HP Mini 311: Netbooks get ION

HP Mini 311 — CUDA on ION

NVIDIA always likes to tout the advantages of their GPUs and CUDA. In theory, if you have a highly parallelizable task like video encoding/decoding/processing, you can accomplish that task a lot faster on a GPU than on a traditional CPU. When you add in the fact that Atom is already a very slow in-order architecture (albeit with Hyper-Threading), boosting performance of certain tasks could be very beneficial.

One of the applications NVIDIA talks about a lot with relation to CUDA is Badaboom. We haven't looked at in quite some time, and with the low performance of Atom CPUs it might be a viable option. We ran a test encode using Badaboom and a 720p source video, targeting the default "YouTube" output: 640x360 with 1000kbit VBR video and 128kbit MP3 audio. We also tweaked the Badaboom settings to enable auto key frame insertion for improved image quality and performance, which we'll see in a moment.

We can't run Badaboom on CPUs, unfortunately, so we used TMPGEnc 4.0 Xpress and similar settings: MPEG4 1000kbit video with 128kbit AAC audio. Interestingly, TMPGEnc also supports CUDA. However, the encoding is all done on the CPU, so the main benefit of CUDA for TMPGEnc is if you use some of the complex filters like noise reduction and smart sharpen. If you're just doing a straight video encode, CUDA doesn't appear to help TMPGEnc right now.

First up, here are the results, with CPU encoding time on an Acer Timeline 1810 as well as a desktop Core 2 Quad 3.2GHz thrown in for reference.

Badaboom ends up being about 2.5 times faster than doing the video encoding on the Atom N270. Hooray! It's also about the same speed as a Pentium SU4300, but an overclocked Kentsfield 3.2GHz CPU is still 2.8 times as fast. Can you do basic video encoding on the ION LE with Badaboom? Yes, but now let's discuss quality for a moment.

The highest quality encodes will always require multiple passes in order to provide a better analysis of how to best use the available file size. Badaboom doesn't support multiple passes, so we did a single-pass encode in TMPGEnc for comparison. File sizes are similar (4213KB TMPGEnc vs. 4205KB Badaboom). Below are the image captures of the encoded video.

Default Badaboom YouTube setting

Badaboom YouTube setting with Auto Key Frames

TMPGEnc 4.0 Express — Single pass MPEG4 AVC 1000kbit VBR (3000kbit max)

The initial encodes using the Badaboom YouTube defaults are, at times, horrific. The content of the source video (and the scene in particular) can make a big difference, but overall the default settings fail to produce optimal results. Fast motion content is the worst offender, but then doing a good compression of fast action content has always been a lot more difficult. Luckily, we determined that making one change — setting the Key Frame Mode to "Auto" instead of "Fixed" (with the default "Baseline" and not "Main" mode) — results in a much better image. We still give the quality lead to TMPGEnc — and if you do multiple passes, the gap widens — but as a compromise the CUDA-based Badaboom encoder can achieve some decent results. As for ION, it's not really fast enough to make CUDA compelling; a moderate CPU can easily produce better results. Move to a faster GPU (NVIDIA's GTX 260M does the same encode at around 82 FPS and GTX 280M is at 100 FPS) and CUDA can flex its muscles.

There are other concerns with Badaboom, however. We had quite a few sample videos that it wouldn't accept (including a standard MPEG2 video). We also had some issues with audio glitches — our test encoded video has a couple small gaps in the audio, and they're there every single time on multiple systems. It's a relatively easy to use tool, but ultimately it can provide encoding performance at the cost of flexibility and quality.

There are other CUDA applications, of course. We'll look at SimHD DVD up-sampling in our battery life tests for example. Another potentially useful tool is called vReveal, which can remove noise, improve contrast/lighting, and reduce shakiness in recorded videos. If you make videos with your cell phone or inexpensive digicam and you want to upload them to YouTube, vReveal might be worth a purchase. A better video recorder would be an even better purchase, of course.

Most of the CUDA tools at present focus on video (and to a lesser extent image) related activities, since these tend to be highly parallelizable. Anyone doing video editing/encoding work will want something far more powerful than an Atom-based netbook, even if it does have ION to help out. Ultimately, while CUDA is certainly interesting, we prefer a vendor-agnostic approach, and DirectCompute will hopefully provide that in the coming months (or years). We also need to maintain a balanced system — certain tasks still rely heavily on the CPU and aren't candidates for GPU acceleration — and Atom is nowhere near fast enough for a modern computer.