AFDS 2012 Day 2

Author: James Prior
Editor: Charles Oliver
Date: June 18th, 2012

AFDS Day 2: Tuesday, June 12th

Tuesday was what AMD considered the first day of AFDS 12 despite the previous day's sessions and demonstrations. For me, it began with powering through bacon, croissants, and assorted fresh fruit from the provided breakfast in the Experience zone. Suitably invigorated, it was time to grab the shuttle bus over to the Meydenbauer center for the keynote addresses. Last year, all sessions were held at the Meydenbauer center, but this year only the keynotes were held there with all the sessions and demonstrations being held at the Hyatt Regency. This is due to the increase in sessions held at AFDS '12, which exceeded the capacity of the Meydenbauer. Rather than split everything across the two locations, it was decided to have only the keynotes at the Meydenbauer and everything else at the hotel. This was largely beneficial, as they two locations are only a few blocks apart and there were frequent shuttle buses between the two locales, plus most attendees were staying at the Hyatt Regency making it roll-out-of-bed easy to get to the content on offer.

Keynote #1 - Lisa Su, AMD

The first keynote of the morning was delivered by Lisa Su, Senior Vice President and General Manager, Global Business Units - which is to say, she's steering the boat. A new hire to AMD, joining just 6 months ago, her presentation was much better than expected from a senior C-level exec. This is largely due to the fact that Dr. Su, equipped with doctorate and masters degrees from MIT, has a strong engineering background rather than one of pure management and corporate schmoozing. Speaking with her in person later, she's a very matter of fact and straight forward person, quick to invite criticism and wants to understand where AMD can do better; an honest and refreshing approach to understanding gaps in AMD's products and owning the improvements that need to do better.

Quote of the day came from Lisa Su's keynote - 'AMD is betting the company that the future is heterogeneous compute' with an 'extreme commitment to performance' and a desire to 'deliver exception experiences' using 'complete synergy between compute units'. Having entered the keynote expecting to be somewhat bored and underwhelmed by the corporate doublespeak, it was refreshing to hear genuine excitement and energy for the future of the company from one of the key people in charge of actually doing it.

Keynote #2 - Tom Malloy, Adobe

Next up was Adobe's Chief Software Architect, Tom Malloy. While the keynote contained the inevitable tour-de-force of Adobe products, it was related to the topic at hand by demonstrating how Adobe has walked away from proprietary acceleration code - CUDA - in favor of OpenCL and OpenGL. Adobe has a long history of optimizing to the hardware and will create the tools that are needed to get there. Adobe has learned many lessons from the needs of innovation driving use of high maintenance, highly complex code that requires 'ninja'' programmers - whom themselves can be high maintenance and high needs to keep.

While Adobe likes the first benefits of Heterogeneous System Architecture (HSA) such as unified address space, which will help bring mainstream developers into the realm of heterogeneous computing, there is a problem; OpenCL (and CUDA) are too low level for mainstream programmers. The current level of expertise needed for stable, optimized, high performance heterogeneous computing is at a high level and only the companies that can afford the ninja programmers and their support can do it. Everybody else waiting for more accessible common platform to emerge, which may end up being born out of C++ AMP and tools to go with it, but will need to be cross-platform and vendor agnostic.

Tom also talked eloquently about the user experience and processing power problem; typically a task that user tries to do is limited by the hardware until the hardware vendors develop ways to speed that task up, which the software guys then leverage. The problem then becomes that the hardware keeps getting faster but the user experience doesn't change, essentially making it a performance surplus. This tends to be the case on user interaction operations, but not always - think about filter processing. Then there's batch processing, where there is no constraint on what level of performance is enough, because the user is batch processing as a 'set and forget'; each iteration takes long enough that the user doesn't want to waste time waiting for each operation to complete before kicking off the next item or process, and the user wants to use the 'dead' time for working on something else. The performance surplus created by the new task being hardware accelerated then goes to being a deficit to batch processing, and is the key to opening mainstream adoption.

AMD's Manju Hegde, co-organizer of AFDS as well as corporate Vice President in charge of Software Development, spoke about how AMD waited to lock in the dates for AFDS '12 until Tom Malloy could confirm his presence. Tom's keynote was very good and enjoyable on both consumer and developer levels, well worth the time.

AFDS Organizers

Keynote #3 - Phil Rogers, AMD

The final keynote was the big one for the day, with AMD Fellow Phil Rogers unveiling the new non-profit HSA Foundation. The HSA Foundation is founded by AMD, ARM, Texas Instruments, Imagination, and MediaTek and is founded with the purpose of creating, developing and furthering the Hetergeneous System Architecture Interpretation Layer (HSAIL) as the new platform for PCs. At this point it's important to remember what PC means - Personal Computer. Apple changed the definition of PC with their ad campaign by saying they weren't a PC, only drab business class Windows boxen are. This was never true, just a great way to get into the collected masses consciousness but the fact remains PCs come in more form factors than ever - phones, tablets, ultra-portable notebooks, notebooks, portable desktop replacements, all-in-one devices, workstations; all made better by HSA. We're not in the Post-PC era, we're in an explosion of PC devices in many new and long awaited form factors that will facilitate some fantastic user experiences.

