Quick Sync: The Best Way to Transcode

Currently Intel’s Quick Sync transcode is only supported by two applications: Cyberlink’s Media Espresso 6 and Arcsoft’s Media Converter 7. Both of these applications are video to go applications targeted at users who want to take high resolution/high bitrate content and transcode it to more compact formats for use on smartphones, tablets, media streamers and gaming consoles. The intended market is not users who are attempting to make high quality archives of Blu-ray content. As a result, there’s no support for multi-channel audio; both applications are limited to 2-channel MP3 or AAC output. There’s also no support for transcoding to anything higher than the main profile of H.264.

Intel indicates that these are not hardware limitations of Quick Sync, but rather limitations of the transcoding software. To that extent, Intel is working with developers of video editing applications to bring Quick Sync support to applications that have a more quality-oriented usage model. These applications are using Intel’s Media SDK 2.0 which is publicly available. Intel says that any developer can get access to and use it.

For the purposes of this comparison I’ve used Media Converter 7, but that’s purely a personal preference thing. The performance and image quality should be roughly identical between the two applications as they both use the same APIs. Jarred's look at Mobile Sandy Bridge will focus on MediaEspresso.

Where image quality isn’t consistent however is between transcoding methods in either application. Both applications support four codepaths: ATI Stream, Intel Quick Sync, NVIDIA CUDA, and x86. While you can set any of these codepaths to the same transcoding settings, the method by which they arrive at the transcoded image will differ. This makes sense given how different all four target architectures are (e.g. a Radeon HD 6870 doesn’t look anything like a NVIDIA GeForce GTX 460). Each codepath makes a different set of performance vs. quality tradeoffs which we’ll explore in this section.

The first but not as obvious difference is if you use the Sandy Bridge CPU cores vs. Quick Sync to transcode you will actually get a different image. The image quality is slightly better on the x86 path, but the two are similar.

The reason for the image quality difference is easy to understand. CPUs are inherently not very parallel beasts. We get tremendous speedup on highly parallel tasks on multi-core CPUs, but compared to a GPU’s ability to juggle hundreds or thousands of threads, even a 6-core CPU doesn’t look too wide. As a result of this serial vs. parallel difference, transcoding algorithms optimized for CPUs are very computationally efficient. They have to be, because you can’t rely on hundreds of cores running in parallel when you’re running on a CPU.

Take the same code and run it on a GPU and you’ll find that the majority of your execution resources are wasted. A new codepath is needed that can take advantage of the greater amount of compute at your disposal. For example, a GPU can evaluate many different compression modes in parallel whereas on a CPU you generally have to pick a balance between performance and quality up front regardless of the content you’re dealing with.

There’s also one more basic difference between code running on the CPU vs. integrated GPU. At least in Intel’s case, certain math operations can be performed with higher precision on Sandy Bridge’s SSE units vs. the GPU’s EUs.

Intel tuned the PSNR of the Quick Sync codepath to be as similar to the x86 codepath as possible. The result is, as I mentioned above, quite similar:

Now let’s tackle the other GPUs. When I first started my Quick Sync investigations I did a little experiment. Without forming any judgments of my own, I quickly transcoded a ~15Mbps 1080p movie into a iPhone 4 compatible 720p H.264 at 4Mbps. I then trimmed it down to a single continuous 4 minute scene and passed the movie along to six AnandTech editors. I sent the editors three copies of the 4 minute scene. One transcoded on a GeForce GTX 460, one using Intel’s Quick Sync, and one using the standard x86 codepath. I named the three movies numerically and told no one which platform was responsible for each output. All I asked for was feedback on which ones they thought were best.

Here are some of the comments I received:

“Wow... there are some serious differences in quality. I'm concerned that the 1.mp4 is the accelerated transcode, in which case it looks like poop..”

“Video 1: Lots of distracting small compression blocks, as if the grid was determined pre-encoding (I know that generally there are blocks, but here the edges seem to persist constantly). Persistent artifacts after black. Quality not too amazing, I wouldn't be happy with this.”

Video one, which many assumed was Quick Sync, actually came from the GeForce GTX 460. The CUDA codepath, although extremely fast, actually produces a much worse image. Videos 2 and 3 were outputs from Sandy Bridge, and the editors generally didn’t agree on which one of those two looked better just that they were definitely better than the first video.

