If you liked the Marvell PXA1928 SemiAccurate told you about at Computex, you will love the new PXA1908 and PXA1938 SoCs. As the names suggest, one is aimed lower and the other aimed higher than their predecessors in the 64-bit LTE SoC stack.
These three chips all have core similarities but are in fact different dies mainly for cost reasons. All use 64-bit ARM A53 cores, all are built on an undisclosed 28nm processes, have a multimode Cat4 modem, and all use Vivante GPUs. From there things diverge for market segmentation reasons so take the 1928 as a mid-point and move as appropriate from there.
The low-end 1908 is aimed at the sub-$100/1000RMB phone arena and it is squarely meant to play there and only there. This differs from the 1928 which could play in that arena but could also go above it, the 1908 isn’t really suited for the >$100/1000RMB market. Why? Features or more to the point, lack thereof.
Starting off with four A53 cores at 1.2GHz make you wonder what low-end means any more. How long ago was it that 2x A9’s were state of the art? I think it was around February or maybe a tad before but I could be a bit off there. The GPU is a Vivante DC7000UL and the video engines are smaller and less capable than their PXA1928/1938 brethren which limits the display to 720p and the camera to 8MP but remember this is for a really cheap phone. If anything the 1908 raises the bar a bit here too.
One interesting cost reduction bit is that the battery charger is integrated into the PMIC. This drops board cost, area, and development time for the OEMs. Not only do they not have to source a charger and components, the one they get is supported and ‘free’ with the SoC. It is things like this that can make enough difference to win a socket in the ultra-low cost phone space.
Moving on to the PXA1938 we go up to the top of the current Marvell stack. This one has eight A53 cores running at 1.5GHz, no asymmetry here. The GPU is a Vivante DC7000L and a much more capable video engine too. We are not sure what the difference between the L and the UL are other than the L being more capable and significantly larger in area. The 1938 will support full 1080p displays and image sensors in the 13-16MP range with a slew of features ‘required’ in the target market.
While it lacks the charger in the PMIC, the 1938 adds two new features to the mix, a sensor hub and a DRM/secutity processor. The sensor hub is nothing amazing, just a very low power controller to aggregate the sensor data and pre-process it to avoid waking the CPU. Most new SoCs have it and it saves a ton of power if you actually use the sensors on the device. Since most devices have them always on, a sensor hub is almost a necessity if you are going to include sensors. A sub-$100/1000RMB phone probably won’t have many if any sensors at all, a higher priced phone will.
The DRM/Security chip is an interesting addition too, it is not the ARM Trustzone MMU based tagging but a separate microcontroller on the die. This Marvell core is basically a gatekeeper for the system that monitors fabric traffic at certain undisclosed points and does certain undisclosed things when called for. Whatever unspecified things this controller does is not proprietary, it is fully TEE compliant so software should not be a big issue.
Moving back to the commonalities we have one more thing that both SoCs, and presumably the 1928 share, a socket. Both the 1908 and 1938 are pin-compatible with a lot of platform level sharing, OEMs love this. Amortizing development costs across both a mid-range and low-end device is key to making the numbers work in the cutthroat low-end emerging markets so Marvell is catering to that.
The 1928 shown at Computex has been sampling for a while and one would presume OEMs will have devices on display at CES. The PXA1908 just stared sampling recently and the 1938 should be sampling really soon. Devices bearing 1908s should start coming out in Q1 with 1938s following a few months later in the sub-$200 phone category. If nothing else, Marvell just reset what the low-end is, 4x 64-bit cores, Cat4 LTE, and all in a <$100 package.S|A
Latest posts by Charlie Demerjian (see all)
- Intel kills off the 10nm process - Oct 22, 2018
- HyperX ships it’s 60 millionth enthusiast memory module - Oct 15, 2018
- Bittware/Nallatech water cools 300W of Xilinx FPGA - Oct 12, 2018
- More on Intel’s 10nm process problems - Sep 17, 2018
- Intel puts out another 14nm 2020 server platform - Sep 11, 2018