BitrateCraft · Presets and operations · bitratecraft.org

Disclaimer: Independent technical documentation and configuration guide. Not affiliated with The HandBrake Team. No software downloads hosted

Hardware Encoding and Throughput

Dedicated encoder blocks on modern GPUs and SoCs move the entropy-coding and transform work off the general-purpose CPU. The throughput gain is real, but it is not free: it changes the quality-per-bit curve, the memory profile, and the power budget of the whole machine. This page covers what each hardware family offers, how it compares with software encoders at matched quality, and how the rest of the pipeline scales around it.

Encoder families in practice

Four families dominate desktop and workstation hardware. Intel Quick Sync Video (QSV) has shipped on integrated graphics for many generations and is exposed through the same media SDK on recent discrete Arc parts. NVIDIA NVENC is present on GeForce and Quadro boards and is frequently paired with NVDEC for decode. AMD ships VCN (formerly VCE) in Radeon and Ryzen integrated graphics. Apple's VideoToolbox drives the media engines in M-series silicon. HandBrake can drive all four, and the codec menu changes the available options: H.264 and HEVC are broadly supported, while AV1 encode support is limited to newer generations such as Arc, Ada, and RDNA 3.

Quality equivalence against software encoders

Hardware encoders are fast because they make simplifying assumptions. Typical silicon uses smaller search windows, coarser rate-distortion decisions, and fewer reference frames than x264 or x265. At the same nominal quality setting the hardware output is usually larger for the same visual quality, or visibly softer at the same file size. The practical rule is to budget roughly 20 to 40 percent more bitrate for hardware H.264 to match software H.264 on difficult content, and a similar or larger gap for HEVC.

Constant-quality modes are not comparable across encoders. A CRF value in x265 and a quality value in NVENC share no scale; they must be calibrated per source. Do the calibration once on a representative clip, then reuse the value. For archival work where bitrate matters, software encoding still wins; for delivery masters that must be produced on a schedule, hardware encoding often wins on wall-clock time even after the bitrate penalty.

Thread scaling

Software encoders scale with cores but not linearly. Above roughly eight to twelve threads, frame-parallel scheduling, lookahead, and rate control become the bottleneck, and adding cores yields diminishing returns. The useful curve is a broad plateau, not a straight line. Two machines with sixteen and thirty-two cores may differ by well under a factor of two in finished wall-clock time.

Hardware encoders invert the picture. The encoder block is a fixed-function unit with its own throughput ceiling, so CPU core count stops being the limiting factor and the surrounding pipeline takes over. Decode, scaling, denoise, and subtitle burn-in are all performed on the CPU unless a hardware path exists. On a fast GPU with a slow CPU, filters become the real constraint.

Memory and power

Memory pressure comes from frames in flight, not from the encoder alone. A 4K pipeline holds several 8-bit or 10-bit frames per stage, and each frame is about 12 to 25 MB before conversion. Lookahead buffers, denoise temporal windows, and hardware staging buffers multiply that figure. On a 16 GB machine, a 4K HEVC encode with heavy filtering can consume the majority of available RAM, so close other workloads and prefer a 10-bit intermediate only when the source needs it.

Power draw deserves attention on laptops and dense nodes. Software encoding pins all cores near their sustained clock and produces the highest package power; hardware encoding shifts work to a fixed-function block that can be more efficient per frame. Watch for thermal throttling on thin laptops: the first two minutes may be fast, then clocks drop and throughput falls. For long unattended runs, measure sustained rather than burst frame rate.

Reading the throughput graph

Choose the encoder from the deliverable, not from habit. Archival masters go to software. Scheduled deliveries and large batches go to hardware, with a bitrate bump to cover the efficiency gap. In HandBrake, keep both presets side by side and compare on real footage before committing a queue.