Video streaming now represents 80% or more of the traffic on the internet. According to a recent study by the Boston Consulting Group, the internet is responsible for around one billion tonnes of greenhouse gases a year. It is in everyone’s interests to make a meaningful reduction in these emissions.
It is the mission of iSIZE to address this issue while enabling our clients to provide reliable, efficient and high-quality user experiences, at scale. iSIZE develops AI-based video preprocessing and denoising technologies: BitSave and BitClear. Running real-world tests to accelerate the neural network inferences with the latest 4th Gen AMD EPYC™ CPU, iSIZE achieves a noticeable performance boost compared to previous generation architectures, greatly enhancing the efficiency for end users.
BitSave is a deep perceptual preprocessor that runs before conventional video encoding in order to make encoding significantly more efficient in terms of bitrate at the same or improved visual quality.
BitSave neural networks are able to isolate areas of perceptual importance, such as those with high motion or detailed texture, and optimize their structure so that they are preserved better at lower bitrate by any subsequent encoder.
Being a codec and content agnostic solution, BitSave also allows for the encoder to be used with simpler encoding recipes, thereby making it faster, saving datacenter cost and significantly improving the sustainability of video streaming. It requires no changes to the encoding, delivery or decoding devices and can run in real-time.
BitClear makes unwatchable videos watchable by removing compression artifacts (blurring, blocking, etc.) from user-generated, or heavily compressed content. BitClear uses our unique neural network designs that learn to disentangle the noise from the image data to remove compression noise and recover as many of the original content features as possible.
User-generated content (UGC) that is viewed by millions of people has been re-shared and re-uploaded numerous times, which tends to make it very degraded in quality due to multiple transcoding iterations happening on non-optimized platforms, like mobile phone transcoding for bandwidth-constrained uploads. For social media or UGC distribution and streaming companies, presenting video content in as high quality as possible is vital to keeping audiences engaged to maximize their video asset.
Video content exposed to multiple transcoding iterations can be revived to the maximum possible quality without affecting the original artistic intent of its creators. The process also allows for video upscaling, all with as little as 30ms processing latency on GPUs or high-performance CPUs.
Performance Benchmarks on 4th Gen AMD EPYC™ Processors
Both BitSave and BitClear rely on neural networks. To deploy them at scale, iSIZE leverages the latest features of high-performance x86 CPUs. So naturally, we are always excited by continuing evolutions in x86 CPU architectures that have the potential to accelerate our neural network inference and achieve the best compute efficiency for our customers. That is why we have been tracking closely the development of the AMD 4th Gen EPYC CPUs.
Our customers tend to use our products on CPUs in two ways:
- as a microservice for video on demand (VoD) services, where a small number of cores (typically as little as two) is used per input video segment; here the key performance metric is video processing volume per single-socket CPU
- as preprocessing or denoising solutions for live video broadcast, where the key performance metric is the number of CPU cores needed to support a live 1080p HD channel; lower is better as this means that more live channels can be serviced with a single-socket CPU.
We have evaluated performance with the 4th Gen AMD EPYC under controlled conditions. We selected two most typical scenarios, VoD and live. In VoD environment typically we see that our customers use a minimal amount of cores to perform iSIZE operations, hence we did comparison for 2 cores environment and compared the performance on new generation AMD EPYC versus the previous generation. For live streaming use cases our customers typically measure minimum number of cores that will allow them to stream live channel in any given resolution and our comparison reflects that. Our performance benchmarks showed the following results for BitSave and BitClear:
For VoD use case (Table 1) performance improves between 2.5x and 3x improvement over previous generation EPYC™ on single socket CPU.
For live scenarios (Table 2) that improvement is between 2x and nearly 5x. iSIZE BitSave can allow for nearly 5x more 4k60fps channels on new AMD EPYC CPUs than on the previous generation. Performance numbers are core/frame adjusted – for example for 1080p we only need 3 cores on new AMD EPYC 4th Gen CPUs to achieve 60fps or over (75fps) while we need double amount of cores (6) on previous generation to achieve 60fps or over (67fps).
Similarly, for our BitClear denoising solution, performance improves between 3x and 5x for denoising only model (no upscaling) in the VoD scenario (Table 3). We have selected lower resolution (up to 720p) as this is typically a resolution where our customers are seeing blocking artefacts created by heavy encoding and hence use our solution to help with quality improvement and recovery.
For live 60fps channels (Table 4), our BitClear solution could do live only up to 360p on previous generation AMD EPYC. On the new generation we can now achieve live performance up to 720p. Performance increase, core/fps adjusted for generation over generation of AMD EPYC™ ranges between 4.4x on for higher resolution and 5.3x on lower resolutions.
To conclude, our testing showed when running benchmarks for iSIZE BitSave and BitClear on 4th and 3rd generation AMD EPYC CPUs, performance improves between 2x and 5x. This will allow our customers to add more channels or improve VoD processing density on new generation AMD EPYC.
