EPYC at a glance
Advanced Micro Devices (AMD) is entering the high-end server market with its new EPYC processor. The processor has a balanced design that offers flexibility and power efficiency. This enables the server to operate at a high frequency even in a power-constrained environment.
The EPYC processor features 32 cores, eight memory channels, and 128 lanes of high-bandwidth I/O (input/output) that can transfer data from the network, drives, and flash storage. However, putting so much memory, I/O, and cores on a single chip would increase the cost of the chip and result in low yields.
AMD addressed these limitations by using Infinity Fabric. EPYC has four dies packed in one integrated unit. Each of these dies has eight Zen cores, and these dies are connected through Infinity Fabric, which allows the cores to scale efficiently.
Infinity Fabric is a scalable control fabric that manages the sensors, voltage, and frequency of every chip, which delivers maximum performance while consuming the lowest amount of power. Infinity Fabric is scalable within a die, across multiple dies, and across multiple sockets. As a result, Infinity Fabric allows AMD to scale while reducing costs.
AMD plans to launch two versions of EPYC—a two-socket platform and a one-socket platform.
Leveraging Infinity Fabric in two-socket platform
AMD is scaling its server processor by connecting two EPYC systems with Infinity Fabric in a two-socket platform. The two-socket platform features 64 cores, 16 memory channels, and 128 lanes of high-bandwidth I/O.
When compared to Intel’s (INTC) highest-end server processor, AMD’s two-socket EPYC platform has 45% more cores, 122% more memory bandwidth, and 60% more I/O. This new scaling approach gives AMD a leadership position in the two-socket server market.
AMD’s single-socket platform
Most server systems depend on NVIDIA’s (NVDA) GPUs (graphics processing units) and other accelerators for their machine learning applications. These systems do not need the CPU power offered by a two-socket server system, but users must still purchase that server system to attain the level of I/O memory bandwidth needed to support GPUs.
If PCIe (Peripheral Component Interconnect Express) switches are used to connect GPUs, this configuration would increase latency and impact the GPUs’ performance.
AMD is offering a one-socket platform with all the memory bandwidth offered by Intel’s dual-socket platform while delivering improved TCO (total cost of ownership).
EPYC’s data center applications
The flexible structure of EPYC processors makes them suitable for a large number of data center workloads. EPYC’s high memory bandwidth and memory capacity enable it to process data faster, making it well-suited for high-performance computing, virtualized cloud computing, machine learning, Big Data, and analytics.
EPYC’s high I/O bandwidth and denser cores speed up the migration of data from proprietary storage hardware onto software, making it well-suited for software-defined storage.
Next, we’ll compare EPYC’s performance against Intel’s server offerings.