Home  ¦ About  ¦ Disclaimer

Intel J3355 - A Gem in the Rough for H.264 and H.265 Encoding

alt

This last decade the world of desktop processors has been, well, boring. The Intel Core processors which launched in 2006 were a significant step forward. They did away with the worst of the Pentium 4 series and moved to a much higher IPC, resulting in a processor that could outperform the Pentium 4s using a significantly lower clock rate while using significantly less power. Those were the days of revolutions. Nowadays, those leaps and bounds are reserved for the mobile space (although those days are coming to an end before long). On the desktop, we have only seen (limited) incremental progress as better core designs are introduced along decreasing fab sizes. The problem is that we are getting closer and closer to a wall, the fabrication size, that cannot easily be broken. Making the fabrication process one step smaller comes with an exponential increase in costs, while only offering a modest increase in performance. In the everyday world this translates itself in a humdrum evolution of the desktop processor. A processor bought 5 years ago is often easily up for the challenge of today’s workload. Yet to find a way to buck this trend, processor manufacturers are increasingly expanding the abilities of their CPUs. These abilities can take various shapes, such as the integrated graphics processors and/or dedicated circuits for common functions such as encryption or video encoding.

The ugly duckling in this story is the Intel Atom processor. Conceived as a low power, low cost processors around the same time the Core processor was introduced, these Atom processors ended up being perceived in a very different way. Due to their low cost and power requirements, they ended up in some of the most dreaded devices of the last decade, including nettops, netbooks, and mobile internet devices. In recent years, Atom processors found their way in low-cost Windows tablets where they benefited from x86 instructions sets unlike their crippled Windows RT brethren. Still, people perceived these Atom devices as (unworkably) slow, not worth the money, and typical examples of “you get what you pay for”. While partially true, the modern day Atom CPUs were not to blame. The real culprits were the horrendous components they were paired with. Minimal amounts of memory, laughable eMMC storage (typically no faster than a cheap SD card), and ancient USB2 connections meant that these devices felt sluggish even on the best of days. It thus comes as a surprise to many that, if you combine these Intel Atom processors with proper components, you end up with a rather capable system. Due to their low-power and cost inheritance it is possible to build them into reasonably cheap system (£250) that are entirely without moving parts (and hence truly silent). A unique mix of traits makes such builds rather exceptional. They are cheap to buy, economical to run, absolutely quiet, and entirely maintenance-free (as no fans means no dust drawn into the case). But how do they perform, and are they powerful enough? As it turns out, the very same dedicated circuits I talked about earlier make these Intel Atom processors true swans.

The new Apollo Lake generation of Atom processors are an excellent example of this. On paper, they look rather dull with a small number of cores and low CPU frequencies, for a low price. What does stand out is their power requirement, with a rather interesting 10W TDP. This is the sweet spot, as a CPU with a 10W TDP or less can be integrated on motherboards without any form of active cooling. Hidden away in the details is another juicy tidbit: the 10W package includes a Gen9 graphical processor. This is the same generation as found in Intel Core processors from the Haswell generation, and it means they come with a long levy of dedicated circuits for decoding/encoding. Notably, it means they come with full hardware acceleration for H.264 and (8-bit) HEVC/H.265 decoding/encoding. While these CPUs are thus not the fastest, they do stand out as excellent devices for playing and converting (home-)movies.

The ASRock J3355B motherboard with an integrated Intel J3355 Apollo Lake processor. The motherboard I ended up purchasing is the ASRock J3355B. As you would expect, this is a barebones board. It comes with a J3355 CPU soldered on, 2 SODIMM DDR3 slots, 4 SATA connectors, and not much else apart from a few USB ports, HDMI, and a few other input/output. I paired the board with 2x 2GB DDR3 memory, a 64GB SSD for booting, and a 240GB SSD for storing files. The entire system is powered by a Chieftec 90W power supply, specifically designed for low-power mini-ITX systems. The operating system is Windows 10 (as QuickSync in Handbrake is so far only supported in Windows). That’s it. There really isn’t anything more to this build.