The theme of Heterogeneous System Architecture, and the HSA Foundation, is to make devices easier to create and deliver great user experiences on. This means making them easier to design with a variety of intellectual property inside, and easier to write applications for. Common standards do this, and the HSA platform is an open standard that members can contribute to, use, and make money from, at both hardware and software levels. When Rory Read said it's not AMD vs. Intel anymore, this is to what he was referring; it's HSA vs. Intel, and NVIDIA. Niether of these companies are locked out from HSA, and publicly AMD and the other HSA Foundation founders encourage them join. Privately, they still wish they would, but are skeptical of such a change in approach to IP and platform sharing from these companies, which is a shame.

The talk covered a lot of different aspects of programming for HSA, with the most resonant message for me being how compilers need to change. For thirty years they have been optimized for instruction count - keep the number of cycles used down, on the basis that provides better performance. Now we're seeing, despite the massive escalation of addressable memory in the 64-bit era, that memory is the most precious commodity. In the middle of the spectrum is the current notebook and desktop PC segment, where if you've got 4GB you've got plenty for 99% of uses. The pinch is felt at the bottom end, and at the top end; cloud computing, HPC, and big data processing all use gigantic gobs of memory which is expensive to buy and hard to get addressed by all your compute. At the bottom end are smart phones, tablets and embedded, where the amount of storage available is limited by the form factor, power consumption, and platform cost. In both these arena's it's becoming more popular to compile for optimized memory use rather than cpu cycles - it's more efficient to recalculate something when it’s needed than to calculate it and then store it, retrieve it, retrieve it, retrieve it, across a slow bus or bottlenecked interface. Another announcement from AMD was the parallel primitives library for HSA, named Bolt. This is a C++ template library for HSA that supports C++ AMP and OpenCL to allow developers to get code working on HSA platforms quickly and easily.

With the morning keynotes out of the way, it was time for a media lunch and round table with the newly-announced HSA Foundation founder members. The panel was led by Insight64's Nathan Brookwood, and from the various questions and comments from the assembled worldwide press it was clear that HSA is simple on the face of it but hard to understand in terms of what it means in specifics for each company. AMD made themselves relevant to a massive market of differentiated products that can be targeted to the programmers that are going to need very different hardware for their specific needs. Commodity devices that were being considered 'good enough' are falling off the back end of the acceptable curve; you've got to do better than be average. All the members of the HSA Foundation benefit from the new organization, as it opens up access to bigger markets. Many developers span multiple platforms these days, so the ability to write once and run on many is a big attraction. Currently, doing this in interpreted runtimes like JAVA and Flash have their own drawbacks, and aren't the best option for performance or taking full advantage of a particular platform features. HSA changes that, with the goal being allowing developers to write high quality code faster and get their product to market in better shape.

Afternoon Session - Mike Wolfe, AMD

After lunch it was session time, and there were literally dozens to choose from in the 1pm – 6pm block. I attended a session from AMD's Mike Wolfe, Chief Information Officer and recently named 'Most Transformative CIO'. Mike spoke about how AMD eats its own dogfood - AMD business runs on AMD products. Obviously they have their own competitive labs to make sure they understand where they are, but the job of the CIO is to enable the running of the AMD business as best it can; AMD is the customer of AMD's IT infrastructure, after all. At the same time, AMD IT's group is the customer of AMD, and that relationship isn't one way - day one of the new CEO's tenure, Rory Read approached Mike and asked him to tell him 'Does this stuff work the way we sell it does?' That honest and open internal feedback loop is critical, and heartening to hear is in place and at such a high level.

Under Mike, AMD consolidated and centralized several small departmental and regional datacenters into larger datacenters. This consolidation, combined with virtualization, allowed AMD to increase utilization as well as reduce project completion time through dynamic allocation; AMD's poster child for this process is the Brazos project, for which they decreased a two month validation job to 5 days. Additionally, more storage was freed up as duplicated files were reduced. Centralizing the data with the storage also reduced time to bring up projects, as now when projects were begun or changed from one team to another, large amounts of data weren't needed to be moved over the wide area network (WAN) to get to the compute - it's already there, and only the access needs to be changed.

Mike is also in charge of the group who design security into AMD's projects, and he's very aware of the needs of customers to operate in a secure manner. AMD need more secure IT infrastructure, so they're going to add it to their products through the use of a co-processor: an ARM Cortex-A5 CPU with TrustZone technology. Given the choice of inventing their own platform and ecosystem, from hardware instructions and interconnects to a programming model and API, or licensing one already in wide use with strong support, the choice was simple especially given AMD's new HSA Foundation. Additionally this supports AMD's message of 'write once, run many' as existing code for TrustZone can be ported to AMD APUs equipped with TrustZone much quicker now. It's interesting to note that Charlie Demerjian of Semiaccurate believes that this announcement is very, very late - the technology block might already exist in certain APUs but isn't turned on yet. Waiting for Windows 8?

All work and no play ...

The day was rounded out with a boat trip for the media and select AMD executives, a chilly but enjoyable ride around Lake Union seeing some interesting architecture. As a way of getting a mixer between execs and press for a quick feedback and light interview setting it was a pretty bad one, given the hugely annoying cover band and rather dire weather for sitting out on a boat, but it was a nice way to unwind a little bit and relax after spending a day cooped up in convention rooms. Attendees and Speakers were free to find their own entertainment for the evening, and there are plenty of small bars and cafes near the hotel (and in the Hyatt Regency) to find some very interesting people to sit and chew the fat with. Walking around were people from Intel, Sony, many major game companies, and some soon-to-be-disruptive GPGPU companies as well as developers from all walks of programming.