Vega is AMD’s latest attempt to battle Intel and Nvidia for data centers and gaming PCs.
GPUs aren’t just for graphics anymore: They’re crunching through artificial intelligence algorithms in Facebook’s data centers and powering self-driving cars from dozens of automakers and tech companies.
Nvidia and Intel have fueled much of this GPU revolution, but AMD is now throwing its hat into the ring with today’s unveiling of the Vega GPU architecture. Vega is AMD’s equivalent of a Swiss Army knife: a next-generation chipset equally at home processing petabytes of data for machine learning and powering the latest virtual reality headsets and 4K video games.
Vega’s chief selling point is its scalability: AMD claims it is the most scalable GPU memory architecture on the market. Instead of the GDDR5 memory used in most graphics cards, Vega will use the more energy-efficient HBM2. Each module has twice the bandwidth and is less than half the size of its GDDR5 equivalent, which could make Vega feasible for use in more (and smaller) applications than earlier architectures.
Other Vega improvements include a high-bandwidth cache controller capable of allowing up to 512TB of virtual address space, a new programmable geometry pipeline to support more realistic gaming graphics, and an updated compute unit that can handle 128 32-bit operations per clock cycle, several times the rate of a conventional GPU compute unit.
AMD says Vega will help its customers “make decisions based on exabytes of data in an instant.” That indicates the company is positioning the GPU for data center use as much as use in high-end gaming PCs. The company didn’t release details about the pricing and availability of Vega units, although it’s likely that consumer cards will follow the traditional GPU model of manufacture and sale by third parties like PNY, Sapphire, and EVGA.