New Arm server processor with 136 cores surpasses Intel and AMD

Arm is taking on Intel and AMD's server processors with the Arm Neoverse AGI. Designed for data centers and demanding AI tasks, the device will begin commercial deliveries in the coming months. With this model, Arm has taken a major step forward in the server market, which has been dominated by x86 processors.

The new Arm Neoverse AGI processor uses a two-die design, manufactured by TSMC on its advanced N3P manufacturing process. Customers will be able to choose from 64, 128 or 136 Arm Neoverse V3 cores, operating at frequencies between 2.8 and 3.7 GHz. Interestingly, each individual die physically contains 70 cores. In the most powerful 136-core version, four cores are disabled, which helps the manufacturer improve manufacturing efficiency. Each core offers two 128-bit vector engines and 2 MB of level two (L2) cache, while the entire processor offers up to 272 MB of system cache.

The two components are connected to each other via a 32 GT/s UCIe interface, enabling throughput of up to 2 TB/s. Unlike some competitors, each component has its own processor cores, memory controllers, and I/O units. This means that access to local memory does not have to pass between the components, which reduces latency. According to the manufacturer, the Arm Neoverse AGI processor achieves a local DRAM access latency of less than 100 ns.

The memory subsystem is one of the main features of this innovation. Each module contains a six-channel DDR5 memory controller, which together provides up to 844.8 GB/s of memory bandwidth. The chip supports DDR5-8800 memory and up to 6 TB of memory per socket. Special mechanisms are also built in to distribute memory traffic priorities between processor cores and other devices.

For add-in card connectivity and expansion, the Arm Neoverse AGI processor offers 96 PCIe 6.0 lanes with CXL 3.0 support, plus four PCIe 4.0 lanes, I3C, I2C, and SPI interfaces. The total thermal design power (TDP) is 300 W.

Arm is targeting the Arm Neoverse AGI processor primarily at AI servers, where memory bandwidth and low latency are key to high performance. Arm claims that servers with these processors could offer up to twice the performance per rack compared to modern x86 processors. However, these are internal estimates from the manufacturer, so it's worth keeping a healthy dose of skepticism about these numbers until independent tests are available.