Linux 4.1 Speeds Up Intel Atom SoCs


Intel Cherry Trail tablet

The release of the Linux 4.1 kernel is more significant than most, and not only because it was designated as a long term stable (LTS) release, or that it included contributions from 1,539 developers, the most in in Linux history. The release improves Btrfs file-system support for massive servers, adds encryption support to the latest ext4 file system, and offers enhanced support for Chrome OS, RAID 5/6 storage, and ACPI power management on 64-bit ARM systems.

Hardware improvements include native Nouveau acceleration on Nvidia GeForce GTX 750 graphics, as well as support for the upcoming “Skylake” Core CPUs. This performance and power-optimized “tock” answer to the similarly 14nm “Broadwell” “tick” line of 5th Generation Cores, expected to be announced in early August.

From an embedded perspective, the biggest hardware news in Linux 4.1, however, is a boost in performance of the 22nm- and 14nm-fabricated lines of Intel Atom system-on-chips. The Atom centers Intel’s Internet of Things strategy along with its lower-power Quark chip, and also aims to cut into ARM’s dominant share of the mobile device market. A Linux 4.1 Git Pull posting, a submission from Intel’s Kristen Carlson Accardi “changes one of the intel_pstate’s P-state selection parameters for Baytrail and Cherrytrail CPUs to significantly improve performance at the cost of a small increase in energy consumption.”

Faster Bay Trail and Cherry Trail 

The performance improvements, which relate to Intel’s 22nm-fabricated Bay Trail Atom and Celeron system-on-chips, as well as its new 14nm Cherry Trail Atoms, are significant indeed, according to recent Phoronix benchmarks. Phoronix tested a pre-release version of Linux 4.1 running on Ubuntu. The targets were the new Intel Compute Stick running the tablet-oriented Atom Z3735F (Bay Trail-T), as well as an Intel NUC mini-PC running on a Celeron N2820 (Bay Trail-D).

As the Phoronix benchmarks show, the kernel’s Atomic improvements actually started with Linux 4.0. Compared to tests run with Linux 3.19, there were modest gains on graphics tests for Linux 4.0 and even more of a spread for Linux 4.1. There were more significant gains on CPU load, especially on the Atom Z3000-based Compute Stick. In almost all the tests, Linux 4.1 beat or matched Linux 4.0 by a considerable margin, and there was usually a larger gap between Linux 4.0 and Linux 3.19.

The Linux 4.1 improvements derived from changing the CPU frequency scaling driver setpoint from 97 to 60. This is said to make the P-State more aggressive in increasing the power/performance states when a heavy load is encountered. Note that the current Ubuntu 15.04 version available for the Compute Stick and NUC still use kernel 3.19, so you would have to load the new kernel yourself.

The improvements bode well for the Intel Atom family, which has traditionally been faster than similarly priced ARM SoCs while struggling to keep up on power efficiency. The Atom has made gains in efficiency, especially with the latest Cherry Trail models, but the new multi-core 64-bit Cortex-A53 SoCs are also more competitive on performance, as well.

The Linux 4.1 speed boost is likely similar on the embedded Intel Atom E3800 (Bay Trail-I), which has had a wider impact in the market than the other Bay Trails. The Atom E3800 is found on Intel boards like the Intel Edison and Minnowboard Max, not to mention countless embedded systems that have adopted the E3800 over the last two years.

The P-state changes have also been made to the 14nm Cherry Trail Atoms — the quad-core Atom x5 and x7 — which are already notable for major improvements in graphics performance over Bay Trail. At Computex earlier this month, Acer demonstrated one of the first Cherry Trail devices with its Atom x7-based Predator 8 Android gaming tablet. Cherry Trail has found faster acceptance on Windows tablets, including Microsoft’s Surface 3, which sheds Windows RT for the ARM-friendly Windows 10.

We saw no mention of the similar Celeron and Pentium branded “Braswell” line of processors, but we imagine the performance gains are about the same. Braswell SoCs like the quad-core, 1.6GHz Celeron N3150 and dual-core, 1.6GHz Celeron N3050 have the same 14nm foundation and kicked-up Intel Gen8 graphics as the Atom x5 and x7, and are otherwise quite similar except for adding desktop-like features such as SATA support.

Next Up: An IoT version of the Atom x3 (Sofia)

It’s unclear whether the Linux 4.1 speedup would affect the 28nm Atom x3 (“Sofia”), which aims to take on ARM’s Cortex-A7 on the low end. The x3 models are aimed at budget smartphones and tablets sold to Asian markets, and include ARM Mali GPUs, and in some cases cellular basebands.

In April, Intel announced a rugged, IoT version of the Atom x3, which unlike the original model supports Linux, as well as the Linux-based Android. The IoT versions, which will be launched with developer kits later this year, feature seven years of extended product lifecycle support and optional extended temperature support.

Like the original Atom x3, which began shipping on a surprisingly affordable $70 Telcast X70 Android tablet in China, and is expected to arrive on some 45 different products by year’s end, the IoT versions include integrated basebands. This will enable long-range communications for sensor devices.

We’re still waiting for a true embedded version of Cherry Trail similar to the Atom E3800. So far, Braswell seems to be playing that role with TDP power efficiency ratings of 4-6 Watts. There has already been widespread Braswell adoption in recent weeks in Linux- and Windows 10-ready COM Express modules and Mini-ITX boards.

Embedded ARM escalation: Freescale’s i.MX7

Whether it’s in mobile or embedded space, Intel’s Atom has some challenging competition from ARM. In addition to the arrival of 64-bit Cortex-A53 SoCs to drive high-end tablets and phones, there has been increasing support for Cortex-A7 in embedded devices. This week, for example, Freescale updated its line of popular, Cortex-A9-based i.MX6 SoCs with a more power-efficient i.MX7. The efficiency gains, including a claimed 15.7 DMIPS/mW performance/power ratio, not only derive from Cortex-A7, but also a 28nm process.

The i.MX7 reinforces how IoT has begun to shape the semiconductor industry. This is Freescale’s first i.MX line of SoCs to actually drop slightly in performance, although it also improves security, and adds a Cortex-M4 MCU and heterogeneous core management, in addition to extending battery life.