To confirm whether or not this was a fluke I set up three different transcodes. Lossy video compression is hard to get right when you’re transcoding scenes that are changing quickly, so I focused on scenes with significant movement.

The first transcode involves taking the original Casino Royale Blu-ray, stripping it of its DRM using AnyDVD HD, and feeding that into MC7 as a source. The output in this case was a custom profile: 15Mbps 1080p main profile H.264. This is an unrealistic usage model simply because the output file only had 2-channel audio, making it suitable only for PC use and likely a waste of bitrate. I simply wanted to see how the various codepaths looked and performed with an original BD source.

Let’s look at performance first. The entire movie has around 200,000 frames, the transcoding frame rate is below:

ArcSoft Media Converter 7—Casino Royale Transcode

As we’ve been noting in our GPU reviews for quite some time now, there’s no advantage to transcoding on a GPU faster than the $200 mainstream parts. Remember that the transcode process isn’t all infinitely parallel, we are ultimately bound by the performance of the sequential components of the algorithm. As a result, the Radeon HD 6970 offers no advantage over the 6870 here. Both of these AMD GPUs end up being just as fast as a Core i5-2500K.

NVIDIA’s GPUs offer a 15.7% performance advantage, but as I mentioned earlier, the advantage comes at the price of decreased quality (which we’ll get to in a moment).

Inte’s Quick Sync is untouchable though. It’s 48% faster than NVIDIA’s GeForce GTX 460 and 71% faster than the Radeon HD 6970. I don’t want to proclaim that discrete GPU based transcoding is dead, but based on these results it sure looks like it. What about image quality?

My image quality test scene isn’t anything absurd. Bond and Vespyr are about to meet Mathis for the first time. Mathis walks towards the two and the camera pans to follow him. With only one character and the camera both moving at a predictable rate, using some simple motion estimation most high quality transcoders should be able to handle this scene without getting tripped up too much.

Intel Core i5-2500K (x86) Intel Quick Sync NVIDIA GeForce GTX 460 AMD Radeon HD 6870
Download: PNG Download: PNG Download: PNG Download: PNG

Comparing the shots above the only real outlier is NVIDIA’s GeForce GTX 460. The CUDA path clearly errs on the side of performance vs. quality and produces a far noisier image. The ATI Stream codepath produces an image that’s very close to the standard x86 and Quick Sync output. In fact, everything but the GTX 460 does well here.

The next test uses an already transcoded 15Mbps 1080p x264 rip of Quantum of Solace Blu-ray disc. For many this is likely what you’ll have stored on your movie server rather than a full 50GB Blu-ray rip. Our destination this time is the iPhone 4. The settings are as follows: 4Mbps 720p H.264.

At only 4Mbps there’s a lot of compression going on, image quality isn’t going to be nearly as good as the previous test. Performance is considerably higher as the encoders are able to discard more data and optimize for performance over absolute quality. The entire movie has 152,000 frames that are transcoded in this test:

ArcSoft Media Converter 7—Quantum of Solace Transcode

The six-core Phenom II X6 1100T is faster than the Core i5-2500K thanks to the latter’s lack of Hyper Threading. Both are around the speed of the Radeon HD 6870.

The GeForce GTX 460 is faster than any standalone x86 CPU, regardless of core count. However once again, Quick Sync blows them all out of the water. At 200 frames per second Quick Sync is more than twice the speed of a standard Core i5-2500K or even the Phenom II X6 1100T. And it’s nearly twice as fast as the GTX 460.

The image quality comparison scene is also far more stressful on the transcoders. There’s a lot of unpredictable movement going on as Bond is in pursuit of a double agent at the beginning of the film.

Intel Core i5-2500K (x86) Intel Quick Sync NVIDIA GeForce GTX 460 AMD Radeon HD 6870
Download: PNG Download: PNG Download: PNG Download: PNG

The image quality story is about the same for AMD’s GPUs and the x86 path, however Quick Sync delivers a noticeably worse quality image. It’s no where near as bad as the GTX 460, but it’s just not as sharp as what you get from the software or ATI Stream codepaths.

The problem here seems to be that when transcoding from a lower quality source, the tradeoffs NVIDIA makes are amplified. Even Quick Sync isn’t perfect here. I’d say Quick Sync is closer to the pure x86 path than CUDA. Given the tremendous performance advantage I’d say the tradeoff is probably worth it in this case.

