Unleash discovery on the world's first 7nm accelerator.
The jump to 7nm allowed AMD to bump up the transistor count for Vega 20, growing from 12.5 billion to 13.2 billion. Although the former isn't of much interest to us since the announced-so-far Vega graphics chips won't be made available to consumers, the CPU news is much more pressing as it furthers rumors that AMD's next-generation architecture will make it not only competitive with Intel on cost, but on overall raw power too.
Beyond the die shrink, the newest Radeon Instinct cards come with a few other upgrades that should benefit professionals, such as PCIe 4.0 support (another industry first), and two additional HBM2 memory controllers for a total of four. The extra HBM2 stacks and memory bandwidth likely won't fix AMD's performance deficit, where the GTX 1080 Ti is about 40 percent faster than Vega 64, never mind the RTX 2080 Ti.
AMD has also introduced the lower-end Radeon Instinct MI50 accelerator that still comes with the same 7nm Vega 20 GPU but features lower amount of Compute Units and stream processors at 60 and 3840 respectively.
The Radeon Instinct MI60 comes rocking 7.4 teraflops of compute power when handling 64-bit floating point data.
How Well Does the MI60 Perform?
In terms maximum theoretical compute power, AMDclaims up to 118 TFLOPS in INT4.
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.