Intel updated its Xeon roadmap yesterday at a virtual DCAI (Data Center and AI Investor Webinar) that confirmed that it is working on a 144-core Xeon processor currently called Sierra Forrest and expected to come in the first half of 2024. What’s surprising is that Intel wants to launch another pair of Xeon products, Emerald Rapids (in Q4 2023) and Granite Rapids, shortly after Sierra Forest.
The other half-surprise is that Intel is embracing Performance/Efficiency cores at the Data Center level, which makes sense because that’s what the market demands. We know that AMD will launch a 128-core Bergamo CPU based on the Zen 4c architecture later this year.
Both P- and E-cores will not live on the same silicon (like on consumer CPUs) as they target different audiences. They also come with their very own challenges; both have to contend with a fixed power and transistor budget.
Fewer Power cores with all bells and whistles (e.g. AVX-512), clocked at higher speeds will generate/consume as much power as more Efficiency cores that have fewer bells and whistles (e.g. single threaded) with lower clock speed.
Then there’s cache (the memory that’s nearest to the cores), latency (moving data to and from various parts of the CPU), the chip yield (how many usable silicon chips are produced), available memory bandwidth (to the main system memory) and so on and so forth.
Does it matter to anyone else?
Rumors of a 120-core Xeon CPU were finally put to bed yesterday but while this race produces entertaining headlines, it won’t matter much to anyone but hyperscalers. Consumers won’t be able to buy these chips anytime soon (ed: well, they might be able to bid for them on Ebay or buy them from Aliexpress once they’re decommissioned though, in a few years) as they will require specialized server systems (including a new breed of DDR memory modules) that would make no sense to end users and will cost hundreds of thousands of dollars per piece.
The high-core count, low-power strategy that both Intel and AMD have adopted will help them contain the Arm threat for now. There’s at least a half-a dozen rivals that either want a bigger part of the pie (Huawei, Nvidia, Ampere) or reduce their dependence on Intel and AMD (Amazon, Tencent, Google).
AWS’s Graviton has 64 cores but we’d expect that to rise depending on market demand. Ampere already has 128 cores and Arm’s Neoverse Poseidon platform which was announced in September 2020, aims to offer higher core density (higher than N1’s 128-core limit) when it rolls out.
Higher core count matters for applications such as web hosting (Virtual Private Servers or VPS), virtual desktop (VDI) and a bunch of other applications that do not consume a lot of power and don’t mind living in a multi-tenancy environment.
