With its New RISC-V Processors, SiFive Bets on Compute Density
Aiming to usurp Arm processors in size-constrained, compute hungry designs like wearables, SiFive has expanded its RISC-V “Performance” line of processors.
The momentum for RISC-V continues to grow, as a variety of technologies and products have emerged this year leveraging RISC-V technology. Feeding that trend, today SiFive announced the addition of two processors—the P670 and P470—to its Performance line of RISC-V IP-based processors.
In this article, we examine the details of the new processors, we compare them to their Arm counterparts, and we share insights from our interview with Drew Barbier, Senior Director of Product Management at SiFive.
SiFive has added the P670 and P470 to its Performance line of RISC-V processor IP.
New Features Expand Application Focus
The P670 IP adds to the company’s P600-series line, the first of which was the P650. The company provided a sneak peak of the P650 at last year's RISC-V Summit. Today’s announcement also includes SiFIve’s P470, which the company has not previously shared publicly.
According to Barbier, the new processors are aimed at meeting the need for performance and efficiency, but in a format that suits small size and high volume system designs. Example applications include wearables, smart home, industrial automation, AR/VR, and other consumer devices.
Wearables is the use case that best exemplifies where Barbier says that these new processors will have an edge over equivalent counterparts from Arm’s Cortex processor line.
There are a number of new features that the P470 and P670 products share. Both support virtualization, including a separate IOMMU for accelerating virtualized device I/O. The processors implement full, out-of-order, RISC-V vector implementation, based on the ratified RISC-V vector v1.0 specification.
Other features the two processors share are RISC-V vector cryptography extensions and SiFive’s WorldGuard system security functionality. Importantly, the processors are the first on the market to support full RISC-V RVA22 profile compliance, says Barbier. The processors also feature a new Advanced Interrupt Architecture (AIA) compliant interrupt controller, which includes better support for Message Signal Interrupts (MSI) and virtualization.
P670 Targets Wearables and More
According to Barbier, the P670 is well suited for applications like advanced wearables, networking, robotics, and mobile. The P650 differs from the P650 by including a vector unit ALU. Specifically, the processor has two 28-bit Vector ALUs that are compliant with the ratified RISC-V Vector v1.0 specification.
The P670 RISC-V processor includes two ALUs that are compliant with the RISC-V Vector v1.0 spec.
P670 offers a maximum frequency of 3.4 GHz in 5 nm, resulting in performance of greater than 12 SpecINT2k6/GHz. The device offers higher single threaded performance and 2x compute density compared to legacy solutions. We’ll discuss the “computer density” aspect more later.
Efficiency-focused P470 Processor
SiFive is positioning its new P470 as its “first efficiency-focused Out-of-Order, area optimized, vector processor.” Like the P670, the P470 is aimed at applications like wearables, consumer, and smart home devices.
The P470 is significantly smaller than competing solutions, says Barbier, and it is optimized for its efficiency and area density. He says the P470 was designed to also serve as a companion to the P670 processor, when there’s a need for demanding computing and a sharing of computing resources.
The P470 offers a maximum frequency 3.4 GHz in 5 nm, and greater than 8 SpecINT2k6/GHz, within a minimal area and power envelope. The device provides one 128-bit RISC-V Vector ALU compliant with the ratified RISC-V Vector v1.0 spec. SiFive also plans to release the P450 in the future—an area-optimized version of the P470 that excludes the vector unit.
All About Compute Density
According to Barbier, the best way to evaluate embedded IP processors like the new P670 and P470 is in terms of compute density. With that in mind, SiFive has done side-by-side comparisons to illustrate how these products stack up.
The charts below compare single thread performance, all measured in 7 nm process implementations. “We do that because that's where we can find the public competitive data with Arm processors,” says Barbier. “There's things like die teardowns and public devices that we can go and buy that are implemented on 7 nm and we can run benchmarks on them.”
Combining SpecINT2006 performance benchmarks and die area gets you a compute density metric, with the P470 at 4x more than the Arm Cortex-A55. (Click image to enlarge)
As the charts above show, you can compare the Arm Cortex-A55 vs. the P470 in terms of peak single-there performance based on the SpecINT2006 benchmark. You can also compare the two processors in terms of die area (mm2). “When you combine those two metrics, we start to see these really significant performance uplifts over the competition,” says Barbier.
By taking the peak single-thread performance number and dividing it by the area, you get a compute density number. “When you do that and you look at the P 470 compared against the Arm A55, it's something like a 4x compute density advantage,” says Barbier. “We think this is compelling.”
P670 Compute Density vs. Arm Cortex-A78
For the new P670 processor, SiFive chose the Arm Cortex-A78 processor to make compute density comparisons. In this case, Barbier points out that they aren’t even claiming the P670 to be higher performance than an Arm A78. It’s actually a little bit less in terms of peak performance compared to an A78—about 5%.
Comparing SiFive’s P670 RISC-V processor with Arm’s Cortex-A78, the P670 offers a 2x improvement in computer density. (Click image to enlarge)
All that said, if you look at the P670 die area, it's significantly less. It’s less than half the area of an Arm Cortex A78. “When you put this together, we still see a nice 2x compute density improvement over something like the Arm Cortex-A78,” says Barbier.
A Reason to Make a Switch?
Because in general RISC-V technology is the new kid on the block, there has to be a compelling reason for design engineers to switch from a tried and true architecture like the Arm family to something new. Comparing simple stats like clock speeds aren’t enough these days. SiFive’s new processors perhaps offer a good argument for applications where size constraints and performance needs go hand-in-hand.
All images use courtesy of SiFive