AMD shows Vega with 32 GB HBM2 memory and 7-nanometer technology – released late 2018

In August 2017, AMD released the then fresh performance circuit Vega 10 for consumers in the form of Radeon RX Vega 64 and RX Vega 56. Like Polaris, the architecture then remained at the manufacturing technology 14 nanometers, but it was not long before rumors of a shrinkage to 7 nanometers began to flourish.

This was later confirmed by official product plans and further details will be revealed during this year’s Computex. The investment revolves around machine learning and artificial intelligence, where the Vega architecture has been adapted for higher computing power at loads typical of deep learning.

In addition to this, the shrinkage to 7 nanometers is the main number, where this enables higher transistor density, improved energy efficiency and clock frequencies. According to AMD, this will be a doubling of the first two parameters and in terms of performance, the 7-nanometer technology will be 35 percent better.

AMD already uses Infinity Fabric for internal communication in Vega 10 and at the conference AMD suggests that the company may now use it for communication between several Vega circuits on one and the same substrate. Specifically, this would mean that AMD can manufacture several smaller circuits and then in practice use them as a large one – similar to how AMD Epyc or Ryzen Threadripper are structured.

In terms of specifications, few details are revealed, where neither circuit area, clock frequencies nor number of calculation units were revealed during the conference. However, AMD confirms that the 7-nanometer Vega comes with 32 GB of HBM2 memory spread over four capsules, which points to a 4,096-bit wide memory bus.

► Read all articles from #computex-2018

Finally, AMD is moving the launch of Vega to 7 nanometers, where it was previously planned to go into production in late 2018, for a launch in early 2019. Instead, test production will begin at Globalfoundries now and AMD aims to release during the second half of the year.

Source: Anandtech