor (CPU) often cannot keep up. Instead of measuring the maximum potential of the graphics card, the benchmark inadvertently measures the limits of the CPU. This forces reviewers to run exponentially more tests—covering native resolutions, multiple upscaling presets, and frame generation toggles—significantly increasing the time and complexity required to accurately evaluate a single piece of hardware.