I’m on a benchmark tear this past month. It’s just my level of excitement around the news around x86 alternatives. There are the ARM chips by Fujitsu running some of the fastest new supercomputers. There is the M1 chip by Apple. Now we have a potential new RISC-V chip by a company called Micro Magic which looks to be finally bringing performance into a range comparable to desktop-ish ARM chips. This article by ArsTechnica really wet my appetite. I wanted to see how this chip’s real world performance, assuming we take the benchmarks at face value, can compare to the CPU in the PineBook Pro (PBP).
I chose the PBP because it is an ARM laptop that I happen to own. As I’ve written elsewhere, I have been dying for a non-x86 laptop since the last PowerPC Macs were sold. It’s nothing truly against Intel or x86, I just like having architecture diversity in the world. Granted in recent years the power per Watt issues have given reasons for people and companies to start looking elsewhere. I never did a full review of the PBP but the first thing I’ll say is that “Pro” in the name is only accurate compared to it’s non-pro sister. The chip in these laptops, the Rockchip RK3399 is not a powerhouse chip. It has two Cortex-A72 high performance ARM cores and four Cortex-A53 low power cores along with a separate GPU unit. That can sound impressive but it’s not. It’s meant for smaller low power single board computer type applications and I’d put it in the same class as a Raspberry Pi. One could hypothetically make a go of it as their machine but it’d be a bit of a stretch.
In fact I did use this as my machine on a month long trip to try out the experience. It was perfectly usable when I cranked the CPU up to it’s max 2 GHz. The laptop has a frequency throttling mode which let’s idle at 400 MHz, yes mega-Hertz, and then just bump up the speed as needed. The responsiveness was so noticeably affected I’d just turn it off if I was doing anything that I wanted the system to feel responsive with. Besides playing around with running software at turn of the century clock speeds (you don’t want to do it) it really needs every ounce of CPU that it has. I’d place the overall feel to something like a 2008-2010 mid-range laptop. I happen to have one such laptop, a 2011 MacBook Air, so can do a more direct comparison. The Air feels a bit more responsive than the PBP. However for a $200 machine it’s a nice little setup.
This isn’t a review of the PBP but of the prototype of this new RISC-V chip. RISC-V is an open source RISC hardware system which some hope will one day give companies a platform to be build competitive CPUs, GPUs, and SOCs around. It was first started in 2010. In recent years there has been some a lot of movement in the RISC-V market and the first computers built around the chips are becoming available. One problem so far though is that their performance has been far behind ARM, which has of course been generally far behind x86 desktop offerings. However there has been a lot of improvements on that front from several venders. This recent announcement by Micro Magic offers some tantiliving power per Watt claims. They literally would be unprecedented levels in this CoreMark embedded computing benchmark. ArsTechnica in the above article run the benchmark against the Apple M1 chip, an AMD Ryzen chip, and a Qualcomm Snapdragon 820 chip. I wanted to round this out with the Rock64 chip at various CPU speeds.
Jim Salter in his ArsTechnica article did a great job of explaining his methodology and configurations. Like him I had problems getting turbostat running on my PBP I therefore had to do a comparable power differential calculation using another Linux tool called
powerstat. Also as in his case because of the noisiness associated with that it is hard to make a one for one claim against the Micro Magic release except to say that it is at least a factor of two lower performing per Watt than the Micro Magic claims. How did it do on raw benchmark performance though?
Because I was able to have precise control over the clock on the PBP I decided to run it through various steps from it’s slowest 410 MHz up to it’s fastest 2.02 GHz. As we can see the MM RISC-V chip at it’s upper speed range is comparable with the PBP ARM chip. I consider that a huge achievement in the RISC-V world. Previous benchmarks I’ve seen have left a lot lacking for my purposes. I was waiting for something in the Raspberry Pi 3/4 or PBP class before starting to experiment with RISC-V. If these are sustainable numbers that hold up in independent tests Micro Magic has produced a CPU which is capable of meeting this level of performance. Therefore I could easily see spending some money to start dabbling with this architecture.
That said, it’s clear that this level of performance is still a far cry from Intel laptop chips. The Intel i7 benchmarked next to it is a mobile(ish) chip from 2016, four years ago. Yes that laptop feels a little sluggish compared to a modern laptop but still nowhere near as slow as the 2011 MacBook Air or the PBP, all in that order. So while this may be a great achievement for RISC-V we won’t be seeing this in a competitive consumer laptop or desktop at this point. That’s not to take away anything from the achievement or my excitement for it. Also, as stated previously, since we need to wait for independent reviews confirm the performance and running other real world benchmarks on it to see how this chip performs in the wild. All that aside this still is a great indication of what can be done with the RISC-V architecture and promise for performance growth as more people start working with it.