The AMD Radeon Vega Frontier Edition is the first Vega GPU, but it's targeted at AI researchers
Updated on 29 June 2017: The Radeon Vega Frontier Edition is now available to buy at US$999, but local customers will need to pre-order it first. The liquid-cooled version will retail later in Q3 for US$1,499.
Originally published on 17 May 2017:
AMD’s first Vega GPU is a workstation card. At its 2017 Financial Analyst Day in Silicon Valley, the company announced the Radeon Vega Frontier Edition, which can be seen as a quick riposte to NVIDIA’s just-announced Tesla V100.
Just like Ryzen in the CPU space, Vega is supposed to be AMD’s resurgence in the high-performance GPU market. However, Vega will also be AMD’s way to make inroads in the data center and enterprise arenas, what with its ability to offer access to terabytes of memory through a scalable memory architecture that lets the GPU utilize HBM2, on-card NAND, and even sources like RAM and network storage.
That said, the Frontier Edition card was designed with machine learning researchers, scientists, and creative professionals in mind, and its specifications are actually closer to NVIDIA’s Titan Xp rather than the newer Tesla V100.
It offers around 12.5 TFLOPS of peak single-precision compute performance, just edging out the Titan Xp’s 12 TFLOPS.
In addition, it boasts 64 next-generation compute units (NCUs) that probably represent the most significant change to AMD’s shader cores since the company rolled out its Graphics Core Next (GCN) architecture. The new NCUs support new data types such as packed math formats, and Vega should prove a lot more adept in applications such as neural network inference.
The Radeon Vega Frontier Edition will also feature a total of 4,096 stream processors and 16GB of High Bandwidth Cache (HBC), AMD’s new-fangled term for HBM2, for memory bandwidth of approximately 480GB/s. It can access up to 256TB of virtual memory as well, setting it up for use in things like deep learning algorithms, intensive video processing, and real-time visualizations.
AMD intends for the Vega card to be paired with its open-source Radeon Open Compute (ROCm) machine learning platform, which will reportedly help developers better utilize Vega’s capabilities for machine learning algorithm development.
This isn’t AMD’s first foray into AI – it introduced the Polaris-based Radeon Instinct accelerators last year for deep learning inference and training, and a Vega-based version is also expected – but the Frontier Edition GPU should help the company shift the conversation away from NVIDIA and put it on the map.
And even though AMD recently announced a refreshed Radeon Pro Duo based on the Polaris architecture, it said we can expect a Vega-based Radeon Pro GPU soon as well.
A liquid-cooled version of the card with a hand-made golden shroud will also be available in more limited quantities.