The purpose of this system in our household is to convert Blu-Ray discs for viewing on all of our devices. None of these devices even have a disc player, which makes the conversion step a necessity. They do, however, all have excellent support for H.264 playback. Cue MakeMKV and Handbrake which are used to rip the Blu-Rays and store them locally for easy access. This allows the files to be easily played on the media player connected to our TV, or taken along on our phones and tablets while travelling abroad. The key when it comes to these Blu-Ray conversions is that they have to be as straightforward as possible, and as fast/economical as possible, while maintaining visual quality that is acceptably close to the original Blu-Ray. The software helps ensure that the whole process is reasonably straightforward. Thanks to Intel QuickSync and the Intel Atom Apollo Lake CPU I can now also tick the boxes for fast, economical, and of a high quality.

So how does it perform? Thanks to the SSDs the system is quick to boot and offers excellent responsiveness. Compared to the J1800 that preceded it, the J3355 definitely feels a lot more responive as well. Purely subjectively, I would be happy to use this system as my everyday PC. While it may not have the raw CPU power for demanding tasks, it handles web browsing, music playing, programming, and whatever else I do on a daily basis with ease. Since I bought the J3355, which is 5 weeks ago, it never felt slow or made me wait. Where it really shines though, is in converting those Blu-Rays.

converting Pixar short movie Piper

For the first test, the Piper short movie from the Pixar Animation Studios was used. The speed and size results can be found in the table below. The baseline uses a software x264 encoding, while Intel QuickSync (QSV) results are shown for Quality/Balanced/Speed settings for H.264 and H.265.

methodspeed (in fps)size (in MB)
x264 baseline 14.67 fps217.1 MB
QSV H.264 Quality 218.84 fps216.5 MB
QSV H.264 Balanced 354.78 fps284.5 MB
QSV H.264 Speed 357.59 fps284.8 MB
QSV H.265 Quality 48.29 fps135.5 MB
QSV H.265 Balanced 411.65 fps136.2 MB
QSV H.265 Speed 441.33 fps148.5 MB

On a subjective level, the quality of the output files is close to indiscernible. To highlight the differences, particularly troublesome still frames were extracted and cropped to 700px width. Frames were only extracted for the x264 baseline, the QSV H.264 Quality, and QSV H.265 Speed. All QSV H.265 files showed similar visual quality for the frames in question. The QSV H.264 showed more detail for the Balanced and Speed setting, in line with their 30% increase in file size.

First we compare the x264 baseline (left) with the QSV H.264 Quality encoding (right). Details in the stone wall are noticeably more absent for the QSV H.264 Quality encoding, while the x264 baseline encode creates more artefacts around the pink clouds and spires in an attempt to retain detail.

When comparing the x264 baseline (left) with the QSV H.265 Speed encoding (right) similar conclusions can be drawn. Details are lost on the stone wall for the QSV H.265 Speed encoding. More artefacts around the spires are visible for the QSV H.265 than for the QSV H.264 encoding.

For the same frame, we also compare the QSV H.264 Quality encoding (left) with the QSV H.265 Speed encoding (right). As indicated above, slightly more artefacts occur around the spires for the H.265 encode. At the same time, the H.265 encode retains slightly more detail when looking at the wall texture.

Overall, banding doesn’t seem to be a big issue. This is an area where previous iterations of Intel QuickSync often performed poorly (especially in earlier frames). The difference in quality is hard to notice when the movie is playing, and whether you prefer more detail (at the cost of more artefacts) are a smoother overall look mostly is up to personal preference.

We now once again compare the x264 baseline (left) with the QSV H.264 Quality encoding (right) for another frame from the Piper short movie. This frame has a lot of small detail that again gets smoothed out in the QSV H.264 encode. The sand (particularly on the right) is less clearly outlined, while the detail on the orange shell has less noise.

When comparing the x264 baseline (left) with the QSV H.265 Speed encoding (right), it’s clear that the QSV H.264 encode preserves more detail. The QSV H.265 encode generates slightly more noise in the dust clouds on the right side of the image.

