Earlier this week, we wrote about SiFive Shield open security platform as the equivalent of Arm TrustZone security technology, but the company had had another important announcement this week with the introduction of SiFive U8-Series Out-of-Order (OoO) RISC-V Core IP with much higher performance than the company’s earlier U7-series core and competing with Arm Cortex A72 core.
At first, the company will offer two standard cores with SiFive U84 RISC-V core optimized for power efficiency and area efficiency, and the SiFive U87 RISC-V core with vector processing.
SiFive U84 Performance & Efficiency
We do not have much information about U87, but SiFive already published some interesting details about U84 cores.
SiFive U84 core offers about 3.1 times higher performance compared to their earlier U74 standard core thanks to a 2.3x increase in IPC combined with a 1.4x increase in maximum frequency. Compared to SiFive U54, U84 delivers 5.3x higher performance (isolated process) and when taking into account the improved process technology, the SiFive U84 standard core (7nm) is 7.2x faster than SiFive U54 (28nm).
Taking a quad-core SiFive U84 CPU with 2MB of L2 cache as an example, SiFive explains the chip would require 2.63mm2 area in 7nm process technology and achieve up to 2.6GHz clock speed. Compared to Arm Cortex-A72 @ 16nm, the SiFive U84 standard core is said to double the area efficiency, and improve performance/watt by 1.5 times.
SiFive did not provide direct performance comparison against Cortex-A72, and Arm certainly holds the lead in terms of performance with their newer Cortex-A76/A77 cores, but SiFive U84 is the first RISC-V core form the company to offer Cortex-A72 class performance.
SiFive U8-Series Scalability and Mix+Match Technology
While the U84 standard core is configured with a 12-stage pipeline and triple-issue capability, SiFive U8-Series Core IP will allow designers to configure options inside the Out-Of-Order design space hyperplane. The Core IP also features composable caches for real-time operation.
Similar to Arm big.LITTLE / DynamIQ, or Intel Foveros, SiFive Mix+Match technology will allow designers to combine various types of cores in a RISC-V SoC, for example mixing U8-Series cores for performance with U7-Series cores for power efficiency, and an S2-series core for real-time I/Os.
The new SiFive U8-Series Core IP is expected to be found in automotive, enterprise, and edge or endpoint AI markets. There’s no product page just yet, but you may still find a few more details in the announcement.
Jean-Luc started CNX Software in 2010 as a part-time endeavor, before quitting his job as a software engineering manager, and starting to write daily news, and reviews full time later in 2011.
Support CNX Software! Donate via cryptocurrencies, become a Patron on Patreon, or purchase goods on Amazon or Aliexpress
Great to see RV pushing in the direction of big cores. Now, we need volume production of the likes of Unleashed boards, so that prices are not sitting in the ‘lab equipment’ range.
They need a Risc- V media player ( needs another companies GPU, VPU ) Roku or Android.
They’re an IP house so they need a partner to release a mass market chip. If they release their own chip, it would kind of be competing against their own customers and make them not an IP house. Design complete to silicon is about a year so I’m guessing we’ll see it at the end of 2020.
Then you have the problem of software optimization. To this day, ARM64 is still not an optimization target for most software. You only have generic compiler NEON optimizations. Given the flexibility of RISC-V, it would also make optimization even more difficult.
I’m sorry but an area comparison of 16nm vs. 7nm and then claiming double area efficiency is just rubbish misleading marketing. If you want to compare, compare apples to apples, not apples to swedes.
Who cares? We buy by specs and cost-performance. Not by the production method. If it’s rolling off SMD reels and has a reasonable power circuitry, for all I care they can make the damn thing out of Cheerios and unicorn glue at 500nm.
Not exactly, Theo is right here. While it’s nice that they managed to create more efficient cores, it’s pointless to compare them to a 3-year old core and claim “twice as efficient”. Either they compare to something recent of the same price range and power envelope (maybe they are twice as fast as a modern A55 for the same price and power) or they compare to something recent of comparable design (e.g half the performance of A77 for half the power and 1/4 of silicon). But comparing with different generations makes no sense. If someone were to etch an A72 at 7nm it might have better characteristics than this new chip and would be available instantly with an optimized compiler plus existing software waiting for it.
If ARM stopped marketing/licensing cores n years after release you might have point but here SiFive is basically advertising they have something anyone that might be considering a design that ARM is actively marketing should also consider. It might not be a fair technical comparison but if they have something that is as good as an A72 and they already have it working on a better process then theorising that the A72 would be better on a similar process is neither here nor there because it doesn’t exist.
It’s not what I said. I basically said that they’re announcing that something not yet available is as good as something old. It’s comparing oranges and apples and means nothing. They’d rather compare to something of similar generation so that users can get an idea of how it compares. And I think that comparing to A55 should give them an advantage. Or maybe comparing to A76/A77 showing they’re not far away could help. But saying “take an old A72 with large features, compare, we’ll soon do better” just leaves me with a negative opinion on what their upcoming product.
>not yet available is as good as something old
The A72 is still available. If you were considering an A72 in your design (because it’s cheaper than an A76/77) would you tell SiFive were to go when they offered you something with equal or better performance, lower cost and so on because you think they cheated?
This is way faster than I expected. I thought they start small into the microcontroller market. This is going to destroy ARM in the long run. Or maybe even faster?