AMD and Cray this week that by 2021 they will launch the most productive supercomputer system in the world, called Frontier. It is quite expected that the customer was the US Department of Energy, although AMD executive director Lisa Su in comments to the resource listed quite peaceful tasks that this supercomputer will have to solve: biological research, genome deciphering, weather prediction and the search for new energy sources.
AMD representatives gave very interesting comments to the site staff , from which it becomes more or less clear which components AMD prepared for Crayβs order. As previously reported, specifically for this project AMD developed not only EPYC central processors, but also Radeon Instinct computing accelerators based on GPUs with HBM memory (generation not specified).
The mystery of the central processors of the new supercomputer
AMD Vice President Forrest Norrod did not explain which processors will form the basis of the Frontier supercomputer, but made it clear which processors will not be used in it. From his words, it is known that these processors will not use either the Zen 2 architecture of the Rome processors being prepared for announcement in the third quarter, or the next-generation architecture inherent in the Milan processors, which should be released in 2020. Frontier's EPYC processors will be custom-tailored. True, Lisa Su could not resist the temptation to explain that the processors for this supercomputer will be based on the architecture that will replace Zen 2. It can be assumed that they will receive a modified Zen 3 architecture. Such processors should be produced using second-generation 7nm technology, with elements of so-called ultra-hard ultraviolet (EUV) lithography.
In this context, by the way, it becomes clear who the head of AMD had in mind at a recent reporting conference, mentioning the emergence in 2020 of a new client in the direction of βcustomβ components, which has nothing to do with the gaming console segment. It can be assumed that this client could well be Cray, because the supply of processors will have to be established before the supercomputer is launched in 2021.
Forrest Norrod allowed himself to joke that if the name of the specialized EPYC processors for the Frontier project is revealed, it will remind everyone of another Italian city. The company names server processors with Zen family architectures in honor of various Italian cities: Naples, Rome or Milan.
Graphics component Frontier also hides its architectural affiliation
In the case of Radeon Instinct computing accelerators, AMD will also have to adapt to the needs of Cray. The Next Platform website reports that these components for Frontier will not use the Vega or Navi architectures, but will be custom built. Special instruction sets will allow GPUs to faster process tasks typical for server configurations and artificial intelligence systems.
Special attention is also paid to the efficiency of data transfer between central and graphics processors in this supercomputer system. AMD has improved its high-speed Infinity Fabric interface. There will be up to four graphics processors per central processor.
Representatives of the Oak Ridge Laboratory, which will operate the Frontier supercomputer, in the most correct terms made it clear to colleagues from The Next Platform website that the cost of purchasing computational accelerators with HBM memory has so far eaten up the bulk of the budget for the construction of supercomputer systems. Until recently, AMD promoted GPUs with HBM memory primarily in the graphics acceleration segment, but recently it has been actively promoting them for computing acceleration needs. In the first quarter, it was the positive dynamics of deliveries of such accelerators that helped AMD raise its profit margin and the average selling price of its products.
In the supercomputer segment, NVIDIA Tesla computing accelerators encountered almost no competitive resistance, and this circumstance did not have the best effect on the pricing policy of this company. Now that AMD has strong support from supercomputer manufacturers, prices may move closer to fairer levels, although HBM memory continues to be expensive.
Source: 3dnews.ru