Overall, the results indicate a very good performance of the Intel QuickSync hardware encoder. While the QuickSync encoder is characterised by its softness (more so for the H.264 encoder), it does not introduce undesired artefacts and overall provides results that are very close to the x264 baseline. Given the significant increase in speed (4-12x faster, depending on setting) and the significant decrease in file size (62-68% of the size when using H.265) this seems like an ideal balance between speed and detail retention for archival purposes.

converting Planet Earth II episode

Next up, I also looked at the conversion of the first episode of the Planet Earth II season from BBC. As with the Piper short film, the speedup is impressive and the results are close to the baseline. This time around, the ICQ rate control algorithm used for the Speed and Balanced setting for the QSV H.264 method results in a smaller file size than the LA-ICQ rate control algorithm used in the Quality setting. Otherwise, the results are entirely in line with the one from the Piper short movie.

methodspeed (in fps)size (in GB)
x264 baseline 13.42 fps4.90 GB
QSV H.264 Quality 220.71 fps4.69 GB
QSV H.264 Balanced 351.26 fps4.51 GB
QSV H.264 Speed 353.96 fps4.51 GB
QSV H.265 Quality 47.92 fps2.77 GB
QSV H.265 Balanced 411.29 fps2.78 GB
QSV H.265 Speed 441.35 fps2.97 GB

This time around, I compare the x264 baseline (left) with the QSV H.264 Speed encode (right). The results are so close as to be indistinguishable.

The same holds when comparing the x264 baseline (left) with the QSV H.265 Quality encode (right). It is again rather hard to notice any differences.

One of the few frames that does cause potential “issues”, is one where the x264 baseline (left) again tries hard to maintain detail, whereas the QSV H.264 Speed encode (right) clearly softens the resulting frame. However, from a purely qualitative perspective, the results are again very close and mostly down to preference.

Wen comparing the x264 baseline (left) with the QSV H.265 Quality encode (right), the same results from the Piper short film encode are verified. The QSV H.265 encoder is better at maintaining detail than the QSV H.264 encoder, although is still slightly softer than the x264 baseline result.

The result after comparing these frames is again too close to call. The speedup obtained from the Intel QuickSync hardware encoder is again impressive and barely slows down, whereas the baseline encode was 27% slower. It again appears clear that the tradeoff made by the Intel QuickSync hardware encoder between speed and detail retention is the right one. Moreover, during these hardware encodes the processor temperature never rised more than 28°C above average and the processor was idling nearly all the time. Clearly, the J3355 satisfies the desire for a high-quality, efficient H.264/H.265 encoding box.

conclusion

The Asrock J3355B is a simple and cheap motherboard without frills. It is powered by the equally mundane Intel J3355, an Atom processor from the Apollo Lake generation. What sets the board and processor apart is the use of the 9th Generation of Intel graphics processor. The Intel QuickSync support on this GPU is impressive, offering results so close to the output of the x264 encoder that they are nearly indistinguishable. Thanks to the tremendous boost in speed from using dedicated hardware to do the encoding, and the surprisingly low energy consumption that allows for an entirely silent system, the end result is an impressive box for video encoding. The processor is cheap, power costs are low, and the results are good enough (if not excellent) for encoding Blu-Rays for use on a variety of devices.


  1. For software H.264, the settings used are: m4v (mp4) container, no audio, no subtitles, no chapter markers, no filters, Medium quality setting, RF value of 20, High@4.2, tune for Film. 

  2. For hardware Intel QuickSync H.264, the settings used are: m4v (mp4) container, no audio, no subtitles, no chapter markers, no filters, High quality setting, QP value of 26, High@4.2. 

  3. Since the Balanced and Speed setting revert to ICQ over ICQ-LA, the QP value chosen for these settings is 22 instead of 26 to better match the output size. 

  4. For hardware Intel QuickSync H.265, the settings used are: m4v (mp4) container, no audio, no subtitles, no chapter markers, no filters, High quality setting, QP value of 23, Auto@Auto.