Pages: 1 2
Roadmap Update
Modern SoC projects have gestation period between 4-6 years, and during the last 1-2 years, the design is essentially frozen. This means that for a management team that has been around for less than 6 months, the only real option available is to a cancel an upcoming project; there is no way to add more performance or accelerate time-to-market. With only a couple of small changes in strategic direction and few levers for near term products, AMD’s 2012 roadmap is essentially unchanged. Longer term, AMD will continue to have three key lines of IP blocks, a high performance CPU (named after construction equipment), a low power CPU (named after feline predators) and a GPU (named after islands where AMD engineers wish they were on vacation).
In 2012, there are 3 products to span the client SoC market. At the high-end is the 32nm Trinity, which packs up to four Piledriver cores (the second generation of Bulldozer with FMA3 support) and a modern GPU with aggressive power management. A second generation of the 40nm Brazos platform will target low-power PC systems. Hondo, a 40nm SoC with a single Bobcat core is too high power (roughly 4.5W) for tablets, but represents a stepping stone to the mobile device market. The 28nm Southern Islands family will cover discrete graphics, workstation and compute-oriented cards.
The new 2013 roadmap moves AMD’s entire client product family to 28nm manufacturing. 28nm is a substantial improvement over 40nm, with a full node shrink and perhaps more importantly, the introduction of high-k/metal-gate transistors that should benefit performance and power. While moving to 28nm in 2013 delays these benefits, it also improves the economics by waiting for stable yields and greater availability. For low cost parts, this trade-off seems reasonable.
In 2013, AMD will have three major client SoCs, which rely on several different IP blocks. At the high-end is Kaveri, which incorporates 4 Steamroller cores that may have full 256-bit AVX support. The precise power targets were unclear; Kaveri can probably reach 25W, but not much lower. Kabini is the successor to Brazos and fills the gap below 25W. It will probably use 2-4 Jaguar cores, depending on the power budget and integrates the I/O.
The biggest improvement for AMD’s roadmap is Temash, a tablet-optimized SoC with 1-2 Jaguar cores and integrated I/O. While Hondo is really not suitable for tablets, it is clear that with a shrink to 28nm, Temash should be able to run as low as 2W. That should be especially compelling for more sophisticated Windows 8 tablets, as opposed to the inexpensive e-reader style tablets.
The graphics for these 3 SoCs appears to be derived from Southern Islands and the recently launched Radeon 7970 and 7950. This represents a lag of 12-18 months behind discrete graphics, which is reasonable given the additional validation challenges for SoCs. Hopefully, AMD can tighten up engineering schedules so that new GPU architectures cascade down in 6-12 months. The new 2013 GPU architecture is called Sea Islands and should feature a number of programmability enhancements and tighter x86 integration.
In servers, the roadmap has actually been redrawn to be substantially less aggressive for 2012-2013. Previously AMD intended to migrate to a new server socket and platform infrastructure, with PCI-E 3.0 integrated into the microprocessor in late 2012 or early 2013. The new server processors were planned with 10 second generation Piledriver cores, which would bring FMA3 support, but probably not full 256-bit AVX execution pipelines. That schedule would have put them about 6-12 months behind Sandy Bridge-EP and would have further focused AMD servers on scale-out workloads.
Instead, AMD has pushed out their new platform architecture to 2014, meaning that the current server and high-end desktop sockets will last for around 3-4 years, rather than the customary 2-3 years. This also puts them nearly 2 years behind Intel in terms of PCI-E integration. However, when the new platform arrives in 2014, it will come with a huge increase in performance. In addition to integrated PCI-E, it is likely that the new platform will have DDR4 and several other enhancements. In the mean time, the next iteration of servers will have 8 Piledriver cores and hopefully use the additional power budget to increase the clock frequency and thus single threaded performance. These new desigsns will be socket-compatible, which is nice for OEMs and the less aggressive roadmap should also be much easier for AMD to deliver to customers.
Analysis
In client systems, AMD’s main goal is to increase market share for notebooks and get into tablets. Overall, the roadmap looks reasonable to that end. Trinity is a substantial improvement in power efficiency and graphics performance, particularly for mainstream 35W notebooks. While it can fit into 17W form factors, it is not a superb fit, but will act as a prelude for future designs. The 40nm Brazos will have to continue to compete on the basis of price, when faced with 22nm and 32nm designs from Intel. 2013 should be much better for AMD, as the 28nm product portfolio can hit tablets at around 2W and should have strong offerings for 17W ultrathin notebooks. The thinnest and most power efficient notebooks will continue to be outside of AMD’s reach, but each generation is improving significantly and opening up new doors and more design wins.
On the other hand, the roadmap suggests that AMD will fall further behind for most of the server market. The transition from the Bulldozer core to Piledriver should have a nice benefit for HPC applications that can take advantage of FMA3 (as opposed to AMD’s current FMA4 which is not supported by Intel). Software that adopts FMA can theoretically double compute performance, but most software will see much smaller gains and FMA is of limited utility outside HPC. It is possible that some server processors will feature four channels of DDR3 without an MCM, which would achieve significantly higher single threaded performance and lower cost. It is clear that AMD will not ramp up performance and power efficiency at the same rate as Intel for general server workloads, but the new Piledriver-based servers might be attractive for HPC. Overall, AMD should be able to mildly increase market share, but 10% seems like the limit and any further improvements will be dependent on the new server platform in 2014.
Discuss (38 comments)