New RISC-V processor claims record performance per watt

Micro Magic Inc., a small electronics design company in Sunnyvale, Calif., Produced a prototype processor that is several times more efficient than its major global competitors, while maintaining reasonable raw performance.

We first noticed Micro Magic’s claims earlier this week, when EE Times reported on the company’s new processor prototype, which appears to be the world’s fastest RISC-V processor. Micro Magic advisor Andy Huang claimed that the processor could output 13,000 CoreMarks (more on this later) at 5 GHz and 1.1V while outputting 11,000 CoreMarks at 4.25 GHz – the latter, while not consuming only 200 mW. Huang demonstrated the CPU – running on an Odroid board – to EE Times at 4.327 GHz / 0.8 V and 5.19 GHz / 1.1 V.

Later the same week, Micro Magic announced that the same processor could produce over 8,000 CoreMarks at 3 GHz while consuming just 69 mW of power.

OK, but what is a CoreMark?

Part of the difficulty in assessing Micro Magic’s claim for its new processor lies in determining what a CoreMark is and how many of them are needed to make a fast processor. It is a deliberately simplified processor benchmarking tool published by the Embedded Microprocessor Benchmark Consortium, designed to be as platform neutral as possible and simple to build and use. CoreMark only focuses on the basic pipeline functions of a processor, including basic read / write, integer, and control operations. This specifically avoids most of the effects of system differences in memory, I / O, etc.

The Embedded Microprocessor Benchmark Consortium (EMBC) is a group with broad industry representation: Intel, Texas Instruments, ARM, Realtek and Nokia are some of its most notable and easily recognizable members.

Now that we understood all of that, the next step to better assess Micro Magic’s claims was to run our own CoreMark benchmarks. All we had to do here was clone its GitHub repository and then issue a make command: optionally, with arguments XCFLAGS="-DMULTITHREAD=8 -DUSE_FORK=1" if we want to test on multiple threads / cores at the same time.

I always have an Apple M1 Mac Mini on hand, as well as an Acer Swift 3 powered by Ryzen 7 4700U, so these are my test systems for comparison. Getting the raw performance scores was considerably easier than getting really comparable power readings. On the Linux system powered by Ryzen I used the utility turbostat to get the kernel and package power readings while the tests are running.

I don’t have access to anything as fine as turbostat for the Apple M1, so for this platform I took the power consumption of the whole system on the wall and just subtracted the desktop slow motion reading from the sustained reading during the test. This is extremely rude, and I caution readers not to be overly reliant on comparing the efficiency of the M1 to that of the Swift 3 on these numbers alone – but it’s good enough to have some perspective on the claim. from Micro Magic for its new RISC-V (pronounced risk five “) CPU.

Let’s go to the tests!

The Micro Magic processor is, at the moment, single-core and single-threaded – although Huang says it could “easily” be built as a 25-core piece. Micro Magic provided figures – and in one case, a screenshot – for performance at 3 GHz, 4.25 GHz, and 5 GHz. At the maximum power efficient 3 GHz clock rate, the Micro Magic processor gets about 1/4 of the CoreMarks of the Ryzen 4700u or Apple M1. At the extremely capable 5 GHz clock, it manages just over a third of their performance.

This is enough to let us know that the Micro Magic chip in its current form is not a world-class competitor to traditional ARM and x86 processors in phone or laptop applications – but it is much closer to them than previous RISC-V implementations. . At the energy-efficient 3 GHz clock frequency, the Micro Magic processor is almost three times faster than, for example, SiFive’s Freedom U540 single-threaded processor. At 5 GHz, it exceeds the four cores of SiFive.

With roughly a quarter of the performance of the world’s leading x86 and ARM mobile processors, the Micro Magic processor doesn’t ring a bell yet. But when you factor in energy efficiency, things get crazy. I gave my Ryzen and Apple processors the benefit of every possible doubt when generating the graphics above – I used base horsepower (not full package horsepower) on the Ryzen 4700U and I performed tests with the Gnome3 desktop turned off. For Apple, I only had access to the power consumption of the whole system, so I subtracted the “inactive desktop” power consumption from the “under test” power consumption.

I tested both single-threaded and multi-threaded Apple and AMD processors during the energy efficiency check. Unsurprisingly, both sides produced more performance per watt when used with one worker thread for each available CPU thread. None of this has made much of a dent in Micro Magic’s dominant lead in energy efficiency.

At 4.25 GHz, the Micro Magic can accomplish the same workload as the Ryzen 4700U with less than a third of the power required. At 3 GHz, that number drops to less than one –eighth the power required.

What’s the point?

The Linux operating system already supports the RISC-V architecture. So, for headless or near-headless controllers that just need to deliver decent performance coupled with extreme power efficiency, Micro Magic’s new processor probably is most of the time. Things get complicated considerably once you start talking about whole, user-friendly systems, of course. Even apart from hardware considerations like GPU and LTE modem, building a full Android phone based on a non-ARM architecture is probably a much bigger undertaking.

That said, it’s worth pointing out that, if we take Micro Magic’s numbers for granted, they already beat the performance of some solid processors for mobile phones. Even at its first clock rate of 3 GHz, the Micro Magic processor outperformed a Qualcomm Snapdragon 820. The Snapdragon 820 is no longer world class, but neither is it – it was the processor. of the US version of the Samsung Galaxy S7.

If we use the single-core score published by EMBC for the Snapdragon 820 with the result of Anandtech’s single-core processor power test, we get around 16,000 CoreMarks per watt. That’s triple the efficiency of the Ryzen 4700u running single-threaded, and a bit better than when the Ryzen is performing a multithreaded workload optimally.

In other words, Micro Magic’s prototype processor is both significantly faster and considerably more power efficient than a reasonably modern and still high performance smartphone processor.

Conclusions

It all sounds very exciting – Micro Magic’s new prototype delivers solid smartphone-grade performance at a fraction of the power budget, using an instruction set that Linux already runs natively on. Best of all, the company itself is no stranger.

Micro Magic was founded in 1995 and was purchased by Juniper Networks for $ 260 million. In 2004, it was reborn under its original name by the original founders – Mark Santoro and Lee Tavrow, who originally worked at Sun and led the team that developed the SPARC 300 MHz microprocessor.

Micro Magic intends to offer its new RISC-V design to customers using an IP licensing model. The simplicity of the design – RISC-V requires about a tenth of the opcodes that modern ARM architecture does – further simplifies manufacturing issues, as RISC-V CPU designs can be built as a shuttle, sharing space on a wafer with other designs.

That said, it would be a huge undertaking to port – for example – an entire ecosystem of smartphones, like commercial Android, to a new architecture. In addition to creating the operating system itself (not just the kernel, but drivers for all hardware, from GPU to Wi-Fi to LTE modem, etc.), third-party application developers should also recompile their own applications for the new architecture. .

We also still take a good chunk of Micro Magic’s claims at face value. While we saw a screenshot of a CoreMark score of 8,200 and saw a power reading of 69mW, it’s not entirely clear that the power reading was representative of the whole. of the reference run.

Yet it is an exciting development. Not only does the new design appear to perform well while massively breaking efficiency records, but it does so with a design that is much more ideologically open than its competition. The RISC-V ISA – unlike x86, ARM, and even MIPS – is open and released under royalty-free licenses.