Last week Intel revealed the new 9500-series (Poulson) Itanium processors that finally replace the previous generation 9300-series Tukwila CPUs which were released nearly three years ago. The 9500-series processors utilise the Intel Itanium Architecture (also known as the IA64 ISA), and are used in large mainframe-like systems for mission-critical applications where redundancy and availability are of the utmost importance.
However, in addition to the error detection and redundancy features, the new Itanium processors bring several performance related improvements. Specifically, Intel managed to pack 3.1 billion transistors into the (32nm) Itanium 9500-series parts [PDF], increased the core count to eight physical cores with two threads per core (coarse multithreading), doubled the amount of memory supported to 2TB in 4P servers, and included up to 54MB of cache (32MB of which is L3). Intel has also managed to ratchet up the clock speeds by 40 per cent to a maximum of 2.53GHz, while lowering idle power consumption by 80 per cent and the TDP rating by 8 per cent to 170 Watts on the top-end parts. According to Intel, the chips will offer up to twice the performance of previous generation parts.
The performance improvements have been a long time coming, and will allow Itanium to maintain some relevance for big number crunching applications. Unfortunately, in the interim between Tukwila and Poulson, Intel’s own Xeon (x86-64) processors have become the king of the server world in terms of performance.
To make matters worse for Itanium, the Xeon chips have integrated some of the specialised RAS (reliability, availability, serviceability) features previously only available with Itanium chips. Further, Red Hat and Microsoft have announced their intention to drop support for IA64 processors in future OS versions. As a result, Itanium has been pushed further into niche status than ever before.
In addition to Xeon and Itanium (already) trading features and sharing a chipset, Intel has also announced that it plans to merge the two server chips even further, going so far as using the same CPU socket and motherboards – with the only appreciable difference being whether the CPU is based on the x86-64 or IA64 ISA (Instruction Set Architecture). This merging of architectures will happen with the 22nm successor to the 9500-series, known as Kittson, and Haswell EX-based Xeon processors. Currently, rumours suggest that an electrically re-engineered LGA 2011 is the shared CPU socket of choice, which will be modified to provide additional power for chips with up to 180 Watt TDP ratings.
The stated intention is to reduce the cost of the supporting hardware components required to use Itanium chips, and encourage adoption by making the platform cheaper. By supporting Itanium on the same motherboards used by Xeons, OEMs and manufacturers can take advantage of economies of scale to produce lower cost motherboards and save money by no longer needing to maintain a separate Itanium-specific motherboard line-up. At least that is what Intel hopes to accomplish with its Modular Development Model.
Of course, the flipside is that the changes will make it easier for Itanium customers to transition to x86-64 Xeon chips and migrate away from Itanium altogether. With Xeons being refreshed more often and continually integrating more of Itanium’s once-unique reliability and redundancy features, it certainly seems possible that Intel is writing itself into a corner with Itanium. As such, the age-old question remains: Is the writing finally on the wall for the pessimistically-nicknamed Itanic line of IA64 CPUs?
To Intel’s credit, despite the factors working against Itanium’s adoption, the company seems unconcerned. Intel has acknowledged that Itanium and Xeon chips will trade wins and leapfrog each other with new iterations, but it has reiterated that it is more committed to Itanium than ever before. The company does continue to have hardware support from HP, NEC, Bull, Inspur, and Huawei for servers, and software compatibility with HP-UNIX, Nonstop, and OpenVMS operating systems (and Oracle software whether they like it or not).
The market for Itanium systems is certainly niche, but it is profitable for Intel to maintain, and necessary for companies and researchers that demand precise calculations with redundancy, error detection, parallelism throughout, and configurations where cores can double or triple check calculations (which is important for engineering and aerospace research where lives may be at stake).
Itanium is also useful in processing transactions and doing so-called big data analysis on massive data sets. Right now, there continues to be a demand for Itanium (however small relative to x86-64 servers), and while the public Intel roadmap only accounts for one additional iteration, Intel is more likely to harness the processor socket convergence to continue support and development of Itanium than it is to use the Modular Development Model as a graceful exit from the market.
Itanium may be shrinking, but the Itanium ship is not going to sink any time soon.
One final note: You might also want to check out our piece on Xeon Phi here.