For our final test we’ve got a 12Mbps 1080p x264 rip of The Dark Knight. Our target this time is a 640x480, 1.5Mbps iPod Touch compatible format.

ArcSoft Media Converter 7—Dark Knight Transcode

Surprisingly enough the 6970 shows a slight performance advantage compared to the 6870 in this test, but still not enough to approach the speed of the x86 CPUs in this test. Quick Sync is almost 4x faster than the Radeon HD 6970 and twice as fast as everything else.

Our Dark Knight image quality test is also the most strenuous of the review. We’re looking at a very dark, high motion scene with a sudden explosion. The frame we’re looking at is right after the Joker fires a rocket at the rear of a police car. The sudden explosion casts light everywhere which can’t be predicted based on the previous frame.

Intel Core i5-2500K (x86) Intel Quick Sync NVIDIA GeForce GTX 460 AMD Radeon HD 6870
Download: PNG Download: PNG Download: PNG Download: PNG

The GeForce GTX 460 looks horrible here. The output looks like an old film, it’s simply inexcusable.

The Radeon HD 6870 produces a frame that has similar sharpness to the x86 codepath, but with muted colors. Quick Sync maintains color fidelity but loses the sharpness of the x86 path, similar to what we saw in the previous test. In this case the loss of sharpness does help smooth out some aliasing in the paint on the police car but otherwise is undesirable.

Overall, based on what I’ve seen in my testing of Quick Sync, it isn’t perfect but it does deliver a good balance of image quality and performance. With Quick Sync enabled you can transcode a ~2.5 hour Blu-ray disc in around 35 minutes. If you’ve got a lower quality source (e.g. a 15GB Blu-ray re-encode), you can plan on doing a full movie in around 13 minutes. Quick Sync will chew through TV shows in a couple of minutes, without a tremendous loss in quality.

With CUDA on NVIDIA GPUs we had to choose between high quality or high performance. (Perhaps other applications will do the transcode better as well, but at least Arcsoft's Media Converter 7 has serious image quality problems with CUDA.) With Quick Sync you can have both, and better performance than we’ve ever seen from any transcoding solution in desktops or notebooks.

Quick Sync with a Discrete GPU

There’s just one hangup to all of this Quick Sync greatness: it only works if the processor’s GPU is enabled. In other words, on a desktop with a single monitor connected to a discrete GPU, you can’t use Quick Sync.

This isn’t a problem for mobile since Sandy Bridge notebooks should support switchable graphics, meaning you can use Quick Sync without waking up the discrete GPU. However there’s no standardized switchable graphics for desktops yet. Intel indicated that we may see some switchable solutions in the coming months on the desktop, but until then you either have to use the integrated GPU alone or run a multimonitor setup with one monitor connected to Intel’s GPU in order to use Quick Sync.

Intel’s Quick Sync Technology Intel’s Gen 6 Graphics


View All Comments

  • ReaM - Tuesday, January 04, 2011 - link

    the six core 980x still owns them in all tests where all cores are used.

    I dont know 22k in cinebench is really not a reason to buy the new i7, I reach 24k on air with i7 860 and my i5 runs on 20k on air.

    Short term performance is real good, but I dont care if I wait for a package to unpack for 7 seconds or 8, for long term like rendering, neither there is a reason to upgrade.

    I recommend you get the older 1156 off ebay and save a ton of money.

    I have the i5 on hackintosh, I am wondering if 1155 will be hackintoshable
  • Spivonious - Tuesday, January 04, 2011 - link

    I have to disagree with Anand; I feel the QuickSync image is the best of the four in all cases. Yes, there is some edge-softening going on, so you lose some of the finer detail that ATi and SNB gives you, but when viewing on a small screen such as one on an iPhone/iPod, I'd rather have the smoothed-out shapes than pixel-perfect detail. Reply
  • wutsurstyle - Tuesday, January 04, 2011 - link

    I started my computing days with Intel but I'm so put off by the way Intel is marketing their new toys. Get this but you can't have that...buy that, but your purchase must include other things. And even after I throw my wallet to Intel, I still would not have a OC'd Sandy Bridge with useful IGP and Quicksync. But wait, throw more money on a Z68 a little later. Oh...and there's a shiny new LGA2011 in the works. Anyone worried that they started naming sockets after the year it comes out? Yay for spending!

