Intel and AMD are both facing delays and product changes on the server front. While we alluded to some of these changes in our AMD earnings coverage from last year, it’s worth exploring them in a bit more detail. Intel has finally let slip some details on Kittson, the follow-up processor to Poulson and the last Itanium processor on any roadmap. Meanwhile, AMD’s roadmap for its Cortex-A57 based “Seattle” SoC has slipped a full year, with volume ramping in 2015.
Let’s kick off with Itanium. As Kitguru details, Kittson, which was originally expected to debut on a 22nm process, has been pulled back to 32nm. This shift underscores the facts that were laid out in the HP-Oracle lawsuit from several years ago — Itanium, as a processor family, is nearing end-of-life. HP is transitioning away from it, even at the highest-end of its server family, and HP is estimated to account for about 80% of all Itanium processors. Four years ago, Kittson was marketed as being 22nm and launching in the same time frame as Ivy Bridge and Haswell. More recently, it was shown as following Haswell.
At this point, killing Itanium probably makes more sense for Intel than trying to resurrect it. The chip was designed based on a series of assumptions about the future of computing and CPU architectures that didn’t prove accurate. It’s not uncommon to see enthusiasts pining for an alternative to x86, or blaming the microarchitecture for perceived deficiencies compared with ARM. But even if an x86 replacement were desirable, Itanium was never going to be that core. The underlying assumptions that drove its architecture, and the belief that pushing instruction scheduling and parallelism entirely to the compiler would result in a higher-performing CPU architecture, never proved true outside the narrow confines of the HPC and Big Iron markets.
Poulson offered a significant performance improvement over previous Itanium generations
Intel hasn’t detailed the improvements that Kittson was meant to introduce over Poulson,which substantially changed the architecture and introduced a variety of new features for both increased reliability and additional performance. Many of the reliability features have been added to Xeons as well, however, and the rise of heavy-duty Xeon E7 chips — now packing up to 18 cores — has limited the future for any Itanium variant. Intel may still bring Kittson to market, but it’s almost certainly the last hurrah.
AMD’s Seattle falls back to H2 2015
This is something we alluded to last week, but it deserves its own mention. During the conference call, when CEO Lisa Su was asked when investors could expect to see Seattle ramp for volume shipments, Su responded: “Seattle we continued to sample, and customers are continuing to develop both systems and software. Relative to the production ramp for volume shipments, I think we’ll see that in the second half of this year.”
This is a significant delay compared to what the company had previously stated. Back in 2013, AMD announced it would sample Seattle in early 2014, with volume shipments beginning later that year. In January 2014, AMD announced it was targeted a March sample date (this actually happened in July) with a Q4 2014 server announce date (this hasn’t happened yet). Now, Lisa Su has said that the chips start ramping for production by the second half of the year.
AMD’s 2013-2014 server roadmap. At no time did the company signal Seattle would still be ramping in H2 2015.
Why the delays? There are several possibilities. While the IP blocks and data from ARM for the basic design should be straightforward, this is still AMD’s first non-x86 CPU in a very, very long time. It’s also a product aimed at the server market, which means the amount of validation that has to occur is going to be higher than it would be for a consumer product. 28nm, however, is a proven node and the Cortex-A57 is a proven CPU design — which means AMD really ought to have had a product out by now.
It’s possible that AMD’s foot-dragging is being driven by market considerations as much as anything. Several years ago, analysts were confidently predicting that ARM would ascend into both the notebook and server spaces, challenging x86 for dominance at every stage. Intel responded to this rhetoric by aggressively positioning Xeon, Atom and Core M products in an attempt to cut ARM off at the pass. With the exception of Intel’s contra-revenue mobile shipments (which could be seen as a defensive move to shore up its market share while it worked on developing cheaper SoC manufacturing techniques), the ARM and x86 ecosystems have not come to blows. This is doubly true in servers, where ARM occupies essentially no space at all. Several companies, including AMD and Qualcomm, have talked about changing that — but there are few products.
AMD’s upcoming Analyst Day should answer some questions on whether the company intends to seriously push ARM as a server architecture. AMD still owns the Freedom Fabric IP, even if it closed SeaMicro, but the one advantage of owning the server retailer was that it gave AMD its own showcase for future products — including ARM processors and the x86/ARM hybrid platform known as Project Skybridge. Shutting the vendor down altogether could be seen as evidence Lisa Su disagreed with Rory Read’s oft-repeated assertion that dense servers represented a huge growth opportunity for AMD.
If AMD is still planning to launch Seattle in late 2015, the core may be short-lived. Project Skybridge was supposed to be built on 20nm technology and to offer HSA support across both ARM and x86. Once that platform debuts (if it does), the advantage to using an older 28nm Cortex-A57 would be small indeed. And while both Intel and AMD are pushing back businesses that account for a very small percentage of revenue, Intel is moving to shut down an initiative that was a distraction from its healthy businesses, while AMD had positioned ARM servers as a vital component of its own recovery.
The early rumors on Zen look promising, but we’ll have to wait a few more weeks to find out if there’s any truth to them.
No comments:
Post a Comment