Taking notes from AWS, Google prepares custom Arm server chips of its own
Aim of the game is to match Intel and AMD on performance
Google is understood to be developing its own custom Arm server processors, following in the footsteps of cloud rival AWS.
The search giant already supports virtual machines powered by Ampere Computing's Arm-based Altra processors, but in the wake of Amazon's in-house designed Graviton silicon, Google is said to be working on a pair of its own Arm-based server CPUs.
According to The Information, Google is preparing one processor codenamed Maple which is based on an existing design from Marvell Technology, and a second codenamed Cypress that is said to be a custom design being developed by a team based in Israel.
Both are said to be designed for 5nm production processes, and as Marvell is a fabless semiconductor company that uses Taiwan-based TSMC to produce its chips, it is possible that this is what Google also has planned.
Mountain View's Google Cloud is aiming for performance from its new Arm server chips that will be comparable with that of server processors from Intel and AMD, according to the report.
But how do you measure this?
It isn't clear whether this means each individual core will be a match for an x86 core, which might be a tall order, or that the entire chip could best the performance of an Intel or AMD part by cramming in more cores.
This was the approach taken by Ampere, which fits 128 cores onto its Altra Max processor, each one focused on maximum single-thread performance to meet the requirements of cloud-native workloads.
However they shape up, producing its own tailor-made chips could be a more cost-effective strategy for Google in the long run than buying in x86 chips, which can cost thousands of dollars apiece, especially at the kind of scale cloud companies operate.
According to the report, Google is looking to have its custom server chips in production sometime in the second half of 2024, with actual deployment in its datacenters by 2025.
- Uber strikes deals with Google and Oracle to cut datacenter dependence
- 5% of the cloud now runs on Arm as chip designer plans 2023 IPO
- Arm still strong despite SoftBank loss as shipments pass a quarter of a trillion
- Fortinet's latest ASIC promises 2.5Gbps of SSL inspection at the edge
Meanwhile, AWS has just introduced more instances powered by its own Graviton3 Arm-based processors. The M7g instances are aimed at general-purpose workloads, while the R7g are memory-optimized, making them a better fit for workloads such as databases, in-memory caches, and data analytics.
These 7th generation instances were designed to deliver up to 25 percent greater performance than the equivalent 6th gen instances making them "the best performers" in its EC2 service, AWS claimed.
The M7g instances can be specified with up to 64 vCPUs and from 4GB to 256GB of memory, while the R7g instances can also have up to 64 vCPUs but 8GB to 512GB of memory. These are currently available in the AWS US East, US West and Europe (Ireland) regions.
Microsoft also has workloads running on Arm servers on Azure. It announced instances last year based on Ampere Altra processors, claiming that these would be able to deliver up to 50 percent better price-performance than comparable instances operating on x86 hardware.
Last week, it was claimed by Arm's parent company SoftBank that as much as 5 percent of the cloud services operated by AWS, Microsoft Azure and Google Cloud are already delivered by Arm chips.
According to a prediction by Steve Brazier, CEO at channel analyst Canalys, by 2026 some 50 pecent fo CPUs sold to the public clouds will be Arm-based, not x86-based. "That is a massive industry shift".
Indeed it is. If correct. ®