GPU acceleration comes to Flash video

The introduction of Intel’s Atom and the rise of the netbook have unquestionably validated the market for low-cost, low-power computing. Compact little systems that are “just fast enough” for the general set of uses people have for computers—surfing the web, chatting via Skype, sending emails, balancing the checkbook, taking notes, conducting Ponzi schemes via Craigslist—have been selling like mad for the past couple of years, and they show no signs on slowing down in the future. Once you exclude a few tough cases like video editing and hard-core gaming, low-cost PCs tend to do very well on the most common sorts of PC applications, which is why consumers have been buying them in droves.

But there’s long been one area of common PC usage where low-cost systems tend to choke: dealing with Adobe’s Flash platform, especially the near-ubiquitous Flash-based videos plastered all over the web. Pull up that latest hilarious YouTube clip of babies and/or kittens on your hot pink MSI Wind, and you’re in for deep disappointment. Those lovable little balls of flubber/fluff slow to a near-standstill. Meanwhile, the Wind works to earn it name; its fan kicks up, shooting a stream of hot air out of its side, to no great effect. You’re stuck in slideshow land, from which the best escape is closing the browser tab—if only the system would respond to your clicks.

This is a tragic outcome for what is otherwise a dynamite little computer, made more infuriating by the fact that such systems will play a range of video formats just fine via Windows Media Player or the like. The difference is that traditional media players often employ some form of GPU acceleration, at least for scaling up the video to the size of the display or window in which it’s being played—and perhaps even for decoding a highly compressed format like H.264. Trouble is, Flash doesn’t make use of the video acceleration logic built into most graphics chips, so the poor low-power CPU is left to fend entirely for itself. A tragic kitten slideshow is the only possible outcome.

Until now, that is. Adobe has recently released a beta version of Flash 10.1 with built-in GPU acceleration. The support for GPUs is preliminary and is largely focused on video decoding alone, but it’s already being facilitated by new video drivers from AMD, Nvidia, and Intel. The promise? Low-power computing just might be able to conquer its one long-standing bugaboo. Amazingly enough, given Flash’s broad popularity, this is a development of pretty serious import in certain quarters. To see how much of a difference GPU acceleration can make, we rounded up a trio of low-power systems with integrated graphics and took the Flash 10.1 beta for a spin.

Intel Atom 330 and Nvidia Ion performance
Nvidia has generated a tremendous amount of press coverage and general goodwill for its Ion platform, which simply pairs up the GeForce 9400 chipset and IGP with an Atom processor. The not-so-subtle message of the Ion platform pitch is simple: that a small, commodity-class CPU will suffice for most users, so long as it’s paired with a competent chipset and graphics solution. The fact that Intel tends to restrict the Atom to the time-worn 945G chipset and its moldy old IGP, with no H.264 decode acceleration and all the graphics power of a Lite Brite, only helps underscore Nvidia’s point. A system based on the Ion platform has substantially more potential than your usual netbook or nettop, and for some practical purposes, it may rival PCs based on much more powerful CPUs.

Flash video acceleration, then, is a great test case for the Ion’s mojo. We checked it out using an ASRock ION 330 nettop. Even its dual-core Atom 330 struggles with Flash video, though not quite as much as your average single-core Atom netbook.

For this test, we installed Windows 7 x64 on the ION 330, along with Firefox 3.5.5, the Flash 10.1 beta, and Nvidia’s ForceWare 195.5 beta drivers with Flash video support. The ION 330 was driving a VGA monitor at 1600×1200 resolution and an 85Hz refresh rate. We used the Flash 10.1 beta for both sets of tests, simply disabling GPU acceleration for the CPU-only portion of our testing.

Those of you who have read one of our netbook or laptop reviews will likely be familiar with our simple red-yellow-green system of rating video playback performance. The rest of you had better be able to figure it out on your own. Heh. We’re simply playing back a video and recording our impressions of the playback’s fluidity, while taking notes about any quirks or problems.

