Meant to house four Vega GPUs
It appears that AMD showed quite a few interesting products at its earlier event, including what is now refered to as the Vega Cube, a Radeon Instinct GPGPU solution.
The device is actually comprised of four VEGA prototype boards that are somehow interconnected and should provide 100 TFLOPs of half-precision compute power, at least according to early reports like the one coming from Hardwareluxx.de.
While there are still no VEGA GPUs on it since the GPU is obviously in its early prototype stages, such a device should outperform Nvidia’s Tesla P100, at least if the interconnect is anything similar to the NVLink.
Bear in mind, that the Tesla P100 offers 5.3 TFLOPs of double-precision and 10.6 TFLOPs of single-precision compute performance, while its NVLink promises 40GB/s interconnect bandwidth per GPU.
Unfortunately, there aren’t a lot of information regarding the Vega Cube but hopefully AMD will share a few more details soon.
