< All posts

Intel’s Xeon Phi Heat Transfer Dilemma

We’ve been spending all of last year around FPGA devices, but lately we spent a bit of time with Intel’s Xeon Phi coprocessor. What took 9,298 CPUs and occupied 72 server racks in 1997 now fits on a tiny little chip the size of an iPod Nano.

Intel showing off Xeon Phi aka Knights Corner

This new chip, formerly called Knights Corner, delivers 1 teraflop of double precision floating point performance. And what’s more, Intel promises we’ll see accurate 28 days weather forecasts (!) within the decade, something we couldn’t imagine just a week ago.

With a price tag of $2600, the Xeon Phi coprocessor is a HPC (high performance computing) product and not available for the every day gaming PC. While it was just released, it has already been installed in a couple of supercomputers around the world. Apparently Intel is also preparing to ship 100,000 units to China within 2013, where Tianhe-2, the world’s fastest supercomputer aiming 100 petaflops, will be built.

Xeon Phi Heatsink and HeatpipesHeatsink, heat pipes, fan

But not just supercomputers are making use of the Phi. Smaller installations, starting from just a few Xeon Phi CPUs up to a few dozens or hundreds can be used to simulate weather forecasts, do medical research (ie. cure for Alzheimer, cancer, etc) and process geological images (earthquake simulations) and more.

80% Heatsink – 20% Hardware

This all would be good news, if Intel wouldn’t be forced to literally surround their chips with outdated and inefficient technology most datacenters still use for cooling today: heat sinks, fans, air conditions and more fans. In fact, the majority of today’s hardware is built around the constraints of this truly outdated technology and this has to change.

While there has been a lot of development (pressure?) in recent years to make datacenters more efficient (greener), we can’t just all start building in Iceland and there is only so much you can do with cold aisle/hot aisle containment, with water pipes running to your CPUs or by submerging your hardware in pumped oil.

Hong Kong – average PUE 2.52 (it’s true)

The ultimate solution to this age old problem is passive 2-phase immersion cooling. These are not just empty words by a company in the business of immersion cooling, we do stand behind our claims and we are here to show just how exactly this can make a better product. Passive 2-phase immersion cooling takes the heat away more efficiently than other methods, uses much less energy for the task, and allows the engineers to focus on more important things than getting the heat out. And it costs much less to build and is easier to maintain. Truly.

Now that our new blog is online and that the site slowly starts to make sense, you can expect more information from our part of the world on this new exciting solution to the heat transfer challenge. Whenever we have a little bit of time, we’ll post right here with information how this technology can be implemented into your projects, no matter what size (ie. something the size of a server rack, or something that occupies whole rows in a data center).


Xeon Phi heat transfer design with lots of parts

80% Heatsink, 20% Hardware


Stampede at TACC, Air Condition Raised FloorStampede at TACC, first supercomputer using Xeon Phi


Stampede at TACC, rack, aircon, rack, aircon

A closer look reveals an aircon for every rack


The same is true for GPUs: Nvidia’s Tesla is more heatsink than GPU accelerator