We do have some approximate CPU utilization numbers taken from the Windows Task Manager for the windowed playback tests, but I wouldn’t put too much stock into them, frankly. Many factors can throw off CPU utilization numbers, including the Atom’s dual front ends used for Hyper-Threading. Anyone who has seen an Atom struggle to play back a video, dropping frames left and right while posting only 80% CPU utilization, will understand why utilization alone isn’t a great indicator of performance. These numbers can help us understand the before-and-after comparison between GPU video acceleration and CPU-only playback on a given platform, though, so I’ve included them for reference.

We tested with a range of video clips on both Hulu and YouTube, ranging from low-res 360p to full-size 1080p HD. 720p videos are still rare on Hulu, but we tried one of the few 720p videos in Hulu’s HD Showcase, the premiere episode of Legend of the Seeker, being sure to use the pop-out feature to display the full 720p image. We used a pair of movie trailers on YouTube, including one for Moon in HQ format and another for Star Trek in 720p. Our 1080p clip was a PureVideo demo posted by Nvidia. At Nvidia’s suggestion, you may notice, I added some format strings to the ends of those YouTube URLs to work around some problems with GPU acceleration when the video is first opened in a lower-resolution format. In all cases, we tested at the native windowed resolution and then again when scaled to full-screen mode.

As you can tell, the ION 330 struggles quite a bit when the dual-core Atom alone is asked to decode and play back our test videos. Even our episode of Fringe at a low 360p resolution dips into iffy yellow territory with dropped frames in full-screen mode, and anything above that on Hulu is hopeless. On YouTube, only the Moon trailer proves workable.

With some help from the Ion’s PureVideo acceleration logic, though, the ION 330 is transformed into a much more competent media playback system. Suddenly, all but the most difficult videos play back smoothly, some of them even ranging into “Perfect” territory, where I consider playback entirely fluid.

There is one disappointing exception in the form of the 720p LoTS episode. Although playback is typically fluid and CPU utilization relatively low, the video pauses briefly every few seconds, though audio is uninterrupted.

I tried a number of things to resolve this issue. I started with Win7 x86 and installed a fresh copy of Win7 x64 hoping to sidestep the issue, but it persisted. I let the Hulu player queue up a 100% full buffer, to see if that would help, to no avail. I compared to playback on my own PC, using the same network and Internet connection; my Core 2 Quad-based system had no trouble playing the video flawlessly. I even tried dropping the display resolution and refresh rate on the monitor, to no apparent effect. In the end, I had to settle for vastly improved Hulu video playback that fell somewhat short of perfect with this video, though not quite into the red.

Still, the overall improvement is startling. The drop in CPU utilization with the 720p and 1080p YouTube clips is especially striking. The Star Trek trailer runs fluidly with Task Manager dipping into the single digits. Interestingly, the Ion is unique on the PC, as far as I know, in that it assists with scaling the video’s size as well as decoding it. No doubt Nvidia made this extra effort due to the fact that the Atom could definitely use the help. That work seems to have paid off nicely.

This first beta of Flash GPU video acceleration remains fragile and incomplete, of course. I doubt most folks will bother with appending format strings onto YouTube URLs manually, for instance. Also, I ran into a problem with 480p Hulu video, unscaled in a window, looking as if it had been poorly scaled to a non-native resolution. Nvidia tells us this problem will be resolved when Adobe releases its next Flash 10.1 beta version.

Here’s hoping subsequent improvements to the software and driver stacks involved here can also overcome the intermittent video pauses we saw in the 720p Hulu video format. If they can, the Ion will have nearly aced this test, which would be quite the revelation.

Intel’s GS45M chipset and CULV processor
My next victim is my new favorite laptop, the Acer Aspire 1810TZ. This little 11.6″ ultraportable is technically not a netbook; it’s a step above, based as it is on one of Intel’s consumer ultra-low-voltage (CULV) processors, the Pentium SU4100. This is a Penryn-based dual-core CPU with a 1.3GHz core clock, 800MHz bus, and 2MB of L2 cache. Like most CULV systems, the 1810TZ has an Intel GS45 chipset with GMA 4500MHD integrated graphics. This is a newer chipset than the 945G found in most netbooks, and it’s capable of accelerating HD video decoding and playback.

We tested the Aspire 1810TZ with Intel’s 15.16.2.1986 video drivers. This laptop has a native screen resolution of 1366×768 and comes with Windows 7 Home Premium x64 installed. For this system, our CPU-only numbers come from Flash 10.0.32, while the GPU-accelerated ones come from the 10.1 beta. Since the dual-core CULV processor is really quite competent, as you’ll see, we’ve omitted the Hulu 360 tests from this round.

