NETINT Argues for ASICs Over GPUs for Video Encoding
Hardware firm NETINT is making the case that specialized ASICs, not general-purpose GPUs or CPUs, are the core infrastructure for video encoding at scale. The company's position challenges the dominant narrative that GPUs are the solution for all heavy compute tasks. They argue that purpose-built hardware offers superior efficiency for high-volume video workflows.
The push for ASICs over GPUs in video encoding is fundamentally an argument for specialization over generalization, echoing similar shifts in high-performance computing. While GPUs are versatile, a significant portion of their hardware is dedicated to graphics or other functions, not video transcoding. ASICs, or Application-Specific Integrated Circuits, are designed exclusively for video processing, leading to higher efficiency in speed, power consumption, and density. NETINT's Codensity line, built on their own ASICs, claims significant performance gains over traditional methods. For example, their T432 Video Processing Unit is said to offer a 10x increase in encoding density and a 20x reduction in carbon footprint compared to CPU-based software encoding. This focus on "watts per stream" is becoming a critical metric as data centers face increasing power constraints. The economic argument hinges on the total cost of ownership (TCO). While the initial investment in specialized ASIC hardware can be higher than for general-purpose CPUs, the long-term savings in power and rack space can be substantial for large-scale operations. One analysis showed that for the HEVC codec, ASICs saved 94% on capital expenditures and 98% on operating expenses compared to CPU-based transcoding. This trend toward specialized hardware isn't unique to video encoding. Google, for instance, developed its own custom "Argos" ASICs for YouTube's vast video transcoding needs, reporting a 20 to 33-fold increase in efficiency over their previous server setups. Meta has also developed its own encoding ASIC, indicating a broader industry recognition of the benefits of purpose-built silicon for massive video workloads. The development of more advanced codecs like AV1 further complicates the hardware decision. While software-based encoding on CPUs can often yield the highest quality, it is not always feasible for real-time, high-volume streaming. NETINT's newer Codensity G5-based products, which they term Video Processing Units (VPUs), are designed to support AV1 and also include onboard Neural Processing Units (NPUs) for AI-based video processing tasks. Ultimately, the choice between ASICs, GPUs, and CPUs depends on the specific use case. CPU-based solutions offer the most flexibility for frequently changing codecs and smaller-scale operations. GPUs provide a balance of performance and flexibility. However, for high-volume, consistent encoding tasks, ASICs present a compelling case for superior performance and lower long-term costs.
