|Introduction and Positioning|
|Today, for the first time since Atom first debuted, Intel is launching an updated version of the core that does more than integrate additional function blocks and lower power consumption. The tablet iteration of the new SoC is called Bay Trail, and it's aimed at the red-hot tablet and handheld market currently occupied by various chips from Apple, Qualcomm, Samsung, and NVIDIA.
Our first thought? Finally. This is what Intel's Atom architecture always aspired to.
Intel has done some impressive things with Atom, including its first serious forays into mobile phones and tablets, but the Bonnell microarchitecture that debuted in 2008 was explicitly designed to emphasize low power consumption, not performance. ARM has released multiple processor updates since 2008 -- the Cortex-A8 debuted that year, followed by the Cortex-A9 and the Cortex-A15, while Intel's in-order Atom contented itself with die shrinks and feature integration. Such efforts gave us Medfield and Clover Trail devices, but they also risked Intel being permanently painted as an also-ran.
After today, that shouldn't be a problem.
We've covered the Silvermont architecture before; here we'll focus on the entire tablet, including performance tests and comparisons with other top-flight solutions from major players in the space.
The Bay Trail launch puts Intel in a position to hit both Android and Windows targets. That's in contrast to last year, when Clover Trail launched as a Windows part first, with Android options appearing later in the development cycle. It's also a testament to Windows 8's general weakness in the tablet market -- Android's share of the 7-10" space has grown significantly, while Windows remains mostly flat. Intel hasn't backed off supporting Microsoft, but it's clearly not putting all its eggs in one basket, either.
The various Bay Trail SKUs give some idea of the spaces Intel wants these parts to hit. At the top of the stack, there's the Z3770, a quad-core part with a Turbo Mode up to 2.4GHz, 2MB of L2 cache, dual-channel memory, up to 4GB of RAM (Clover Trail topped out at 2GB) and a maximum display resolution of 2560x1600. There are multiple alternates stepping downwards -- the single-channel parts use faster DDR3-1333 to compensate -- before we hit the dual-core chips. The Z3680 and Z3680D are dual-core chips with burst modes up to 2GHz, and lower memory bandwidth. Memory complements are fixed at 1GB for the Z3680, 2GB for the Z3680D. The Z3680, with its limited 1280x800 resolution, is obviously aimed at low-end devices, with the rest of the hardware aimed at a wide band of potential products and use cases.
*Note: Portions of this article were contributed by Marco Chiappetta and Dave Altavilla
|Design and Power Gating|
|Like Clover Trail, Bay Trail is a full system on a chip, or SoC. This time, however, Intel hasn't just integrated an advanced CPU core -- it's using its own GPU cores as well. Bay Trail includes four of the same Execution Units (EUs) that were used in Ivy Bridge. These chips can clock up to 667MHz. As far as we know, these are direct ports from Ivy Bridge, which means each EU can handle up to eight threads (total of 32) and has two texture samplers for a total of eight.
A great deal of work has gone into re-architecting the various components of the SoC, including a central storage block that ties the SDIO, SDcard, and eMMC storage controllers together, support for USB 3.0 baked into the chipset, added support for DisplayPort 1.2, and the ability to power gate the rest of the graphics chip while decoding video.
Four EUs isn't a great deal of graphics horsepower, compared to other current gen tablet SoCs, but it's a significant leap forward for Bay Trail. Previous iterations of Clover Trail used either a single SGX545 or a pair of SGX544 cores. The EUs inside SoC support DX11, OpenGL ES 3.0, and, while Intel doesn't specifically call it out, should also support Shader Model 5.0 (based on Ivy Bridge's GPU). It's not clear if this iteration of the core includes a separate graphics-specific L3 cache or not. Ivy Bridge chips with HD4000 cores had a separate cache hooked directly to the GPU, but Intel couldn't tell us if Bay Trail retained this feature.
Bay Trail can output audio and video simultaneously over either HDMI 1.4 or DisplayPort 1.2. HDMI 1.4 output is limited to 1920x1080, but DisplayPort can output at up to 2560x1440. 1920x1080 should be more than sufficient for the vast majority of tablet owners, but it'll be up to OEMs which standard they integrate for video out.
Like Clover Trail and Intel's high-end products, Bay Trail is also capable of monitoring multiple clock domains simultaneously and adjusting the SoC's clock speed to put MHz where they're needed most. That's fairly standard these days, but Intel's graphs suggest that they've further enhanced the flexibility of the underlying system for finer-grained control.
In this slide, the first CPU core is clocked three steps above standard while the other three CPU cores are quiescent. The GPU and Display power domains are both running faster than standard, while the camera is deactivated. This suggests that Bay Trail has added additional flexibility by allowing for variable clock speeds in areas beyond the CPU and GPU. How much of an impact this has on total power consumption remains to be seen -- while it's good to have the option, idle onboard cameras don't typically draw much power.
