ADFS 2011 Wrapup & AMD Lynx Platform Tests

Company: AMD
Author: James Prior
Editor: Charles Oliver
Date: July 18th, 2011

Summary and Conclusions

With the new Accelerated Processing Unit (APU), AMD has a new platform and Fusion Controller Hub (FCH) with native USB 3.0 and a full complement of high-speed SATA-3 6Gbps storage ports, in mainboards sporting dual-graphics support. This first generation of APU uses existing designs and technology to set a high baseline for heterogeneous compute and provide a solid affordable mainstream platform, with our Gigabyte GA-A75M-UD2H mainboard retailing for ~$100USD and the A8 3850 for ~$135USD.

AMD has now differentiated their platforms into three strata, two of which are APU based and the third yet to fully appear. At the low end, for appliances and small form-factor desktops, is the Brazos platform. Attacking the mainstream desktop market is Lynx (Sabine in notebook form), where it is priced competitively. While price comparable Intel second generation Core i-series processors may have single-thread or benchmark advantages in CPU heavy tasks, AMD believes these circumstances are less and less relevant to the modern computing expectations of today's consumers. AMD's graphics performance with Llano far exceeds that of Intel's HD graphics in gaming performance and image quality.

The Lynx platform offers great value for system builders looking to provide a great desktop experience, especially as the era of GPU accelerated applications continues. Microsoft Office 2010, Corel, Arcsoft, and Cyberlink products, and web browsers, now leverage the GPU to enhance basic computing tasks. It's not perfect, as the second PEG slot is x4 speed, making it a general purpose expansion slot instead of a nice platform for CrossfireX performance or enthusiast class cards. This is also reflected in how the headers and sata ports are placed along the lower edge of the mainboard, where using a dual slot full length card will prevent use of front panel audio, usb headers, etc. etc.

The inclusion of legacy connections like parallel and serial port, and a PCI slot, show Gigabyte is considering as broad a market as possible, although removing the IDE port and including a TPM module header shows a forward looking design, too. Four USB 3.0 ports is a nice addition, although we'd have like to see a backplane header included for those that don't have a case with USB 3.0 header installed; perhaps higher models will do so as a 'value add'.

The A-series APUs are competent gamers at mid to high settings and half-HD resolutions like 1360x768 and 1600x900. For high definition media playback they are superb, with the Radeon core horsepower of the Vision Engine providing plenty of image quality enhancements, including Picture Perfect HD and Smooth Video. Dual Graphics Crossfire with the new HD 6500 and 6600 series cards offers a low power, cheap performance upgrade and really is the knife to Intel's throat - add a Turks card to a Core i3 2105 equipped system and you get the same gaming performance as an A8 3850, at higher cost. Add the same card to the Llano desktop and you get an approximate extra 75% performance, for the same money. That's pretty compelling.

Llano's GPU (Sumo) relies on system memory for performance, and AMD have equipped Lynx with support of up to 1866MHz DDR3. It does make a difference, most noticeably to DDR3-1600. By offering official support for DDR3-1600 in four DIMM and DDR3-1866 in two DIMM configuration, AMD have granted a lot more flexibility and scalability than previously was available on platforms with integrated graphics. UMA + Sideport dedicated memory can't match this flexibility or performance.

While evaluating the Lynx platform, the experience flip-flopped from 'I can't believe this is one chip doing this!' to 'Meh, seen it before. What's new?' This stems from the fact that this is 'old' technology, brought together in a new package. Now, every innovation has to start somewhere, and the wow here is the squeezing of ~170W worth of technology into a 100W TDP package, and selling it for around the same price; plus the new form factors that you can squeeze that power into. That's the problem, too, as you've got what you had before, until something really leverages the GPU compute capabilities and desktop apps take off again, as x86 did five years ago with the multi-threaded evolution. All signs point to that being in about 24-36 month's time, as developer toolkits get better, more accessible and the Fusion system architecture evolves with lots of software over the next 12-18mo becoming better at leveraging Fusion APUs. So why buy now, if 2-3 years is when the awesome happens?

The people who buy mainstream platforms tend not to replace them every year. They buy OEM prebuilt machines, or pre-assembled ones from value-added resellers locally. They may upgrade them every once in a while, and that's usually something basic like RAM or a GPU, not a platform upgrade. For them, AMD's Lynx platform is very interesting, as the machine they buy is likely to get faster over time, as the compute power is unlocked. Businesses gain higher fidelity graphics and rendering capabilities for the same entry cost, meaning more powerful software can be considered for daily work flows. Users receive better baseline media consumption and gaming experiences.

Is the A8-3850 that badly named, then? If we think back to another 3800 series branded product, it faced a larger, more powerful and more expensive competitor. Its next evolution was significantly more powerful and forced a lot of change, forcing competitor price drops. The evolution after that, chewed gum and kicked ass. APUs are on that same cadence now and we've seen AMD promise 50% more performance next year, with a new CPU and GPU core architecture inside. The year after that, when the APU inflection point looks ready to skyrocket, is when the really good stuff is due - like DDR5 on the mainstream desktop. Problem, memory bandwidth?

AMD, ARM, Intel and NVIDIA are all heterogeneous compute processor companies now, but they know it to different degrees. Intel is way over on the left, trying to turn the x86-behemoth with small, fixed function dedicated hardware for graphics like tiny bow thrusters on the Titanic, and bypassing CPU socket power and TDP limits by putting dozens of x86 cores on an add-in board. All the way over on the right is NVIDIA, making a Denver omelet with lots of GPU compute clusters and some ARM technology, transitioning to being a software company making their own hardware to run on. In the middle, AMD and ARM are looking at a huge market and opportunities for partnership (heyhowareya, Big Blue?), making a big pie for them both to get double helpings of with cream and cherries on top using Fusion System Architecture.

Around 2013, mainstream software will be ready to rely on personal supercomputer power, and it's a race to get hardware platforms, ISAs and development tools in place for lift off. The future is Fusion, and AMD's heterogeneous compute monster called Llano is Fusion executed almost perfectly. The only thing missing is desktop switchable hybrid graphics and APU Eyefinity, although if we're wishing we'd want a Black Edition with tunable TDP balancing and unlocked TDP limit, as well as unlocked multipliers.