In 2024, Microsoft will use an in-house successor to its Ampere Altra Arm Neoverse N1 chips. The Microsoft Azure Cobalt 100 is a 128-core Arm Neoverse N2 design that is using Arm’s accelerated path to chip design to create a lower-power chip.
Microsoft Azure Cobalt 100 128 Core Arm Neoverse N2 CPU Launched
A few weeks ago we covered the Arm CSS N2 announcement. The basic idea is that Arm is providing not just the Core and fabric IP, but also ready-built IP blocks of many cores. The key building block Arm showed during that announcement was a 64-core IP block arranged in dual core compute tiles.
It then showed the capability to link multiple 64-core CSS N2 blocks together using chiplet interconnects.
Arm also showed a 128-core theoretical device in each socket of a dual-socket server.
Its proof point Arm also discussed that it helped a hyper-scale customer go from project start to silicon in 13 months. Our best guess is that this is Microsoft where several ex-Arm folks now work.
Deployment is apparently happening today, but it seems like GA is going to be in 2024. It is quite interesting that Microsoft is using the 2021-era Arm Neoverse N2 for this. We first started seeing products like the Marvell Octeon 10 Arm Neoverse N2 DPU that rivaled 2017-era Intel Xeons. A year later we get a Microsoft announcement, and a year after that we should get GA of parts.
The Microsoft Cobalt 100 is really interesting. From what we have seen from the Arm Neoverse N2, it is roughly like a Skylake-era Xeon in terms of Integer performance at a lower power point and a significantly higher density. Microsoft will also get new features like DDR5 memory. From a performance perspective, AMD’s 2023 Bergamo will be significantly faster than Microsoft’s 2024 Cobalt 100, but our big question has been whether Bergamo is too fast for cloud providers looking for low-cost and lower-power cores.
It is also interesting that Microsoft is going with a custom N2 design since it also has AmpereOne using custom-designed Arm cores. That begs the question is Microsoft simply cutting out the middle-man or instead planning to offer Arm at multiple performance and capability points? An Arm Neoverse N2 design in 2024 is not even going to be the highest performance per core cloud Arm CPU when it is available, and maybe not even in the top half.
For a few years now at STH, we have been talking about just how big the cloud-native processor will be. Microsoft apparently thinks this segment will be big enough that it warrants building its own CPUs.