Both the newly announced Radeon Instinct MI60 and the Radeon Instinct MI50 share a lot of similarities, with both being based on the same 7nm Vega GPU, and bring significant improvements compared to the previously available Radeon Instinct MI25.
When it comes to specifications, the 7nm Vega GPU packs 3.2B transistors on a 331 mm2 die, with flexible, high-performance compute engine capable of covering different workloads, scalable interconnect and communications and end-to-end ECC enablement. It also comes with up to 32GB of 2nd generation High Bandwidth Memory (HBM2) on a 4096-bit memory interface, offering 1TB/s of memory bandwidth.
The 7nm Vega is also the industry-first PCIe 4.0 capable GPU, offering 64GB/s of bi-directional CPU-to-GPU bandwidth, and comes with new Infinity Fabric Links, offering 100GB/s per Link for GPU-to-GPU bandwidth.
In terms of general performance, AMD says that the 7nm Vega GPU offers up to 2x more density, 1.25x higher performance at the same power, and 50 percent lower power at the same frequency, offering higher performance per Watt compared to the MI25.
AMD's CEO, Dr. Lisa Su, and Forrest Norrod, SVP and General Manager for Datacenter and Embedded Solutions at AMD, showed a benchmark comparing the new AMD EPYC Rome with Radeon Instinct MI60 with previous generation EPYC Naples and Radeon Instinct MI25 GPU, showing significant improvements in Resnet-50.
The Radeon Instinct MI60 packs 64 Compute Units, for a total of 4096 Stream Processors and offer a peak compute performance of up to 7.4 TFLOPS for FP64, up to 14.7 TFLOPS for FP32, up to 29.5 TFLOPS for FP16, up to 59.0 TOPS for INT8 and up to 118 TOPS for INT4. The Radeon Instinct MI60 comes with 32GB of HMB2, and has both RAS support, MxGPU (HW SR-IOV) support for secure virtualized workload, and has a 300W TDP.
The Radeon Instinct MI50, comes with a cut-down GPU with 60 Compute Units, leaving it with a total of 3840 Stream Processors and only slightly lower compute performance, but by packing 16GB of HBM2, is able to target a lower price point for those that still need compute performance but not that much memory.
The big part of the new Vega 7nm Radeon Instinct announcement was the announcement of a new version of the ROCm open software platform, which will bring both support for new Radeon Instinct accelerators as well as updated libraries, support for 64-bit Linux operating systems, a bunch of optimizations, and support for latest versions of deep learning frameworks like TensorFlow 1.11, PyTorch (Caffe2) and others.
According to details provided at the event, AMD plans to ship its Radeon Instinct MI60 accelerator by the end of this year, while the Radeon Instinct MI50 is coming in Q1 2019. AMD also mentioned the "MI-Next" GPU accelerator, stating that it is on track.