AMD's Arcturus CDNA data center GPU features 7,680 cores.

AMD's Radeon Instinct Arcturus GPU, featuring the CDNA architecture, has been spotted by Rogame and is poised to target the data center server market. The compute-optimized GPU design is intended to deliver the highest performance compute capabilities for data centers. AMD's next-generation HPC and AI GPUs have been designed separately from consumer-end chips, with the Arcturus GPU exclusive to the datacenter market. AMD has confirmed in its Radeon CDNA architecture roadmap that all CDNA-based GPUs will be exclusively designed for the HPC and data center markets, while Radeon RDNA GPUs will power the consumer segment.

AMD's Arcturus GPU was initially leaked in 2018, before the company introduced any 7nm GPUs. The Radeon VII and Navi lineups launched in 2019 and featured 7nm GPUs, with Navi aimed at the mass consumer market. The new architecture represents a strategic shift to develop separate product lines for enterprise and consumer applications, allowing AMD to optimize each segment independently.

The Arcturus GPU specifications include increased cache and double the compute units compared to Vega. Data center-specific features such as XDLOPs, Rapid Packed Math, New Vector ALU, and BFloat16 support are expected in the new Radeon Instinct cards. The previous Radeon Instinct MI100 prototype board featured the Arcturus-XL die with a rated TDP of 200W and 32 GB HBM2 VRAM clocked at around 1000–1200 MHz. A spotted test board based on the Arcturus CDNA GPU offers 120 compute units for a total of 7,680 stream processors and a GPU clock speed of 878 MHz with a 750 MHz SOC clock. This variant also features HBM2 memory clocked at 1200 MHz, which with a 4096-bit bus would deliver around 1.2 TB/s bandwidth, comparable to Aquabolt's performance. Both AMD and NVIDIA are likely to utilize the faster HBM2E 'Flashbolt' standard, which can deliver up to 1.8 TB/s bandwidth.

The 878 MHz clock speeds on the test board are relatively slow compared to past variants reaching 1334 MHz, yielding around 13.5 TFLOPs of FP32 compute power—lower than the Radeon Instinct MI60 and the 21 TFLOPs achieved on a previous prototype. The first iteration of CDNA GPUs is expected to reach approximately 25 TFLOPs FP32. Based on leaks of NVIDIA's Ampere GPUs, expected to be announced later this year, NVIDIA may hold the upper hand in compute performance with speculated performance reaching almost 36 TFLOPs of FP32 and 18 TFLOPs of FP64 compute power from their next-generation Tesla 7nm GPU lineup.

AMD's stated focus for CDNA is performance, efficiency, features, and scalability in the data center market, replacing the legacy GCN architecture with GPUs specifically optimized for high-performance compute, machine learning, and HPC applications. The first-generation CDNA GPUs will feature a 2nd Gen Infinity Architecture and utilize the ROCm (Radeon Open Compute Platform) to power data centers with key optimizations and enhanced scalability. The 2nd Gen Infinity Architecture will enable 4–8-way GPU connectivity in a single node, allowing the new Radeon Instinct boards to operate in harmony. AMD has demonstrated the ability to offer competitive FLOPs at competitive pricing, which may be the strategy behind Arcturus. While timing remains unconfirmed, AMD has hinted at a Radeon Instinct product release later this year featuring its first-generation CDNA architecture.