    AMD..please save us!
  • MrCrispy - Tuesday, January 04, 2011 - link

    Why the bloody hell don't the K parts support VT-d ?! I can only imagine it will be introduced at a price premium in a later part. Reply
  • slick121 - Tuesday, January 04, 2011 - link

    Wow I just realized this. I really hate this type of market segmentation. Reply
  • Navier - Tuesday, January 04, 2011 - link

    I'm a little confused why Quick Sync needs to have a monitor connected to the MB to work. I'm trying to understand why having a monitor connected is so important for video transcoding, vs. playback etc.

    Is this a software limitation? Either in the UEFI (BIOS) or drivers? Or something more systemic in the hardware.

    What happens on a P67 motherboard? Does the P67 board disable the on die GPU? Effectively disabling Quick Sync support? This seems a very unfortunate over-site for such a promising feature. Will a future driver/firmware update resolve this limitation?

  • NUSNA_moebius - Tuesday, January 04, 2011 - link

    Intel HD 3000 - ~115 Million transistors
    AMD Radeon HD 3450 - 181 Million transistors - 8 SIMDs
    AMD Radeon HD 4550 - 242 Million transistors - 16 SIMDs
    AMD Radeon HD 5450 - 292 Million transistors - 16 SIMDs
    AMD Xenos (Xbox 360 GPU) - 232 Million transistors + 105 Million (eDRAM daughter die) = 337 Million transistors - 48 SIMDs

    Xenos I think in the end is still a good two, two and a half times more powerful than the Radeon 5450. Xenos does not have to be OpenCL, Direct Compute, DX11 nor fully DX10 compliant (a 50 million jump from the 4550 going from DX10.1 to 11), nor contains hardware video decode, integrated HDMI output with 5.1 audio controller (even the old Radeon 3200 clocks in at 150 million + transistors). What I would like some clarification on is if the transistor count for the Xenos includes Northbridge functions..............

    Clearly PC GPUs have insane transistor counts in order to be highly compatible. It is commendable how well the Intel HD 3000 does with only 115 Million, but it's important to note that older products like the X1900 had 384 Million transistors, back when DX9.0c was the aim and in pure throughput, it should match or closely trail Xenos at 500 MHz. Going from the 3450 to 4550 GPUs, we go up another 60 million for 8 more SIMDs of a similar DX10.1 compatible nature, as well as the probable increases for hardware video decode, etc. So basically, to come into similar order as the Xenos in terms of SIMD counts (of which Xenos is 48 of it's own type I must emphasize), we would need 60 million transistors per 8 SIMDs, which would put us at about 360 million transistors for a 48 SIMD (240 SP) AMD part that is DX 10.1 compatible and not equipped with anything unrelated to graphics processing.

    Yes, it's a most basic comparison (and probably fundamentally wrong in some regards), but I think it sheds some light on the idea that the Radeon HD 5450 really still pales in comparison to the Xenos. We have much better GPUs like Redwood that are twice as powerful with their higher clock speeds + 400 SPs (627 Million transistors total) and consume less energy than Xenos ever did. Of course, this isn't taking memory bandwidth or framebuffer size into account, nor the added benefits of console optimization.
  • frankanderson - Tuesday, January 04, 2011 - link

    I'm still rocking my Q6600 + Gigabyte X38 DS5 board, upgraded to a GTX580 and been waiting for Sandy, definitely looking forward to this once the dust settles..

    Thanks Anand...
  • Spivonious - Wednesday, January 05, 2011 - link

    I'm still on E6600 + P965 board. Honestly, I would upgrade my video card (HD3850) before doing a complete system upgrade, even with Sandy Bridge being so much faster than my old Conroe. I have yet to run a game that wasn't playable at full detail. Maybe my standards are just lower than others. Reply
  • aviat72 - Tuesday, January 04, 2011 - link

    Though SB will be great for some applications, there are still rough edges in terms of the overall platform. I think it will be best to wait for SNB-E or at least the Z68. SNB-E seems to be the best future-proofing bet.

    I also wonder how a part rated for 95W TDP was drawing 111W in the 4.4GHz OC (the Power Consumption Page). SB's power budget controller must be really smart to allow the higher performance without throttling down, assuming your cooling system can manage the thermals.

Log in

Don't have an account? Sign up now