Read on Twitter
Apple’s in-house Mac silicon, AWS’ Graviton, and Fujitsu’s A64FX initiatives are a clear sign of the writing on the wall. Intel has done irreversible damage to x86 — that not even a hot AMD can fix — and hugely accelerated the migration to Arm by 5-10 years with the 10nm fiasco.
Graviton and Apple’s silicon are very symptomatic. There’s nowhere to hide, not in DC, not in PCs... AMD did right by focusing on x86 for up to Zen3 (possibly Zen5), because they didn’t have volume, resources, or even SW talent to drive the foundational work.
But with the hard foundational work having been done elsewhere in the last decade, they’ll be able to ride the Arm wave the right way this time. The only question is, who will step up and lead the way in Arm PCs? AMD or Intel? Qualcomm? Nuvia? I think AMD will be the one.
Unless AMD’s management actually has tunnel vision (a different kind from Intel’s, but still), and I don’t think they do, they’ll do client Arm. Huge opportunity and easier for them than for Intel. Project SkyBridge was the right vision at the wrong time for AMD.
x86 patents are expiring either way, so, it isn’t even like being part of that very small club will still be an advantage. The rest of the industry at large has already chosen its way forward. 20%, 30%, 40% better or whatever IPC with Zen3, it isn’t about that.
Think about it, unless you’re Intel/AMD, why would you want to incentive x86? It’s a huge potential liability, as Intel has shown in the past 5 years. AMD was also a big letdown for a decade. Why would a x86 duopoly — or AMD domination — be really better than an Intel monopoly?
The “oceans of cores” strategy is not even that, it’s a short-term trick and nice side effect of the MCM and then chiplets move. Core counts ain’t increasing by the Xeon/Naples -> Rome factor or close to that again in every other year, no.
The truly major speedups and innovation are all happening in dedicated silicon. 10%, 50% better IPC, it’s all the same — it isn’t even consistent across workloads. We’re talking orders of magnitude — think 10x, 20x — by means of specialization with the slow death of Moore’s law.
When we’re talking orders of magnitude faster, it’s about enabling things that were unfeasible in computing before. It’s about windows and doors into new possibilities. 10% faster or 50% faster are about gains in what already exists and is already being done either way.
