Thanks to its 7nm manufacturing process, the Radeon Instinct MI60 features a 2x uplift in density for a total of 13.2 billion transistors (with a die size of 331mm2). At the same power levels, you'll see a 1.25x performance boost over current Vega-based Instinct cards while power consumption has been reduced by 50 percent at the same clock frequencies.
The Radeon Instinct MI60 is built on an "enhanced" Vega architecture, includes 64 compute units, 4096 stream processors, 32GB of HBM2 (with ECC), and 1TB/sec of memory bandwidth. There will also be a slightly less potent Instinct MI50 with 60 CUs, 3840 SPs, and 16GB ECC HMB2. These will be the first PCIe 4.0-capable GPUs, doubling bandwidth over PCIe 3.0 solutions. Next-generation Infinity Fabric Link technology allows for peer-to-peer GPU communication at speeds of up to 200 GB/s.
When it comes to performance, AMD says that Radeon Instinct MI60 offers 14.7 TFLOPS FP32 and 7.4 TFLOPS FP64 compute. The Instinct MI50 pares those numbers down to 13.4 TFLOPS and 6.7 TFLOPS respectively.
These Radeon Instinct cards are strictly for HPC and deep learning applications -- in fact, they don't even have display connectors so you're out of luck for gaming. But for its intended audience, these new 7nm GPUs are looking to score a bullseye.
“Legacy GPU architectures limit IT managers from effectively addressing the constantly evolving demands of processing and analyzing huge datasets for modern cloud datacenter workloads,” said Wang. “Combining world-class performance and a flexible architecture with a robust software platform and the industry’s leading-edge ROCm open software ecosystem, the new AMD Radeon Instinct accelerators provide the critical components needed to solve the most difficult cloud computing challenges today and into the future.”
Datacenter customers can expect to see Radeon Instinct MI60 cards by the end of 2018. However, Radeon Instinct MI50 won't be shipping until towards the end of Q1 2019.