It’s been a while since we last heard about Vega, AMD’s new GPU architecture that builds on the success of their Graphics Core Next (GCN) engine. It’s one of AMD’s better-kept secrets, and details have been very scarce. However, it looks like AMD will be making an appearance at the 2017 Consumer Electronics Show (CES), and they’re bringing Vega along with them. A promotional website popped up this past weekend with the URL of http://ve.ga.
The early stages of the site’s appearance lacked a countdown timer, which is now online, as well as any kind of security controls, so eager fans could easily gain access to the hidden content that had yet to be published. A word cloud of all the important bits about Vega was one of those things, and it leaves almost nothing to the imagination, but I’m still excited!
A couple of the statistics in the cloud allude to the use of second-generation High Bandwidth Memory (HBM), a replacement for GDDR5 and GDDR5X in graphics cards. HBM was first shipped in AMD’s Radeon Fury and R9 Nano family, but it did not see much mainstream support because GDDR5 was significantly cheaper. For the Vega family, AMD is still using HBM in the high end, but it is now in its second generation. There’s twice the amount of bandwidth available, and up to eight times more memory via stacked memory dies. The Radeon Fury X topped out at 4GB with 512GB/s of bandwidth, so AMD is promising up to 1TB/s of bandwidth and up to 32GB of HBM memory. Good golly!
The virtual address space line is a nod to their Radeon Pro RX graphics cards launched in late 2016 that had a solid state drive strapped into the circuit board, used for caching large amounts of data. This is possibly a better version of that technology, and it might mean that AMD would allow very large storage arrays to act as a memory pool for professional GPUs based on Vega. Keep in mind, also, that the 4x power efficiency claim might be for the HBM chips themselves, rather than the GPU, given its strategic placement next to the higher bandwidth claim. VRAM is a sucker for power draw, and reducing it is always a good thing.
The rest of the cloud is dedicated to some specifics about the architecture. There’s a claimed doubling in peak throughput per clock with the Vega NCU (we don’t know what that stands for yet), and it counts as yet another new version of GCN. The new rasteriser might be mimicking NVIDIA’s improvements in their Maxwell and Pascal architectures which change how the GPU colours images (according to findings by Real World Tech’s David Kanter, which you can read about here). Being able to selectively cull parts of the image that don’t need to be drawn is a trick copied from mobile graphics chips found in phones and tablets, and it’s a great way to save on power.
AMD is expected to detail Vega for the first time publicly during their CES presentation. Stay tuned for more news about new tech coming out of CES 2017!