New entrants to the CPU market have been popping up in the past few years as multiple firms tap the ARM CPU architecture to challenge Intel and AMD.
New entrants to the CPU market have been popping up in the past few years as multiple firms tap the ARM CPU architecture to challenge Intel and AMD. Nuvia, which came out of stealth mode last fall, hasn’t said much about their designs, but now that the company has unveiled its performance expectations, it’s clear they aren’t aiming small.
Nuvia compares the performance and power consumption of multiple solutions based on Geekbench 5 data. They argue that this application can properly model per-core performance in a way that’s representative across iOS, Android, and Windows platforms, and that it’s a broad enough test to have relevance in server markets given how workloads have converged between client and server systems. As always, we highly recommend taking these estimates with a grain of salt: I’ve never been sure if Geekbench is a good cross-platform comparison between Windows versus iOS versus Android, if only because we have so few of them it’s difficult to evaluate their relative accuracy.
Here’s what the image shows in aggregate: Skylake is far less efficient than Ice Lake and Zen 2, which are more-or-less equally efficient to each other. ARM cores are positioned in a fundamentally different spot of the graph and they’re far more efficient than their x86 counterparts. One thing I want to point out, however, is that the ARM SoCs Nuvia measured have big.Little core implementations that shift them to low-energy cores at idle. Nuvia does not discuss this difference between the x86 and ARM SoCs, but forcing the Apple and Android devices to idle exclusively on their “big” cores the way the x86 CPUs do might have changed the starting point for those chips.
Even if true, however, this doesn’t change the fact that the ARM CPUs scale much more efficiently than the x86 CPUs do in terms of performance gains per power expended. The question is: Is this relevant?
Here’s what I mean: According to a 2014 analysis, ISA is not a significant factor in CPU power efficiency. Above the microcontroller level, MIPS, ARM, and x86 were all more-or-less equal in terms of power efficiency. What matters, according to the researchers, is the architectural decisions a company makes to build its chips.
The biggest difference between the x86 CPUs fielded by AMD and Intel and the ARM-based chips Apple and Qualcomm are fielding is the frequencies that they hit and hold. According to Anandtech, Apple’s A13 holds a clock speed of ~2.6GHz under load. AMD and Intel build CPUs that can maintain 4.2 – 4.5GHz clocks under full load. Hitting those kinds of frequencies requires making tradeoffs on efficiency at lower clocks, and it likely explains some of the reason the curves look different for ARM versus AMD hardware. Nuvia’s entire argument is that while x86 CPUs outperform existing ARM solutions by 40-50 percent, they burn far too much power to do it.
I’m curious to see if the CPU manufacturers scaling up ARM solutions can deliver anything like the performance gains they project. Both AMD and Intel continue to promise generation-on-generation improvements with Zen 3 and Tiger Lake, respectively, but Nuvia believes its upcoming hardware will compete effectively even when these are taken into account.