AMD launches new Beema, Mullins SoCs: Higher performance at almost-low-enough TDPs

 

AMD’s Beema and Mullins, both of which officially launch today, are an iterative improvement to the low-power Kabini (notebook) and Temash (tablet) SoCs it shipped nearly a year ago. Normally, follow-up launches like these implement minor frequency boosts or offer slightly better power consumption. With Beema and Mullins, AMD is promising a great deal more than that. These two new cores (Mullins is the tablet chip, Beema is for laptops) offer vastly improved performance per watt.

Neither SoC is a major departure from the previous architecture. Cache structure, branch prediction, and instruction set compatibility is the same between Kabini/Temash and Beema/Mullins, and the new chips do not support HSA. What AMD has done instead is innovate around the edges. Turns out that if you do that right, you can still turn in some impressive gains, year-on-year.

New features

Turbo Core: AMD’s Kabini and Temash could reduce their own clock speeds to save power but didn’t have a Turbo Mode for additional performance in single-threaded workloads. Beema and Mullins both add this capability to certain chips — Beema, the notebook processor, can burst up to 2.4GHz while Mullins, the tablet SoC, can ramp as high as 2.2GHz.

ARM TrustZone: Mullins and Beema are the first AMD processors to integrate a Cortex-A5 on-die for additional system security and management. TrustZone is analogous to Intel’s Trusted Computing technology — ARM’s own website says that the feature is analogous to the Intel standard. This is essentially a corporate or government-oriented feature; there doesn’t seem to be much consumer software that actually uses the TrustZone system.

Reduced leakage: AMD claims that it’s reduced leakage current loss by 19% in Mullins as compared to Kabini. This isn’t’ the same thing as reducing total power consumption, but it should still have a measurable impact. The on-board GPU has improved even more; Beema and Mullins have 38% reduced leakage compared to Kabini/Temash.
A number of additional improvements were made to reduce power consumption in other areas — the display controller now draws less power when using DisplayPort and low-power DDR optimizations allowed AMD to reduce memory controller power consumption by 600mV compared to standard DDR modules.

New power management: This ties into the Turbo Core feature but is distinct enough to deserve its own mention. According to AMD, Beema and Mullins will include the ability to directly measure the skin temperature of the laptop or tablet and will adjust frequency based on how warm the chassis is — not just according to their own Tmax. Because heat dissipates out to the chassis rather slowly, AMD can therefore run their cores at a higher frequency for a longer period of time. According to AMD, Tskin will be a user definable variable (it’s not clear how this capability will be exposed in software).

Performance and TDP

AMD wasn’t able to provide us with a notebook for testing, so our insight into performance is somewhat constrained. The company did provide a robust set of data for three benchmarks — 3DMark 11 (graphics), PCMark 8 Home (traditional workloads) and Basemark CL (GPGPU compute).
AMD’s published results show modest gains from Kabini to Beema and Temash to Mullins. Graphics workloads gain the most — Beema is up to 10% faster than Kabini in 3DMark 11 and the new A4-6210 is 2% faster than the A4-5000 in PCMark 8. The gains are strongest at the lower end; the E1 Micro-6200T is 24% faster in 3DMark 11 and 9% faster in PCMark 8 than the 3.9W Temash it replaces.

 

 

The real improvements Beema and Mullins offer are on the power consumption side. The A6-6310 APU offers roughly the same performance as the A6-5200, but in a 15W TDP envelope, not 25W. The A4-6210 replaces the A4-5000, keeps that chip’s 15W TDP, but offers a 1.8GHz clock speed and 600MHz GPU (up from 1.5GHz and 500MHz respectively). The tablet improvements are even more impressive; the old Temash family had one dual-core chip at 3.9W but the other entrants were all 8W chips. The new Mullins chips, in contrast, have a maximum TDP of 4-4.5W and substantially more CPU and GPU headroom.
The full range of Beema and Mullins processors are listed below:

 

AMD’s “Max” column for Beema is a bit misleading; it implies that every Beema SoC has a defined Turbo Core. That’s not actually the case — only the A6-6310 has a Turbo Core of 2.4GHz with a baseline clock of 2GHz. The A4-6210, E2-6110, and E1-6010 all operate at a steady clock of 1.8GHz, 1.5GHz, and 1.35GHz. It’s not clear why AMD didn’t enable the feature for these chips, but the 1.8GHz A4-6210 should still offer a significant improve over the older A4-5000 thanks to a higher clock on both CPU and GPU.

 

All of the new Mullins cores have a Turbo Mode with base frequencies in the 1-1.4GHz range and maximum frequencies as shown.

Will this refresh stave off Intel’s massive tablet push?

Beema and Mullins still draw too much power to fight Intel’s Atom across the entire range of its market, but AMD has clearly focused on pushing the chip into that 4.5W space in order to enable fanless systems and tablet designs. A quick check of Newegg shows that Kabini did fairly well over the past 12 months; there are multiple systems packing various iterations of the processor. The company has virtually no presence in tablets; there’s a lonely MSI tablet with an A4-1200 in it as compared to 229 separate Atom SKUs.

With Intel openly advertising the fact that it’s shipping tablets at contra revenue (meaning below the total cost of manufacture) in order to build sales volume and market share, AMD runs the risk of being steamrollered by its larger, richer rival. The charts and graph below show AMD’s historic improvement in power management and gives us an intriguing look at what the company is going to implement in future processors.

 

The purple bits are the most interesting, though none of them have timelines attached. AMD is apparently working on integrating its voltage regulator on-die, as Intel did with Haswell, and adapting inter-frame power gating to better manage power consumption in specific workloads. The idea of inter-frame gating is to use software to estimate the minimum power and frequency targets needed to deliver a workload within an acceptable amount of time, and then set those levels before processing begins. It’s another component of the “race to idle” concept.

Hopefully these improvements will help AMD gain traction in the tablet space, but it’s going to be facing an uphill fight on that front. The company has yet to announce a follow-up SoC for 2015 in this space — presumably we’ll see HSA integrated at that point, but it’s still not clear.