Processor Whispers - About Partnerships and Partner Chips
by Andreas Stiller
The IT economy is doing better and better, despite the renewed spitting – and now even cloudy twittering – of the Icelandic volcano Eyjafjallajökull. This positive development could very well lead to new partnerships and almost-forsaken projects like Intel’s Larrabee graphics chip might make a comeback.
Things look bad for the contemplated partnership between Intel and TSMC – Intel clients want products manufactured by Intel, says Intel boss Paul Otellini regarding the plans to have the contract manufacturer TSMC produce some of Intel’s chips. The partnership with Infineon – the planned purchase of Infineon’s mobile communications chip branch – is also unlikely to happen as the German corporation isn't showing any interest in the deal.
Infineon has just financially rehabilitated this branch, is looking at better times ahead and is expanding its capacity in Dresden. But the corporation still has to make amends for past sins and will probably have to pay a cartel fine of 300 million euros to the EU, together with other “conspirators” including Samsung, Elpida, NEC, Hynix and others. Only “traitor” Micron will likely get off cheaply – because of the leniency policy.
However, other interesting partnerships are in the pipeline, for instance between IBM and Nvidia. IBM has decided to offer the Dataplex servers dx360 M3 preconfigured with Nvidia’s Tesla cards M2050 – provided that Nvidia is able to deliver the Fermi chip offspring. “This is the first time that GPUs are part of a mainstream, high-volume product line from a Tier 1 OEM,” rejoiced Nvidia’s HPC boss Sumit Gupta.
Source: Nvidia Previously, Supermicro, Tyan and Appro had announced that they were going to equip their servers with M2050 cards – but when? The previously planned May launch of the GPU versions for High Performance Computing (HPC) has already been postponed until July. Although Nvidia’s last financial statements – with a sales volume of almost exactly one billion dollars (an increase of 51 percent in comparison to the previous year) and a profit of 128 million dollars (in contrast to a loss of 201 million dollars a year ago) – were very positive, the continuous delays as well as the (probably still big) problems with the production of the Fermi chips forced Nvidia to announce poor projections for the next quarter. On the next day, the stock price dropped by 11.5 percent.
At least some M2050 prototypes should find their way to the International Supercomputer Conference ISC10 in Hamburg that begins at the end of May. The stock price could quickly recover then, seeing how the whole HPC scene is so eagerly awaiting the new Tesla cards.
Numerous speeches on the application of GPUs are scheduled to form part of the scientific sessions of the ISC10, along with tutorials and discussions – primarily about Nvidia and CUDA. While Nvidia’s chances in the graphics sector don’t look too good because of the strong competition from ATI Radeon, Tesla is very interesting for HPC, especially once the new chips with the promising features are available: high performance for double precision floating point, ECC-protected memories, large memory capacities thanks to 64-bit addressing, concurrent kernels and more – a list of features that ATI’s current Stream line can’t really compete with. However, ATI isn’t idle and expects to become stronger in the HPC sector with its new architecture Northern Islands next year.
Also Intel is still in the game, with its Larrabee GPU; the Larrabee graphics chips are not dead, Otellini insisted at an analyst meeting. According to him, the number of developers eagerly working on the project is the same as six months ago, when the first version – intended for the desktop graphics market – was largely canned. Some survived however, and are destined for application on developer boards, which Intel will probably present at the Supercomputer Conference in New Orleans in November. With their support for the new instruction set for the 512-bit wide vector unit, these are supposed to pave the way for new Larrabee generations. They will first have to prove themselves in the HPC scene before Intel dares an attack on the graphics sector against the superiority of ATI and Nvidia. Initially, Intel plans to score with a shared memory model for CPU and GPU – a feature that Nvidia won’t be able to offer in the near future as it lacks a suitable CPU.
The PowerXCell processor was also meant to participate in this race, but IBM has taken it out of its roadmap. What’s more, IBM will stop supplying replacement parts for the Cell blade server QS21 at the end of June and the days of the newer QS22 with the PowerXCell 8i should soon be numbered, too.
Source: Mark Barnell / Mayer@wpafb.af.mil In April, Sony accelerated the death of Cell for HPC with the Playstation 3’s new operating system, which bans Linux as a boot option. A class action against this decision is already in place in California: PS3 owners demand that the full original functionality be restored. Sony's decision also affected the US Airforce, which runs more than 2000 PS3s in large clusters and is now bitterly complaining about the Linux removal, which makes life difficult for them when they replace defective systems. Some people are even saying that Sony is threatening American security...
ISC10 Life Online
Thanks to the University of Hamburg, everyone with an interest in the application of HPC in science and technology will be able to follow highlights like the keynote by Intel’s server boss Kirk Skaugen (“HPC Technology Scale-up&Scale-out”), the announcement of the new Top500 list of supercomputers and the “Hot Seat Sessions” – short 10 minute speeches by company representatives with subsequent interviews by “inquisitors” – online. The live stream will be available from the 31st of May at http://lecture2go.unihamburg.de/live.