Amd insight mi250x accelerator card has entered the global top 500 supercomputing companies, and has become the first "frontier" and the first public supercomputing machine with 10 billion times. Intel has also cooperated with the U.S. Department of energy to create a 10 billion times supercomputing "aurura" by using its first acceleration card Ponte Vecchio. It is said that the maximum performance can exceed 20 billion times. It will soon be launched. The next list in November should be available.
Now, intel can't wait to announce the next generation of accelerator card, code named "Rialto Bridge" (Rialto Bridge, Venice, Italy)
Rialto bridge will adopt Intel IDM 2.0 manufacturing mode, with more advanced technology (Intel 4?), It integrates up to 160 Xe cores, 32 more than the current Ponte Vecchio, that is, an increase of 25%, bringing higher floating-point performance and IO bandwidth. The official says that the actual application performance can be improved by 30%**
It seems that the performance improvement mainly comes from more computing units.
Schematic diagram of Rialto Bridge
The Rialto bridge will also adopt the new generation shape specification OAM 2.0 (open accelerator module), which has not been published yet, compared with the current OAM 1 x. The maximum power consumption increased from 700W to 800W
Of course, the actual product power consumption does not necessarily have to reach 800W, but there is this space.
While changing the form, Rialto bridge is downward compatible on the platform and can still form a node with four cards in parallel.
*Rialto Bridge plans to complete the sample * in 2023, but considering the development cycle of Ponte Vecchio, it is estimated that it will be at least 2-3 years before it is put into commercial use.