These changes, however, are part of a comprehensive Intel plan to introduce more efficient clocking and power management at every stage of the device. We've talked before about how the company is pushing plans for DevSleep on SSDs, SATA link power management, deeper CPU sleep states, and optimized PCI-Express power consumption alongside display panels with self-refresh. Many of the benefits introduced here, with Bay Trail, are further steps along the path towards power-optimized platforms at every level, rather than solutions that rely solely upon intelligent CPU/GPU power gating to hit targets.
That's the high level overview of Intel's Bay Trail architecture and its new capabilities. Let's look at performance.
|Bay Trail Performance on Android|
|At Intel Developer's Conference (IDF) this week in San Franciso, Intel gave us direct, hands-on benchmarking and testing time with new Bay Trail prototype vehicles. HotHardware's Marco Chiappetta slaved away in a conference room for hours to bring you the following actual, hard Bay Trail test results. Enjoy!
Bay Trail's performance in Android was rougher in spots than we saw with Windows, but there are more devices to test against here and a tougher operating environment overall. After W8's weak launch, all the momentum in the tablet industry is currently focused on Android-based designs and this is where Bay Trail will have to cut its teeth as well. However, you can of course also expect the WIntel alliance to be strong with next generation Windows 8.1 tablets flowing freely from various OEMs.
We'll start with AnTuTu 4.0, which we recommend taking with a grain of salt. While this isn't the version of the program that caused a furor earlier this summer when it was found to be erroneously handing Clover Trail an easy victory, this version of the benchmark doesn't appear to have been validated independently.
AnTuTu's performance puts Bay Trail ahead of everything but NVIDAI's Shield. Shield, it should be noted, is something of an unusual device. While it runs Android and is designed as a tablet alternative, it has a fan+heatsink. It's got the most aggressive cooling of any device on this list, which means it can clock substantially higher than many of its counterparts. Here, we see the Shield's GPU outstrip all others, with the Z3770 and the Galaxy S4 - itself no slouch - tying for second place. The Z3770 leads the S4 in integer and floating point performance and its memory is significantly faster.
Browsermark is another strong show for the Z3770. Chrome is somewhat faster than the default Android browser, but the 3770 is faster than Shield, which beats out the other hardware on the list. The gap is significant given that Shield has active cooling, and every other device (including Bay Trail), doesn't.
MobileXPRT is an Android benchmark that measures performance in five performance scenarios -- applying photo effects, creating photo collages, slideshows, encrypting data, and facial detection. In the chart below, the top line is the total score, while the individual tests are listed in time-to-complete. Higher scores are better in the Overall Score, while lower scores represent faster times to complete the tests in the detail below it.
What we see here, again, is that SHIELD (NVIDIA's Tegra 4 SoC) and the Intel Atom Z3770 are very closely tied, with other Android devices lagging by a significant margin.
That definitely continues to be the case here, where SunSpider performance breaks the 600ms mark. As a side note, two years ago, seeing Medfield run these tests in ~1300ms using Gingerbread was a major coup for Intel then. Now, single-thread performance is running more than twice as fast, though part of that is definitely due to optimizations from Google. Even the relatively pokey HTC One and Galaxy Note 10.1 are faster than the Medfield reference platform we saw at the end of 2011.
Finally, some 3D benchmarks on Android: GLBench and 3DMark.
We used GLBenchmark's Egypt off-screen test to ensure an equal playing field for every device. Here, the Z3770 SoC puts in respectable, though not top-of-the-line figures. 41 FPS is enough to tie Google's excellent Nexus 7, while the iPad 4 (Apple A6X) and SHIELD (Tegra 4) are both significantly faster, at 54 FPS and 56 FPS respectively.
3DMark Ice Storm:
In the newer 3DMark benchmark Ice Storm, we measured framerates in both of the two game tests. Here, the Z3770 is slower than most of the competition in "Unlimited" mode, but actually loses far less performance in "Extreme." This suggests that Intel's Android drivers need additional tweaking. Still, based on what we've seen in Android, it's fair to say that the CPU/GPU complex is very competitive with what we've got from any other vendor, beating out all but the actively-cooled SHIELD.
Intel's GPU is under more competitive pressure from companies with longer histories in this market. Still, Intel has always played this as a long-term game. Medfield, Clover Trail, and Clover Trail+ were all designed as products that could compete well at the midrange with Bay Trail aimed at taking competition to a higher level. Based on what we've seen, the company has succeeded.
Now let's talk Windows.
|Bay Trail Performance on Windows|
|All of our Windows tests were run using Windows 8 32-bit. There's still no Connected Standby support for Windows 8 64-bit and it's not clear why. One thing to note is that Bay Trail's overall performance is significantly higher in Windows 8, due to better driver support. These benchmarks need less introduction as they're tests we've run for years, but we'll call out the more interesting performance figures. AMD's Kaveri makes an appearance here as well.
Cinebench 11.5 gives us our first look at comparative computational performance between Kaveri and Bay Trail. Single-threaded performance is particularly interesting here. Remember, in single-threaded mode, the A4-5000 tops out at 1.5GHz while Bay Trail surges up to 2.4GHz.
