The Atom Z3000 SoC Architecture
Alright, so we have an idea of where Intel is coming from and where it hopes to go with the Atom Z3000 series for tablets. Now let’s take a look at the six SKUs launched at IDF, along with their specifications:
|Process Tech.||22 nm||22 nm||22 nm||22 nm||22 nm||22 nm|
|L2 Cache||2 MB||2 MB||2 MB||2 MB||1 MB||1 MB|
|Core Clock Rate||Up to 2.4 GHz||Up to 2.4 GHz||Up to 1.8 GHz||Up to 1.8 GHz||Up to 2 GHz||Up to 2 GHz|
|Memory||2-ChannelLPDDR3 1066 MT/s||1-ChannelDDR3L1333 MT/s||2-ChannelLPDDR31066 MT/s||1-ChannelDDR3L1333MT/s||1-ChannelLPDDR31066 MT/s||1-ChannelDDR3L1333 MT/s|
|Peak Memory B/W||17.1 GB/s||10.6 GB/s||17.1 GB/s||10.6 GB/s||8.5 GB/s||10.6 GB/s|
|Memory Capacity||Up to 4 GB||2 GB||Up to 4 GB||2 GB||1 GB||2 GB|
As we already know from our look at Silvermont, Intel is leveraging a version of its 22 nm process optimized for the SoC architecture. Four models sport four cores (or two modules—remember, this is a module-oriented design that couples two cores and 1 MB of shared L2 cache). Two are single-module designs, with two cores and 1 MB of L2.
Rather than simply referring to base clock rate and specifying a frequency ceiling, Intel decided it was a good idea to identify each SoC’s maximum Burst Technology speed. So, the quad-core Atom Z3000s are split between processors that top out at 1.8 GHz and those able to hit 2.4 GHz.
They’re available in a couple of different memory configurations, as well. A pair of the four-core SKUs support two channels (4 GB total) of LPDDR3-1066, pushing up to 17.1 GB/s of bandwidth. That’s key to enabling display resolutions of 2500x1600. The other quad-core SoCs and one of the dual-core models accommodate a single channel of DDR3L-1333. Two gigabytes of memory running at that speed does 10.6 GB/s, which is good enough for 1920x1200. The remaining Atom Z3680 takes one channel of LPDDR3-1066, is limited to 1 GB, and only offers enough bandwidth to support 1280x800.
All of this information is new. Back when Intel introduced us to Silvermont, it avoided product-specific questions. We knew about the module-based design of course, and we suspected that tablet-oriented SoCs would be dual- and quad-core configurations. But the engineers didn’t go into detail about the memory controller, for example—and for good reason. The specifics are implementation-dependent. Bay Trail exposes up to two 64-bit channels of LPDDR3 (last generation, you were stuck with two 32-bit channels of DDR2-800). But in Intel’s server-oriented Avoton, also built on a Silvermont foundation, faster ECC-capable DDR3 is supported.
We also now know that the Bay Trail SoC hosts an HD Graphics engine with four EUs. It runs at up to 667 MHz, though the base clock rate is 311 MHz. Intel says the architecture comes from Ivy Bridge, though it’s worth noting that four EUs is fewer than the GT1 implementation of Sandy Bridge, which featured six EUs. Even still, company representatives say to expect three times as much graphics performance as the PowerVR SGX545-equipped Atom Z2760. HD Graphics adds DirectX 11 and OpenGL ES 3.0 support, whereas Imagination Technologies’ solution was limited to DirectX 10.1 and OpenGL ES 2.0.
Heterogeneous compute via APIs like OpenCL is not supported on Bay Trail for tablets—Intel says its focus was to get the platform out with a focus on compute through the x86 cores. However, its system agent was designed to support a heterogeneous architecture in the future. In fact, the Celeron we're testing today does include the OpenCL runtime in its driver package, and we were able to get OpenCL working on the HD Graphics component, too.
As with its Core architecture, Intel’s Atom Z3000-series shares one power budget between on-die subsystems. So long as the SoC is operating under its power, current, and temperature limits, the x86 cores and other IP blocks are allowed to run at higher clock rates when more performance is needed. Although you won’t realize those peak Burst Technology figures in a thermally-restricted workload, you won’t slam into a ceiling and watch the chip throttle either, as we've seen from competing SoCs. There’s certainly something to be said for more elegant power management, made possible in large part to the central system agent.
The memory controller, cores, and graphics engine all tie together through this critical component, which maintains coherency between the shared caches. Image signal processing, fixed-function video decode, and display control attach to the system agent as well.
Beyond simply beefing up 3D potential, Bay Trail also gets the power and performance benefits of Quick Sync’s fixed-function and programmable media pipeline. H.264, VC-1, MPEG-2, and MVC are all decoded in hardware, while H.264 and MPEG-2 encoding can be accelerated.
Bay Trail supports HDMI 1.4, DisplayPort 1.2, eDP 1.3, and MIPI-DSI output across a pair of display pipelines. Given enough bandwidth, Intel says its DisplayPort connectors will do 2560x1440 at 60 Hz, while HDMI maxes out at 1920x1080. Both digital interfaces do support integrated audio.
An enhanced image signal processor is rated for up to 275 MP/s with support for auto-exposure, -focus, and –white balance, along with 1080p60 video capture, video stabilization, burst mode, continuous capture, and low-light noise reduction. The outward-facing switch supports sensors up to 13 MP, while the forward-facing switch supports up to 2 MP.
In a modern Intel desktop architecture, everything mentioned above is built onto the CPU die. The processor communicates over DMI to a Platform Controller Hub loaded up with storage, USB, audio, and networking—all of the I/O typically associated with a south bridge. Bay Trail takes that functionality and moves it up into the die through a switching fabric attached to the system agent. This fabric manages bandwidth and access priority for some of the new on-board subsystems. For instance, as you saw in the block diagram above, Bay Trail’s storage hub supports SDIO 3.0, SD 3.0, and eMMC 4.51. Native USB 3.0 is a big deal as well.