HSA, heterogeneous computing: Intel and Nvidia iso
Written by Damien Triolet
Published on October 4, 2012
At the beginning of June, AMD set up the HSA Foundation in partnership with ARM, Imagination Technologies, MediaTek and Texas Instruments. This foundation, remember, aims to come up with standards for heterogeneous computing, whether with respect to programming or hardware implementation. It rapidly garnered some heavyweight support.
Before being swallowed up by AMD, ATI was the first to profess the ambition to use GPU processing power for ends other than real time 3D rendering, the reason for which it was originally designed. No doubt through a lack of means, these developments advanced very slowly and we had to wait for over a year, with the concretisation of a similar initiative from Nvidia, for GPUs finally to manage to get a foot in the door of high performance computing. Available since the beginning of 2007, CUDA relegated any similar ATI/AMD initiatives to the background.
Several twists and turns in terms of technological choices and programming languages, as well as the time taken for ATI and AMD to merge, meant that progress was slow on the AMD side. With the AMD Fusion project, the aim wasn’t simply to exploit GPU power but to create a symbiosis of GPU + CPU. Moreover, AMD chose, probably by default, to base its work on open standards. In contrast, Nvidia went down the proprietary route, giving more agility and enabling it to develop a lot faster.
What with the x86 world, largely dominated by Intel, and GPU computing, dominated by Nvidia, AMD found itself in a delicate situation in which it became difficult to have much influence over the technological choices of developers and persuade them to program for heterogeneous solutions.
To get out of this impasse, AMD needed to rally other players to its cause. Offering an architecture standard for heterogeneous computing was a natural solution, especially as such a solution was becoming essential for many other players: ultra low energy consumption SoC designers. When the thermal envelope is limited, as is the case for all mobile peripherals, being able to use different types of core designed for processing (sequential or massively parallel) allows you to maximise performance in more instances. In other terms, the whole ARM ecosystem was predisposed to the use of heterogeneous computing and was going to come up against the same issues as AMD in terms of finding the best approach to put it into place.
In mid-2011, AMD put forward FSA, Fusion System Architecture, as a starting point, with ARM in the wings. A year later, with a change in name to HSA, Heterogeneous System Architecture, AMD passed all its initial work over to a foundation whose initial founder members also included ARM, Imagination Technologies, MediaTek and Texas Instruments. The foundation's statutes however left open the possibility for other players to become founder members if they were to come on board within three months as of June 1st 2012.
A few days from the deadline Samsung joined as the sixth founder member, along with Apical, Arteris, MulticoreWare, Sonics, Symbio and Vivante as secondary members. Although the arrival of a heavyweight such as Samsung was good news for HSA, the absence of Qualcomm was very surprising. With very significant objectives in terms of its SoC capacities and the arrival of Eric Demers, the ex-head of AMD GPU architectures, as head of its engineering department, there could be no doubt that Qualcomm wanted to join HSA… and not only as a secondary member.
The negotiations were no doubt more complicated than expected but in the end the HSA Foundation changed its statutes to remove the deadline for entry for new founder members. Qualcomm thus became the seventh founder member yesterday.
By linking up with all the major players in the ARM ecosystem, AMD has no doubt come up with the best means for developing a heterogeneous computing standard and the graphical presentation of the foundation’s site leaves no doubt that the door remains open for an eighth member. While certain important players such as Apple and Microsoft are still to be convinced, the major absentees are Intel and Nvidia.
These two, on the one hand because of their rivalry with AMD and on the other because they don’t want to facilitate the arrival of competition in what are very juicy markets, remain hostile to such a standard. Intel wants to keep total control over its platform, offer its own solutions for massively parallel computing and favour the use of x86 cores which are now greatly improving in energy efficiency. On its side, Nvidia doesn’t intend to sabotage the commercial success of its Tesla division and the CUDA proprietary architecture and is getting its own heterogeneous solution ready.
There’s no doubt however that Intel and Nvidia will be following HSA developments very closely to avoid finding themselves isolated from the rest of the industry. The first specifications were initially announced for the end of 2011 but were delayed and are now in the hands of all the members of the foundation, with publication expected before the end of the year.
Copyright © 1997-2014 BeHardware. All rights reserved..