AMD has allegedly decided to delegate the responsibility of producing its hotly anticipated Zen CPU chips to TSMC on its 16nm FinFET process. This development has allegedly unfolded amidst AMD fears of Globalfoundries' slow 14nm FinFET yield and capacity progress. And while these claims do seem plausible I should still provide a note of caution as this is very much a rumor at this moment in time and as such was tagged accordingly.
Globalfoundries was famously responsible for the delays to AMD's last round of make-or-break product launches. Which included AMD's first ever mainstream APU code named Llanno and its first high performance Bulldozer based server and desktop FX chips. These first gen 3000 series APUs and FX 8100 series CPUs were supposed to land in Q1 2011 but were delayed to Q4 2011 as a result of multiple challenges that Gloablfoundries' faced in bringing its 32nm SOI manufacturing process to full operation.
AMD Has Been Very Coy When Going On Record About Whether Globalfoundries or TSMC Will Produce Zen
Analysts and journalists have repeatedly asked AMD on various occasions about whether Zen will be manufactured at TSMC or Globalfoundries. I can recall two or three occasions off the top of my head of when this occurred and the answer was always the same. Zen will feature FinFET technology, both TSMC and Globalfoundries are partners of AMD and that was that. It's becoming much clearer now why AMD chose not to commit to any one of the two early on.
Globalfoundries has reportedly suffered a setback recently as a result of its owner, the government of Abu Dhabi, deciding to cut back on expenses after oil prices crashed, which have so far shown no signs of bouncing back. As such GloFo is reportedly attempting to retool its 28nm equipment for 14nm production. This retooling however has been considerably slow and is the reason behind the recent delays in bringing volumes and yields up to snuff by the pure-play foundry.
TSMC has undoubtedly faced its fair share of setbacks but it has successfully managed to iron out most of the kinks and its 16FF+ high performance process is set for volume production this quarter. NVIDIA confirmed that it will partner with TSMC to produce its next generation Pascal GPUs next year. This past August Digitimes confirmed that TSMC will in fact be making chips for AMD on the 16FF+ process. And as AMD GPUs have historically always been made at TSMC it's safe to assume that AMD's next generation Arctic Islands will also be produced at TSMC.
AMD’s Zen CPU Core Coming in 2016 – Features FinFET, SMT and New Cache System
Zen will be AMD’s successor to the Bulldoze family of cores. It will be the first ever CPU core from the company to adopt SMT , Simultaneous Multithreading. Zen also represents AMD's very first brand new from-scratch design since the introduction of the company's CMT based, Clustered Multithreading, design featured in the Bulldozer family of processors.
In addition to SMT, Zen also features a new high-bandwidth low latency cache system, a pivotal design improvement over the previous generation of cores. Since subpar cache performance was one of the primary pitfalls of AMD’s Bulldozer CPU microarchitecture. AMD announced that Zen will be available starting next year. The new CPU core will be fabricated using FinFET technology, however the company has not yet specified whether it will be Globalfoundries' 14nm Finfet process or TSMC's 16nm FinFET process.
Zen Can Process 40% More Instructions Per Clock Compared To AMD's Latest CPU Core "Excavator"
Mark Papermaster, AMD’s Chief Technology Officer, revealed that Zen will have a huge improvement in IPC, Instructions Per Clock, vs "Excavator". Which is AMD’s latest ,and last, Bulldozer family based CPU core. A 40% increase in IPC would represent the largest jump in IPC ever for the company. And the biggest jump in performance that Papermaster claims to have ever seen in his three decade long professional career.
Mark Papermaster also made it a point to highlight that this 40% performance improvement figure is independent of the manufacturing process. So it’s a permanent architectural performance improvement over the previous generation that will always be carried over regardless of the process node flavor chosen to make any of the products featuring Zen.
To put this massive leap in performance into perspective, Intel only managed to increase IPC by single digit percentages with each generation for the past four years. And a 40% IPC improvement would instantly put AMD in a very competitive position with Intel's fastest X99 enthusiast processors. According to the aggregate single threaded PassMark database a 40% IPC improvement over the Steamroller based A10 7850K would put Zen on par with Intel's latest Skylake i7 6700. Which is a paramount improvement, a comparison that uses AMD's Excavator core as a base for the 40% figure would put Zen at an even more favorable position.
If you take that sizable IPC improvement and couple it with Zen's feature set namely SMT, DDR4 support and high core counts - allegedly up to eight cores - it's easy to understand why AMD believes it's got a winner on its hands. The same level of performance and set of features is available today only with Intel's enthusiast class $1000 i7 5960X CPU and the X99 platform.
Zen will form the backbone of AMD's enthusiast FX CPU product line in 2016. Lisa Su confirmed that the new CPU architecture will be arriving to desktop FX CPUs first and to servers second. Succeeding Zen will be Zen+ cores, which will feature incremental evolutionary improvements over Zen. The company will introduce a new socket in 2016 dubbed AM4 that will house products spanning high performance CPUs to mainstream APUs based on Zen and next generation FinFET GCN based GPUs.
The new family of high performance FX CPUs which we've seen labeled unofficially as the “Summit Ridge” family of products will officially feature high core count CPUs with multi-threading, DDR4 compatibility and a brand new platform dubbed AM4. Previous leaks indicate that the Summit Ridge family will include CPUs with up to 8 Zen cores and 12MB of L2 and L3 cache, however AMD has yet to confirm any of those details.
