by Anton Shilov
11/22/2012 | 11:58 PM
The recent round of restructuring at Advanced Micro Devices hurt not only sales and marketing professionals, but also developers of new products. Besides, the new ambidextrous strategy brings a number of new things into the microprocessor development process. As a result, AMD is currently revising its high-performance x86 roadmap. The consequence of such review could be further delay of competitive x86 chips.
It is not a secret that fully-fledged Steamroller micro-architecture chips – such as AMD Opteron for servers as well as FX-series for desktops – have always been planned for 2014, the main question was whether the products were set to be released early in 2014 or late in 2014. Recently, it also transpired that AMD decided to cancel Steamroller-based Kaveri accelerated processing unit for mainstream market and produce Piledriver-powered Richland chip instead, meaning that there will be no Steamroller in 2014 at all. To make the matters even worse, SemiAccurate web-site reported that AMD actually killed Steamroller and Excavator projects as they are today. While AMD was quick to deny the information, this does not mean that there is no massive reconsideration of plans ongoing, especially considering the new plans to introduce ARM-based server solutions in 2014 and changes with the process technology-related roadmaps at the foundries.
Originally, Steamroller-based accelerated processing units and then central processing units were meant to be made using 28nm process technology at Globalfoundries (even though AMD never confirmed that officially). However, according to PC Watch web-site, there were issues with test APUs made at 28nm node, which was a reason why AMD decided to terminate exclusivity agreement with Globalfoundries so that to gain ability to choose the right partner for high-end APUs and CPUs. Nonetheless, since Steamroller designs are already tied to Globalfoundries 28nm fabrication process, the release dates of the formers depend on the ramp up and yields of the latter.
Globalfoundries is about to begin the ramp up of its Fab 8, which means that its abilities to produce 28nm chips in volumes will be quickly improving starting this December, but AMD decided to minimize the risks. The company, according to media reports, cancelled its Steamroller-based performance-mainstream Kaveri APU and substituted it with Richland APU (which features Piledriver x86 cores, but comes with latest-generation Radeon HD GCN cores) made using 32nm SOI process technology.
All-in-all, it looks like Globalfoundries is unable to start volume manufacturing of products based on Steamroller micro-architecture using 28nm process technology, whereas after the recent round of layoffs, AMD cannot redesign its chips quickly to improve yields in order to release them next year. Partly, this explains why AMD will not be talking about Steamroller at International Solid-State Circuits Conference 2013, where the company traditionally revealed details about its high-end x86 cores. Instead, AMD will reveal some additional information about its low-power Jaguar x86 micro-architecture.
In any case, without competitive x86 high-performance Steamroller core in 2013, the gap between general-purpose performance offered by AMD and Intel chips will get even wider next year as Intel’s Haswell is reportedly on-track for mid-year release.
What is interesting is what’s next for AMD? Globalfoundries vowed to start production of 20nm chips in the second half of 2013 and 14nm hybrid process tech in 2014. Will AMD be able to take advantage of those processes and catch up with Intel?
Even though many talented developers were either fired or left AMD in the recent years, the amount of projects that the company is simultaneously working on now has increased thanks to Rory Read’s ambidextrous strategy. The developers now have to work on high-end x86 server/desktop chips (FX-series, Opteron), mainstream x86 APU chips (A-series), low-power x86 APU chips (E-series) and various server and consumer ARM-based products. Ultimately, there are at least four major designs in development at AMD. In addition, AMD also has to work on improvements for its high-performance and low-power x86 micro-architectures.
The ambidextrous strategy implies usage of similar building blocks inside completely different chips. While it clearly lowers development costs, it also means that AMD will be unable to leap forward one particular product line. Each building block needs to be developed for a particular process technology to be efficient in terms of performance. At the same time, each building block needs to be tailored for different product designs to ensure high-performance across the board. Hence, to boost the speed of one product by taking advantage of a new manufacturing, AMD needs to redevelop virtually all building blocks it has and tailor them for all possible designs to ensure maximum performance and power-efficiency across the product family.
Given the limited amount of engineers at the company, the firm will hardly be able to quickly respond to changing conditions and adopt the latest process technologies among the first ones, simply because its CPU/APU product lines will be too broad. AMD will have to be rather conservative with transitions to newer process technologies as they will require redevelopment of building blocks for numerous chips.
To some extent, this is a reason why AMD has to revise its Steamroller, Excavator and post-Excavator/post-Bulldozer plans. Unfortunately, this roadmap change will likely bring a number of delays, hence, the Excavator may emerge even later than expected today.