The semiannual update of the Top500 list of the world’s fastest HPL supercomputers was released at SC in Reno this week.
In possibly our longest report yet, InsideHPC takes a look at the phenomenon of the Top 500 list and the latest top 20 systems.
Updated in June and November of each year, the Top500 is now in its 30th edition. Whilst there are many arguments about the all round value of the list and the HPL benchmark it is based on, the Top500 plays a valuable role in the HPC industry, largely due to its consistency in tracking HPC industry trends over the last 15 years.
And in spite of its detractors [those without high ranking systems?] the HPL benchmark is representative of many applications that users run on these systems.
As ever, the vendors are scrambling to issue press releases (too many to link to here) each claiming their leadership of the list by whatever metric they can find that supports their case (number of systems in list, number of systems in top 10 (or top 20/50/whatever), aggregate FLOPS, top system, etc). Customer sites that have ranked high likewise make sure we all know how high up the list they are, how they are leading their country, industry sector, etc. This mutual scramble shows the importance the industry places on the list, and how useful it has become to track what is going on.
The scramble is closely followed by a wave of ‘politically correct’ utterings along the lines of ‘HPL doesn’t measure everything about a system’, ‘we must focus on the users/science/capability jobs’, ‘we must avoid buying systems just to be high up the list’. Weaker vendors, frankly sounding a bit pious, assure us they ‘are focused on next 50,000 systems not the top 50’. Even the strong vendors aren’t averse to the odd bit of complete daftness, such as IBM/LANL touting ‘next year’s system, Roadrunner, will smash the top of the list’ or some such.
Some analysis of the latest list is available from the Top500 people here.
In number one spot (since 2004) is the IBM BG/L at LLNL, now upgraded to 596TFLOPS peak (478 TFLOPS HPL Rmax). Europe’s fastest supercomputer (a BG/P system at Julich, which we discussed earlier this week) takes 2nd spot. A total of 4 BlueGene’s can be found in the top 10 (positions 8 & 10), and many more as your move down the list, showing the continuing march of this leading supercomputer design.
The SGI Altix ICE supercomputer in New Mexico (details from our earlier story here) takes 3rd spot. It is good to see SGI back at the top after their hard time of late, and showing (as they did with the Columbia system) that they are among the best at deploying large scale systems with minimal fuss. I speak from personal experience that this applies equally at the smaller scale too.
Perhaps the most surprising entries in the top 10 are the next two at 4 & 5. Why surprising? Well, first they are HP system, and HP have been noticeably absent from the high end for some time and indeed weren’t giving much evidence to potential customers that this would change. However, over recent months HP have been making noises towards being serious about supercomputing again and have followed up with these two systems, and another at PNNL (which we reported on earlier but didn’t make this list). HP’s return will be welcomed if it is sustained, as they hold a reputation for quality of build and reliability. Further notability about these two systems surrounds where they are: India, showing the much anticipated emergence of India as a serious supercomputing nation; and Sweden (not traditionally associated with the top 10) with a classified system (apart form the nuclear systems, classified systems tend not to make it to the heights of the Top500).
Cray completes the top 10 vendors with XT4/Red Storm MPP systems at positions 6 (Sandia), 7 (ORNL) and 9 (NERSC – our report here). This is both good and bad for Cray – dropping down to 6th is a significant fall from previous years for Cray, but with 3 systems in the top 10, Cray continues to show they are focused on the high end with among the most scalable HPC products available.
Looking further down the list I see the UK’s HECToR Cray XT4 system (which I have a personal good wish for) entering at number 17 with a score of 54 TFLOPS. This capability class service, led from the University of Edinburgh, has funding through to 2013, with the next upgrade to quarter of a PFLOPS in 2009 already planned. The Top500 analysis goes on to show that the UK has retained its European leadership in HPC with 48 systems listed (see what I mean about reporting metrics about leadership!).
Intel’s much slighted Itanium2 processor shows its presence with 3 systems in the top 20 (15 – SGI Altix15 at LRZ, 19 – Bull/Quadrics at CEA, and 20 – SGI Altix Columbia at NASA).
One final statistic to leave you with, getting a top 10 system this year meant you had to set your sights for 100TFLOPS peak. Looking at what is coming up over the next year, a Nov 2008 top 10 spot could cost you half a Petaflops.