The new, 7nm Vega-based AMD Radeon Instinct line-up is released
AMD Radeon Instinct GPUs are extremely high-powered accelerators for applications such as machine learning. Their most recent generation has graduated to 7nm status, with Vega 20 architecture as anticipated. It pushes these cards (the Instinct MI50 and MI60) into the teraflops in terms of advanced peak precision.
The MI50 and MI60 also have a maximum bandwidth of 1024 gigabits per second (Gbps), whereas the Instinct series had only achieved 512 Gbps at most (via the MI8 card) in the past. The MI60 has 4096 stream processors and 64 computational units, which is not unusual for pre-existing Instinct accelerators. However, the new models' performance is unknown territory for this series.
The MI60 has a peak half precision (or FP16) value equivalent to 29.5 teraflops (TFLOPs), whereas the best the Instinct line-up could do prior to this was 24.6 TFLOPs (in the MI25). Furthermore, the double-peak precision (FP64) of the MI60 goes up to 7.4 TFLOPs, whereas the MI25 managed 768 gigaflops. The MI60 has 32 GB of HBM2 memory, whereas the older models (and the MI50) have 16.
On the other hand, we expected that the new Radeon Instincts would make it to a TDP of 400W. This has not entirely been realized, although the MI50 and 60 are associated with a respectable 300W. Finally, the new Instincts are capable of PCIe 4.0 (although this may require pairing with the EPYC Rome CPU lineup). All in all, it is clear that the new Radeon Instincts are shaping up to be super-powered GPUs for next-level deep learning.