Mobile Sandy Bridge QuickSync and 3DMarks

Anand has provided plenty of coverage of transcoding quality in the desktop SNB review, using Arcsoft’s Media Encoder 7. For the mobile side of things, we’ll turn to CyberLink’s MediaEspresso 6—a similar package that’s useful for quick encodes of movies for YouTube or mobile device consumption. NVIDIA has been touting the benefits of GPU acceleration for such tasks for over a year now, with CUDA making a fairly decent showing. MediaEspresso also supports CUDA acceleration, making for a nice head-to-head, though I’m limited to hardware that I still have on hand.

For the encoding test, I’ve grabbed two other recently reviewed notebooks to show how they compare to Sandy Bridge. The first is ASUS’ mainstream N53JF notebook, sporting an i5-460M and GT 425M GPU. For the higher performance notebook offering, we’ve got ASUS’ G73Jw with i7-740QM and GTX 460M. [Ed: Sorry for the delay in shipping it back, ASUS—it will go out this week now that we’re done with Sandy Bridge testing!] I used a 720p shot with an iPod Touch and transcoded it to a 2Mb 720p YouTube compatible stream. MediaEspresso also has some video quality enhancement features available, dubbed TrueTheater AutoLight, Denoise, and HD. I ran the transcode tests with and without the enhancements enabled, with and without QuickSync/GPU acceleration. Since MediaEspresso also supports ATI GPUs, I tossed in results from my i7-920 with CrossFire HD 5850 as well.

Accelerated MediaEspresso Encoding

CPU-Based MediaEspresso Encoding

Accelerated MediaEspresso Enhanced Encoding

CPU-Based MediaEspresso Enhanced Encoding

First things first, I’d say it’s fair to state that the GPU acceleration for AMD GPUs (at least in this particular instance) isn’t as good as NVIDIA’s CUDA or Intel’s QuickSync. Perhaps future driver, hardware, and/or software updates will change the picture, but the HD 5850 cards in my desktop fail to impress. The CUDA results for GTX 460M are quite good, while the GT 425M was roughly on par with CPU encoding on a quad-core (plus Hyper-Threading) processor. Finally, Intel’s Sandy Bridge manages to easily eclipse any of the other systems—with or without QuickSync.

Using pure CPU encoding, the 2820QM finishes the transcode in 15% less time than a desktop i7-920, and 44% less time than the i7-740QM. Enabling all of the extra TrueTheater enhancements definitely has an impact on performance (and depending on the video source may or may not be worthwhile). Sandy Bridge still required 8% less time than i7-920, and 36% less time than i7-740QM, never mind the i5-460M that requires 134% longer to accomplish the same task.

Switch on all of the GPU acceleration support (including QuickSync, which isn’t technically a GPU feature) and all of the times drop, some substantially. The basic transcode on SNB finishes in a blisteringly fast 10 seconds—this is a 1:33 minute clip with 30FPS content, so the transcode happens at roughly 280FPS (wow!). GTX 260M comes in next at 17 seconds (174FPS), then CrossFire 5850 ends up needing three times longer than SNB and almost twice as long as the mobile GTX 460M, and GT 425M brings up the rear at twice the time of the HD 5850. With the TrueTheater features enabled, the CPU appears to do a lot more work and the GTX 460M and Sandy Bridge are both over an order of magnitude slower.

This is obviously a huge in for Intel, but of course it all depends on how often you happen to transcode videos—and how patient you happen to be. I do it seldom enough that even running encodes on my old quad-core Kentsfield CPU doesn’t particularly bother me; I just set up the transcodes in TMPGEnc Express and walk away, and they’re usually done when I return. If on the other hand you’re the type that lives in the social networks and Twitter feeds, being able to get your video up on YouTube five to ten times faster (without a significant loss in quality, at least based on my iPod Touch experience) is definitely useful.

Futuremark 3DMark Vantage

Futuremark 3DMark06

Futuremark 3DMark05

One final item to quickly cover is synthetic graphics performance, courtesy of 3DMark. Sandy Bridge places in the middle of the pack, and obviously desktop solutions are far out of reach for the time being, but according to 3DMark we could see performance actually surpass some of the entry-level GPUs. Maybe 3DMark just has heavy optimizations from Intel…then again, maybe they actually do have a GPU that can compete.

