Processor Whispers: About Parties and Party Poopers
by Andreas Stiller
That’s the way the cookie crumbles. Just when – after years of struggling with the Itanium – Intel is finally able to roll out the long-delayed product, the bad blue competition from Armonk spoiled the merry launch party.
Shocked, Intel didn’t present a single concrete benchmark result for their new 9300 line (Tukwila); no mention of a “World Record Performance” in a standard industry benchmark as former business boss Pat Gelsinger had promised (I’ll have to get in touch with him concerning our bet).
In the past, it was common to announce a record whenever a new server processor launched, as was the case with the Itanium series 9000 (Montecito and Montvale), although those world records primarily referred to the pure mass of processors and hard discs –7762 for TPC-C – rather than the performance of the processors. Broken down into four socket systems, like in the HP Integrity rx6600, the Montecito and Montvale also mostly fail to keep up with the competition from IBM and Sun and with Intel’s own Xeons.
What made Intel’s Tukwila launch even stranger was the total absence of complete systems. While before at least major partner Hewlett-Packard used to show off various Infinity servers and Superdomes and often released TPC or SAP-SD numbers beforehand on such occasions, now even HP only promised to get Tukwila servers ready within 90 days. Neither could Intel’s partners Bull, Hitachi, NEC, Supermicro and Inspur present anything tangible and former Itanium protagonists like SGI and Fujitsu are not even on the list of partners any more and Unisys withdrew from the Itanium business a year ago.
So, essentially, Intel’s manager of the Data Center Group, Kirk Skaugen, only advertised what Gelsinger had announced a long time ago: more than twice the performance of the predecessor Montvale, which – with double the number of cores – is no miracle, a six times faster memory performance due to a pair of integrated memory controllers with four memory channels (4.8 GT/s) and an up to nine times faster interconnect performance through four full and two half QuickPath links (4.8 GT/s). It’s only by those two additional half QP-Links, which make well connected eight socket systems possible, that the Tukwila is outwardly distinguishable from its colleague slated for the end of March, the Nehalem-EX. Otherwise, their infrastructure, which allows for low-priced systems, is the same (Boxboro chipset, DDR3 memory of up to 1 terabyte via a scalable memory interface).
The new family’s top model Itanium 9350 – manufactured in the reliable 65nm process – runs with an unhurried 1.73 GHz clock speed, features 24MB of L3 cache and consumes 185 watts TDP; a turbo mode can selectively increase the clock speed of single cores by a "Bin" (133MHz).
According to the fuzzy claim “more than twice the performance”, an Itanium 9350 four-processor system could maybe get some 205 points in the SPECint_2006_rate (peak) – about one fifth of what the weakest new Power7 system gets. However, Hewlett-Packard has also mentioned applications that are supposed to run up to nine times faster and HP’s general manager Martin Fink even went as far as saying that “old-established customers” would be able to take advantage of a 40 times higher performance. One might wonder why he didn’t use the many remaining DEC-PDP-11 users for his comparison – he could have boasted a performance multiplication by many thousands then.
And while Intel emphasised the reliability, long-time compatibility – the Tukwila successors Poulson and Kittson are supposed to be socket compatible and the software binary-compatible – and use in "mission-critical" applications, IBM mocked its competition in an uncommonly fierce manner. Obviously, IBM is rather proud of its new high performance 45nm 8-core processor with 3 to 4.14 GHz clock speed and quad multithreading. In contrast to its predecessor, Power6, IBM has now gone back to intelligent out-of-order technology again, which is why the clock rate is a bit lower. The 32MB L3 cache acts as eDRAM, which slightly slows down the access speed but saves transistors and energy. And the Power7 processors have a turbo-core-mode, too. Their cores can assign their cache and memory channels to other cores and go to sleep. The remaining processors can then take advantage of a higher clock speed, like with Intel’s new processors.
IBM is getting bolder concerning its pricing, too. The small systems 755 and 750 Express – with 4 sockets and 32 cores or 128 threads – start at $34,152 in the on-line shop, equipped with a 3.0GHz processor.
The Armonkers claim up to four times higher performance and an increased virtualisation capability in comparison to equally priced products from Sun/Oracle and Intel – the cherry on the cake being a three to four times higher energy efficiency. Through the pricing, they explicitly position the Power 750 Express against the HP-Integrity and SPARC-Enterprise servers in terms of performance and get considerably better numbers. At its website, under “Compare Unix Systems”, IBM is happily showing off big SPEC-CPU2006 numbers in great detail, the most impressive being the SPECint/fp_rate2006/watt – most Power7 systems surpass the competition’s numbers by a factor of four to seven.
As for larger systems with 32 and 64 cores, IBM is in the race with the Power systems 770 and 780, which – with 2530 SPECint_rate2006 and 2240 SPECfp_rate2006 – also relegate their competition (SPARC Enterprise 8000, HP Integrity rx8640).
Although the competing systems used for the comparisons are not as new – the Tukwila systems will surely do a little better and the yet unknown pricing will play an important role – IBM’s Power7 is a real hot potato right now. Who else is anxious to see the challenged competitors’ reactions – especially the one of Oracle, IBM’s favourite opponent?