All by its lonesome, the Aspire’s dual-core Pentium handles the 480p Hulu clip and the Moon trailer perfectly. The 720p Hulu and YouTube videos are also quite good in windowed mode. Only when scaling the 720p clips to full screen do some small cracks begin to appear in its armor. This is way better than a netbook, but even so, decoding the 1080p clip is clearly beyond the Pentium SU4100’s capabilities.

The results with GPU acceleration enabled are intriguing. The problems are evident immediately with the 480p Hulu video. CPU utilization is actually higher with GPU acceleration enabled, and the frame rate drops to maybe half of what it should be—perhaps less. Clearly, Intel and Adobe haven’t got this one sorted out yet. The issues continue with the 720p Hulu video, where CPU utilization is lower, but so are frame rates, dropping performance into the iffy yellow territory. Once again, playback is smoother without GPU acceleration.

The bad news in Hulu is offset by unmitigated success with the first two YouTube clips, where CPU utilization is lower and playback is pristine. Thanks to GPU acceleration, the 1810TZ can play back the 720p Star Trek trailer perfectly at its full screen resolution (which, incidentally, is just a tad larger than the 720p format and the exact same aspect ratio). The CPU alone didn’t quite pull off that feat.

Then again, the GMA 4500MHD isn’t any practical help with the 1080p clip, which has the 1810TZ overwhelmed yet again.

Unfortunately, GPU acceleration on the GS45 chipset doesn’t appear to be a clear win. You’re better off overall sticking with the CPU-only release versions of Flash at this point in time. Here’s hoping Intel can manage to get video decode acceleration working properly. If they do not, some forms of Flash HD video will play back smoother on an Atom + Ion platform combo than on Intel’s 4-series chipsets and CULV Penryn dual-cores, amazingly enough.

AMD Phenom II and 785G performance
Next up is my home theater PC, based on AMD’s “Maui” platform, which has been serving faithfully in my living room while awaiting a proper write-up for a shamefully long time. (It’s half done, I swear!) This box is by no means a slow computer, based on a 785G chipset and a low-power Phenom II X4 905e processor, a 2.5GHz quad-core. Yet it still has a few problems with Flash video from time to time, which matters a lot to me since I watch shows on it via Hulu Desktop with some frequency. Accordingly, I’ve added another test for this system: full-screen playback of 720p video view the Hulu Desktop application.

This system runs Windows 7 x86 and is hooked up to a 1080p television via HDMI. We tested with the Catalyst 9.11 drivers and Flash version 10.0.32 (CPU only), along with the Flash 10.1 beta for GPU acceleration. Here are the results.

As I said, even this system with a relatively fast quad-core processor stumbles slightly over Flash video at times. Fortunately, playback of video with the Flash 10.1 beta and integrated Radeon graphics smoothes over those few bumps in the road, turning in a near-perfect performance. CPU utilization drops quite a bit with GPU acceleration, too.

The 785G’s Flash acceleration is so good that I’ve decided to leave it enabled for everyday use. The primary impact of doing so is that run-of-the-mill 480p Hulu videos played via Hulu Desktop are now more fluid. Previously, playback was acceptable but not stellar, with occasional obvious dropped frames. Now, Hulu Desktop looks as good as one could want—save for the fact that no Flash-based solution offers the kind of high-quality scaling routine one would expect from Windows Media Center or other fully-DXVA-accelerated media players.

AMD has a few extra tricks up its sleeve, too. In the Catalyst Control Center, one can enable and modify the values for a host of color enhancements and post-processing effects, and GPU-accelerated Flash videos will be modified in real time. In fact, AMD has a split-screen preview mode to show how the video output is being modified.

A few of the controls available in the CCC

The Star Trek trailer in split screen with edge enhancement enabled

This is beyond nifty. I’ll admit I’ve only used such controls with other video players occasionally, but from time to time, an interesting video from a poor source can be redeemed through judicious use of the right filters. Having such capabilities available once Flash video becomes GPU-accelerated only makes sense, and it’s good to see AMD making it happen.