Mobile Sandy Bridge Application Performance Mobile Sandy Bridge Gaming Performance
POST A COMMENT

66 Comments

View All Comments

  • JarredWalton - Tuesday, January 04, 2011 - link

    Definitely a driver bug, and I've passed it along to Intel. The HD 4250 manages 7.7FPS, so SNB ought to be able to get at least 15FPS or so. The game is still a beast, though... some would say poorly written, probably, but I just call it "demanding". LOL Reply
  • semo - Monday, January 03, 2011 - link

    Thanks for mentioning USB 3.0 Jarred. It is a much too overlooked essential feature these days. I simply will not pay money for a new laptop in 2011 without a single USB 3.0 port. Reply
  • dmbfeg2 - Monday, January 03, 2011 - link

    Which tool do you use to check the turbo frequencies under load? Reply
  • JarredWalton - Monday, January 03, 2011 - link

    I had both CPU-Z and the Intel Turbo Monitoring tool up, but neither one supports logging so I have to just eyeball it. The clocks in CPU-Z were generally steady, though it's possible that they would bump up for a few milliseconds and then back down and it simply didn't show up. Reply
  • Shadowmaster625 - Monday, January 03, 2011 - link

    On the other Sandy Bridge article by Anand, right on the front page, it is mentioned that the 6EU GT1 (HD2000) die has 504M transistors, while the 12EU GT2 (HD 3000) die has 624M transistors. Yet here you are saying HD Graphics 3000 has 114M. If the 12EU version has 120M more transistors than the 6EU version, then does that not imply a total gpu transistor count well north of 200M? Reply
  • JarredWalton - Monday, January 03, 2011 - link

    AFAIK, the 114M figure is for the 12EU core. All of the currently shipping SNB chips are quad-core with the full 12EU on the die, but on certain desktop models Intel disables half the EUs. However, if memory serves there are actually three SNB die coming out. At the top is the full quad-core chip. Whether you have 6EU or 12EU, the die is the same. For the dual-core parts, however, there are two chips. One is a dual-core with 4MB L3 cache and 12EUs, which will also ship in chips where the L3 only shows 3MB. This is the GT1 variant. The other dual-core version is for the ultra-low-cost Pentium brand, which will ship with 6EUs (there will only be 6EU on the die) and no L3 cache, as well as some other missing features (Quick Sync for sure). That's the GT2, and so the missing 120M includes a lot of items.

    Note: I might not be 100% correct on this, so I'm going to email Anand and our Intel contact for verification.
    Reply
  • mino - Monday, January 03, 2011 - link

    Nice summary (why was this not in the article ?).

    Anyway those 114M do not include memory controller, encoding, display output etc. so the comparison with Redwood/Cedar is not really meaningful.

    If you actually insist on comparing transistor counts, semething like (Cedar-Redwood)/3 shall give you a reasonable value of AMD's SPU efficiency from transistors/performance POW.
    Reply
  • mino - Monday, January 03, 2011 - link

    "After all, being able to run a game at all is the first consideration; making it look good is merely the icing on the cake."

    If making it look good is merely icing on the cake, why bother with GPUs ? Lets just play 2D Mines!
    (While for the poor souls stuck with Intel IGPs it certainly is just the icing, for Christ's sake, that is a major _problem_, not a feature !!!)

    After a few pages I have decided to forgo the "best-thing-since-sliced-bread" attitude, but, what is too much is too much...
    Reply
  • mino - Monday, January 03, 2011 - link

    Regardless the attitude, HUGE thanks for listening to comments and including the older games roundup.

    While I'd love to see more games that actually provide playable frame-rates (read: even older ones) on SNB-class IGPs like Far Cry or HL2, even this mini-roundup is a really big plus.

    As for a suggestion on future game-playability roundup on IGP's, it is really simple:
    1) Take a look at your 2006-2007 GPU benchmarking suites
    2) Add in a few current MMORPGs
    Reply
  • JarredWalton - Monday, January 03, 2011 - link

    Anand covered several other titles, and most of the pre-2007 stuff should run fine (outside of blacklisting problems or bugs). Time constraints limit how much we can test, obviously, but your "reviewer on crack" comment is appreciated. 2D and 3D are completely different, and while you might feel graphical quality is of paramount importance, the fact of the matter is that SNB graphics are basically at the same level as PS3/Xbox 360 -- something millions of users are "okay" with.

    NVIDIA and AMD like to show performance at settings where they're barely playable and SNB fails, but that's no better. If "High + 1680x1050" runs at 20FPS with Sandy Bridge vs. 40FPS on discrete mobile GPUs, wouldn't you consider turning down the detail to get performance up? I know I would, and it's the same reason I almost never enable anti-aliasing on laptops: they can't handle it. But if that's what you require, by all means go out and buy more expensive laptops; we certainly don't recommend SNB graphics as the solution for everyone.

    Honestly, until AMD gets the Radeon equivalent of Optimus for their GPUs (meaning, AMD GPU + Intel CPU with IGP and automatic switching, plus the ability to update your Radeon and Intel drivers independently), Sandy Bridge + GeForce 400M/500M Optimus is going to be the way to go.
    Reply

Log in

Don't have an account? Sign up now