This suggests that Kaveri retains a clock-per-clock efficiency advantage over BT, even though their raw performance figures are identical. We don't know what clock speed the Z3770 maintains during multi-threaded rendering, but Bay Trail's speed-up factor of 3.68 is slightly lower than Kaveri's 3.82. The implication here is that BT backs off its Turbo Mode clock modestly, but doesn't drop all the way back to its 1.47GHz base clock. Kaveri, meanwhile, runs close to 1.5GHz in both cases.
Both cores are far faster than the original Atom Z2760 and even the Core i7-2377M, single threaded.
In our LAME MT MP3 encoder tests, the Z3770 is faster than Kabini in single-threaded benchmarks while slightly slower in multi-threading tests. Again, this suggests that the SoC is aggressively power gated, and hits peak Turbo Mode speed only when running 1-2 threads. Again, the speed-up compared to Z2760 is enormous. The Core i3-2377M is still faster in both metrics, but the gap has narrowed from "painful" to "tolerable," particularly if the system in question had other desirable features.
In total system benchmark PCMark 7, the Z3770 outperforms the A4-5000 overall, despite losing the Creativity, Productivity, and Entertainment sub-tests by significant margins. That's... odd. But again, the jump over plain old Atom is enormous, and well received.
Here's a test where the Windows 8 / Android benchmark figures are substantially different. The exact same Z3770 that scored 578ms in Android 4.2.2 performs the test in 331ms using Windows 8, or just 57% the time. Whether this speaks to lingering issues with Intel's Android support or performance improvements from Microsoft is unclear. The new Bay Trail completes the test in 79% the time it takes AMD's Kabini, but is still significantly back from the Core i3. The original Z2760 brings up the rear, crying into its jar of paste.
3DMark: Ice Storm:
Finally, there's 3DMark: Ice Storm for Windows.
The "Basic" test wouldn't run on Android and the Unlimited Test isn't available on Windows, so we were left to compare against the two Extreme flavors, and vs the W8 Clover Trail tablet. The Android vs. Windows scores were essentially identical, implying the graphics core is well-utilized across both operating systems. Again, Bay Trail shows a monumental improvement over Intel's previous generation Clover Trail+ Atom architecture.
|A Few Words on Power|
|Power consumption with Intel's new Bay Trail architecture is a bit more nebulous. Understand that this is an area where Intel is obviously going to optimize for absolute peak efficiency when it comes to building test tablets for the media to evaluate. Such tablets, however, are just that -- test devices. While they should give a strong indication of overall performance, power consumption at the device level is going to come down to whether or not OEMs maximize power efficiency on every component. Swap out a different panel, or use components that can't sleep as readily to save on manufacturing costs, and the end result is a device with higher power consumption; conversely as well of course, if component selection is tailored for efficiency.
Bay Trail fully powered.
As such, these power consumption figures should be taken with a grain of salt. While we believe they are representative of what OEM hardware can offer, they are not the final word. It should be noted that while the figures below are both for the entire platform as measured at the battery, it's difficult to believe they include the actual display. A 2560x1600 display draws far more power than a 1280x800 -- IGZO and advanced power management techniques may well shrink that gap, but it's hard to believe they've eliminated it entirely.
Bay Trail, with CPUs mostly shut down but graphics core active
Intel's gaming benchmarks suggests that Bay Trail spends a great deal more power on the GPU than Clover Trail ever did, and these are the workloads where BT drives past CT. Performance per watt, however, is far better thanks to 2-3x the performance, period. In the two gaming benchmarks Intel showed, Bay Trail tied Clover Trail power consumption in one test and was worse (1.5W vs. 1.35W) in the other.
|Conclusion: A Winning Solution From Every Angle|
|Bay Trail, in most respects, the Atom solution we've been waiting five years to get our hands on. The Windows numbers show this in particular -- Clover Trail, even at its strongest, left an enormous gap between "netbook/tablet" and "notebook" performance. Intel's repeated insistence on improving battery life and integration at the expense of everything else made it increasingly hard to accept the original Atom's architectural flaws and performance drawbacks. Bay Trail fixes them and then some.
To be fair to AMD, Kabini has done a fine job fixing a lot of the same problems, too -- but a full comparison between the two platforms will have to wait until we've got our hands on shipping hardware. Based on what we've seen thus far, Bay Trail crushes Clover Trail. It dramatically boosts performance compared to its predecessor and its serious competition for AMD's own tablet and netbook hardware. This could be the chip that bridges the gap between thicker, hotter x86 tablets or convertibles and the silent, svelte devices using ARM or Clover Trail. That's 100% upside for the consumer.
None of this means Bay Trail will automatically be successful, as such. Right now, Windows tablets are facing stiff headwinds. If they weren't, Intel wouldn't have moved Android up from second string (its inarguable position 12 months back) to equal player. But what Intel has done here is create a tablet that can hang with the competition in Android and offers its own set of significant benefits to the Windows platform. There are still a lot of questions swirling around Windows 8.1's uptake, battery life and power consumption of shipping devices, as well as what Qualcomm and NVIDIA might unveil over the next few months, but Bay Trail is a huge leap forward for Intel on virtually every front. This should be an interesting Q4 indeed.
*Note: Portions of this article were contributed by Marco Chiappetta and Dave Altavilla