For the past few months, details on the Goldmont architecture (that’s the Atom microarchitecture refresh that replaced Silvermont this year) have been extraordinarily scarce and hard to come by. After doing full deep dives on past architectures, including the Silvermont update that powered Bay Trail and Cherry Trail devices, Intel refused to talk about Goldmont in any significant degree at all. While it’s true that these chips are being relegated to the lower-cost Pentium and Celeron lines, Goldmont hardware will still drive millions of devices over the next few years — and it’s been interesting to watch the evolution of Intel’s small-core chips in relation to their big-core counterparts.
Now, thanks to Intel’s updated x86-64 programming guides, we’ve gotten a look at what the new chip can do and how it differs from Silvermont, which first debuted in 2013. Goldmont incorporates a number of improvements over Silvermont, though some of the diagrams are a bit bare-bones compared with what Intel typically creates for consumer publications:
Where Silvermont (and AMD’s Kabini / Jaguar / Puma) were all dual-issue decoders, Goldmont has three decoder units, and a maximum of 20 bytes decoded per cycle. The fetch and instruction cache pipelines are no longer coupled, large page support have both been added, and there’s a small L2 “precode” cache (16K) that didn’t exist on prior Atom processors. Goldmont’s triple-wide decoder is matched by its ability to retire up to three instructions per cycle, and the chip is capable of executing one load and store per clock cycle (Silvermont could only perform one load or store per clock cycle). Three simple integer operations can be executed per cycle and address generation is now out-of-order in Goldmont (Silvermont generated and scheduled memory addresses in-order, but could complete them out-of-order.)
Goldmont also has generally improved instruction latencies (how much depends on the instructions in question, but some of the gains are considerable) and can decode a maximum of two branches per cycle (Silvermont was limited to one). Overall, Goldmont is a much smaller gain over Silvermont than Silvermont is compared with the original Atom core, Bonnell — but it’s only fair to note that significantly less time has passed between the debut of Bonnell (2008) and Bay Trail (2013) as compared to Silvermont (2013) and Goldmont (2016). Intel also hasn’t gone through nearly as many process node transitions. Bonnell was a 45nm core compared with Silvermont’s 22nm, whereas Goldmont is a 14nm chip.
The Goldmont-Silvermont shift is conceptually similar to the upgrades AMD made to its 40nm “Bobcat” CPU when it built the 28nm Jaguar follow-up. While the two companies made different changes to their underlying architectures, in both cases, Intel and AMD chose to enhance and upgrade the basic designs they’d previously deployed rather than making a huge set of changes or taking a dramatic leap forward.
Earlier this year, I speculated that Intel may have been forced to pull Goldmont CPU clocks downwards compared with Silvermont, because the CPU performed more work per cycle and had more difficulty hitting high frequencies within previous TDP ranges as a result. This could still be true — generally speaking, we’d expect Goldmont to be at least modestly faster than Silvermont on a clock-for-clock basis, and higher efficiency CPU architectures often use more power at the same clock speeds than lower-efficiency cores.
How much any of these issues specifically impacted Goldmont, or whether the new architecture’s performance influenced Intel’s decision to kill Atom’s smartphone and tablet hardware divisions is still open to discussion. Checking Intel’s various Atom pages, it’s clear that “Atom” is being phased out as a separate brand — new chips haven’t been launched in the desktop or conventional mobile markets for years, and there are only a handful of rebranded 28nm Rockchip designs on the Smartphone and Tablet Ark page. That said, Intel is offering a 1.2GHz 28nm chip as part of its Rockchip agreement, where it previously topped out at 1.1GHz. Presumably it created the higher-end SKU for a reason, but whether or not the company is shipping any volume on these parts at all is an open question. It probably isn’t.
As for whether the Goldmont architecture has a future past 14nm or not, Intel really hasn’t said. On the one hand, having a low-end Pentium and Celeron-class core available lets Intel position those systems as meaningfully different (and much less expensive) than Core i3/i5/i7, and it may continue to design Goldmont-derived CPUs for that reason alone. On the other, however, both Intel and AMD have moved away from multi-CPU architectures as they’ve abandoned the tablet and smartphone market. It may be that in the future, Intel will simply address this space with derivatives from Core chips, just as it used to do before Atom came on the scene back in 2008.