Advanced Micro Devices Inc. displayed a new swagger at last week's AMD Fusion Developer Summit (AFDS) in Bellevue, Wash. Weary of trudging through life as perennial second fiddle to Intel Corp., AMD outlined a bold new direction and displayed flexibility that may not have seemed possible just a few years ago.
AMD is seeking to build an ecosystem around its open HSA to increase the number of developers using its APUs and profligate heterogeneous computing into the mainstream.
"We have bet our company on heterogeneous computing," said Lisa Su, senior vice president and general manager of AMD's global business units. "We believe this from the core of who we are—that this is how we can change the world, that this is the future of computing."
Not to be outdone in the gambling analogy department, Mark Papermaster, AMD's chief technology officer, later told attendees that the company was "doubling down" on its strategy to leverage its IP in flexible and more agile ways to deliver value to customers.
Mark Papermaster, AMD senior vice president and chief technology officer, speaks at the AMD Fusion Developer Summit last week.
AMD has already shipped some 40 million APUs since the first was released in early 2011. Su said 11 of the top 12 OEMs are currently shipping APU-based platform. But heterogeneous computing in its current form is difficult to program. AMD estimates there are perhaps 100,000 developers with the skills and experience required to develop applications for parallel computing, which requires multiple languages and separate versions of a program. These developers are such a rare breed that they are sometimes referred to as "ninjas," said Tom Milloy, senior vice president and chief architect at Adobe Systems Inc., who delivered a keynote address at AFDS.
AMD's goal is to increase the ranks of ninjas—or, more accurately, to decrease the complexity of heterogeneous computing programming to the point where one need not be thought of as something as grandiose as a ninja to do it well.
According to Su, AMD, largely through the HSA Foundation, wants to make APUs as easy to program as mainstream CPUs, thus increasing the size of the pool of developers who can use them. The bottleneck, she said, is that this point mostly in software, something that the HSA Foundation—which thus far also includes Texas Instruments, ARM, MediaTek and Imagination Technology—aims to alleviate. AMD's vision, which began taking shape last year when the company took the bold step of opening up the architecture, is to drive the creation of open source APIs while at the same time enabling AMD and other chip vendors to offer proprietary hardware under the hood. AMD expects HSA member companies to bring to create developer tools, software developer kits, libraries, documentation, and more.
"We want to be very clear on a standards front that we want to drive an open ecosystem," Su said.
I think they're on their way to finding their own niche in the market (and not just based on price). AMD's current A and E series APUs are successful products that accomplish quite a few objectives:
1. Good integrated GPU performance that surpasses Intel's offerings. AMD is right about CPU performance: beyond a certain point, you just don't need anymore. The value is in the graphics and gaming capability.
2. Competitive power consumption. This is the first time AMD has been competitive in this area.
That being said, Intel has taken note and is planning a counterattack. Their next generation CPU, Haswell, will have potentially 2.5x the performance of current Ivy Bridge processors. They might actually exceed AMD on pure GPU performance.
Given this reality, I think AMD's new focus on software through HSA is really the only logical next step. While Intel is busy brute forcing CPU, GPU performance and process nodes, AMD is executing its (shared) long-term vision for system architecture. And I do believe this will pay off for them. I am looking forward to Kaveri, the first APU to have a unified memory and a gpu with the GCN arcitecture.
I have always preferred AMD's processors and graphics for price performance in desktops for software that needs an x86 and Windows (normally used Mac or Linux). However, after a stint on Alchemy Semi Au1000 families, when AMD took that over, it withered away and almost died. MIPS vs ARM is a story of greed for licensing, otherwise why would an architecture that was 64-bit and multi-core long ago fail so badly. ARM gets a lot of hype, and I have programmed plenty of them, but AMD has the lost ground of Alchemy to recover before banking on ARM converts. Tell developers how the graphics work so we can use them without NDAs like Imagination or NVIDIA. Bare-metal is not Linux, so a binary and an API is useless. Proper data sheets are required.