AMD’s Vega GPU Cube Is A 4×3 Inch 100 TFLOPS Monster – Meet The Mini-Nuke Of Supercomputers

Author Photo
Dec 16, 2016
15Shares
Submit

AMD debuted its new family of Radeon Instinct GPUs for AI and deep-learning earlier this week, including the powerful MI25 featuring Vega. The company’s most advanced graphics architecture to date. One quite peculiar contraption was showcased on stage by Radeon Technologies Group’s head, Raja Koduri, which has caught our attention.

A small, 4×3 inch, quad GPU cube with one hundred TERAFLOPS of FP16 compute and half that in FP32. To put that figure into perspective, Nvidia’s liquid cooled Drive PX2 AI supercomputer for autonomous, driverless, cars delivers 16 TFLOPS of FP32 compute and 8 TFLOPS of FP16 compute. That’s less than one sixth of what the Vega cube is capable of. And while the Drive PX2 is a large box meant to go inside the trunk of a car, the Vega cube fits into the palm your hand.

amd-nvidia-feature-2RelatedAMD Wins Best Brand, Ryzen Wins Best PC Component, NVIDIA GTX 1080 Ti Wins Best GPU – TR Awards 2017

Meet The Mini-Nuke Of Supercomputers, AMD’s Vega Cube

AMD Vega Cube, held by RTG’s Chief Architect Raja Koduri

The prototype that Koduri showcased on stage is made up of four individual Vega 10 graphics processors, each residing on its own little circuit board. Vega cubes are designed to be stacked up vertically one on-top the other via a unique interface. This ,in theory at least, would enable the creation of extraordinarily powerful supercomputers orders of magnitude smaller than what we see today.

We can only imagine the challenge associated with cooling such a dense device. Especially considering that each Radeon Instinct MI25 accelerator, powered by a single Vega 10 chip, is rated at 300W of power. Nvidia’s own Telsa P100 deep-learning accelerator is rated at 300W However, it doesn’t come in a configuration that allows anywhere near the same computing or thermal density as the Vega Cube.

Liquid cooling would be an obvious route. One that Nvidia has already taken with its Drive PX2 box, in fact. Although it’s not the only one. We’ve seen AMD take its liquid cooled 275W R9 Fury X, powered by Vega’s older sibling Fiji, down to 175W in the form of the R9 Nano. And do it in exchange for less than 15% of the performance. A similarly power-optimized variant of Vega 10 with lower clock speeds is unquestionaly in the pipeline.

Replicating The Power Of The Human Brain, We’re Getting Close

Back in 2001, prolific futurist and one of the world’s biggest proponents of the technological singularity hypothesis, Ray Kurzweil, predicted that by 2019 a typical $1000 computer will match the processing power of the human brain. As things stand today, AMD’s Vega cube will be able to hit that performance mark next year, all be it at more than $1000.

amd-radeon-vega-feature-wccftechRelatedAMD Launching RX Vega 32, 28 & A Dozen New Vega 11 Cards, GPU Passes Certification

As Vega is introduced into the larger consumer gaming-focused market next year and as costs come down over the next couple of years, it’s entirely feasible that by 2019 we will actually have $1000 computers that match the processing power of the human brain. It’s important to note though that machines can only be as smart as the software running on them. Advancements in AI will be the key to converting that processing power into actual intelligence. Whether that’s going to happen within the next three years or not is yet to be seen. One thing’s for sure though, we’re getting incredibly close.

AMD Vega Lineup

Graphics CardRadeon R9 Fury XRadeon RX 480Radeon RX Vega Frontier EditionRadeon Vega ProRadeon RX Vega (Gaming)Radeon RX Vega Pro Duo
GPUFiji XTPolaris 10Vega 10Vega 10Vega 102x Vega 10
Process Node28nm14nm FinFETFinFETFinFETFinFETFinFET
Stream Processors40962304409635844096 (?)Up to 8192
Performance8.6 TFLOPS
8.6 (FP16) TFLOPS
5.8 TFLOPS
5.8 (FP16) TFLOPS
~13 TFLOLPS
~25 (FP16) TFLOPS
11 TFLOLPS
22 (FP16) TFLOPS
>13 TFLOLPS
>25 (FP16) TFLOPS
TBA
TBA
Memory4GB HBM8GB GDDR516GB HBM2TBATBATBA
Memory Bus4096-bit256-bit2048-bit2048-bit2048-bit4096-bit
Bandwidth512GB/s256GB/S480GB/S400GB/STBATBA
TDP275W150WTBATBATBATBA
Launch20152016June 2017June 2017July 2017TBA

 

 

 

Submit