At its Next Horizon event, alongside the Zen 2 architecture preview, AMD took the time to share a few details about its 7nm accelerator cards aimed at high performance computing workloads.
Based on the Vega20 architecture, the Radeon Instinct MI50 and MI60 GPUs are in essence AMD's equivalent for Nvidia's Tesla lineup. Geared towards deep learning and AI, these two deliver up to 6.7 TFLOPS and 7.4 TFLOPS respectively of peak dual precision performance. This is a massive increase in comparison to the previous top-of-the-line Instinct GPU - the MI25 - with its peak dual precision performance of 768 GFLOPS.
Despite keeping the same 300W TDP, Vega20's shrinking of the die has brought with it some improvements too, like an increased bandwidth of up to 1TB/s compared to the MI25's 484GB/s. A few of the specification highlights are present in the table below:
|Radeon Instinct MI50||Radeon Instinct MI60|
|Full Chip ECC
|Bandwidth||Up to 1TB/s||Up to 1TB/s|
|Bus interface||PCIe Gen 3, PCIe Gen 4||PCIe Gen 3, PCIe Gen 4|
|Infinity Fabric||2 links, 100GB/s peak bandwidth||2 links, 100GB/s peak bandwidth|
Peak half precision
|Up to 26.8 TFLOPS||Up to 29.5 TFLOPS|
|Peak single precision
|Up to 13.4 TFLOPS||Up to 14.7 TFLOPS|
|Peak dual precision
|Up to 6.7 TFLOPS||Up to 7.4 TFLOPS|
|Card type||Double slot, full height||Double slot, full height|