Again on the announcement of RDNA, AMD made it clear that it was at a fork within the highway for its sole graphics structure on the time, GCN. On the one hand, it had gaming necessities to fulfill. On the opposite, datacentre fits demanding massive quantity crunching. To placate each events, Radeon created two totally different architectures: RDNA and CDNA. What we’re seeing right this moment is the primary graphics card to make use of the latter: the Intuition MI100.
The MI100 is a critical quantity crunching GPU, and is meant to be positioned amongst the mess of cables present in any good datacentre. Or lack of mess on the actually good ones. The cardboard itself affords no graphics output—or any fixed-function graphics blocks in any way—which means you could not join this card as much as your monitor for slightly again of the warehouse gaming should you wished to. Sorry.
It is a disgrace, too, as a result of the MI100 homes 120 Compute Models. For comparability (tough comparability, thoughts) the so-called ‘Large Navi’ GPU discovered throughout the RX 6900 XT comes with 80 CUs. They’re utterly totally different architectures, in spite of everything, however that does not make the MI100 any much less of a GPU monster.
That chip is designed to speed up HPC and AI workloads like no different, and AMD says it is top-of-the-line round. By its personal numbers, it places the MI100 forward of Nvidia’s A100 by a big quantity because of a brand new ‘Matrix Core’ (HPC will get all of the cool named stuff) that accelerates sure workloads.
That Matrix acceleration helps take the playing cards customary 23.1 TFLOPS of FP32 efficiency to 46.1 TFLOPs when utilizing MFMA directions, a brand new household of wavefront-level instruction from AMD. As such, the efficiency uplift will not be speedy for all workloads.
The enjoyable does not cease there for some fortunate datacentre engineer, although. The MI100 comes with 32GB of HBM2 reminiscence (for a whopping 1.23TB/s reminiscence bandwidth), PCIe 4.0 help, and all at a 300W TDP—the identical because the upcoming RX 6800 XT and RX 6900 XT.
One motive for that easy-going TDP, no less than for a chip of this measurement, is the truth that a lot of the graphics-specific silicon has been ripped from the chip to make method for extra number-crunching package. Waste not, need not.
There’s not lengthy to attend earlier than we get to blow the lid off the second-generation RDNA 2 graphics cards (solely two extra days!), however within the meantime the datacentre world is getting its personal style of AMD’s red-tinted model of the great life.
Oh and Nvidia additionally occurred to launch an 80GB A100 right this moment, too. That is not as a result of some intelligent tactic by both aspect: All of those bulletins have been made at SC20, or Supercomputing 2020, a HPC convention going down right this moment.