Advanced Computing in the Age of AI | Friday, March 29, 2024

IBM, Google Show Off Power8 Systems And OpenPower Efforts 

IBM was planning to do the announcement of the first bunch of Power8-based servers at next week's Impact2014 event in Las Vegas, but the Power Systems division at Big Blue and the OpenPower Foundation that is steering the opening up of Power8 technology decided to reveal some of the feeds and speeds of the new systems ahead of the event next week.

IBM has created five of its own machines based on the Power8 chips with variants targeted at specific operating systems and workloads. At the Open Innovation Summit event in San Francisco this morning, which was hosted by the OpenPower Foundation that Big Blue created last year to open up and promote Power chip and system technology, the foundation showed off the first reference system board based on Power8 chips as well as projects that will ultimately bring GPU, FPGA, and network acceleration to Power8 systems.

Thus far, 26 companies have now joined the OpenPower Foundation with hyperscale datacenter operator Google being the most interesting one. Google is secretive about its plans for Power8 processors but as EnterpriseTech previously reported, the company has been testing Power8 systems of its own design for the past several months. Nvidia is also working to more tightly integrate its Tesla GPU coprocessors with Power8 processors and IBM has designated Teslas as the GPU accelerator of choice for Power8 iron and the two companies are working on a framework for accelerating Java applications. FPGA makers Xilinx and Altera are also key OpenPower partners and they are working with IBM to accelerate memcached caching workloads for hyperscale web applications and Monte Carlo simulations at financial firms using FPGAs working in conjunction with Power8 chips.

Google did not show off whatever Power8 machines it has created to test its software stack but Gordon MacKean, engineering director for the platforms group at the company and currently chairman of the OpenPower Foundation, did talk a bit about why Google was interested in Power-based machines as an alternative to X86 systems. And he did present a reference system (presumably in use at Google for its testing) that was developed by the foundation.

"We came to the conclusion that to make a big impact in the scale-out server space – warehouse-scale computing – that there were going to need to be some big changes at the system level," MacKean explained. "Optimizing the interfaces between the components was really how we were really going to affect the performance per dollar and TCO of these systems. And as a group, we decided that we needed to create an environment that would allow us to work at the device level and at the system level to try to optimize these systems."

The plan for OpenPower Foundation is to work on processors and system-on-chip designs as well as interfaces for accelerators and networking and the accelerators are a big part of the story. "Google has always considered itself an innovation leader in platform technology," MacKean added. "We have built our datacenters around that innovation and we have always pushed innovation at the platform level. We looked at OpenPower as an opportunity to launch a third generation of warehouse-scale computing, to break down the barriers that exist between the components – the whole is really greater than the sum of its parts – and to really fix those bottlenecks that exist between the different components of the system: networking, memory subsystems, I/O subsystems, storage subsystems. We are looking to advance that, and we believe that we are technology leaders to steer that."

MacKean said that Google was going through a "pretty detailed investigation" of the suitability of the Power architecture to run the search giant's applications. As part of that process, Google is helping to develop and harden various bits of code in the Power code base and development tools.

The very first thing that the foundation had to do was create a reference server board that OpenPower Foundation members could use to build and test software. MacKean admitted that it looked a bit like "something you would get at Fry's," referring to the famous electronics chain headquartered in San Jose. MacKean said the reference board would have a lower cost than developers and system builders were used to seeing for something with as much performance as this board packs.

openpower-tyan-reference

Chuck Bartlett, worldwide technical support director for Tyan, said at the summit that this ATX form factor motherboard will be sold to software developers and to whitebox server makers, hyperscale datacenter operators who design their own systems, as well as to IBM for future systems. Moreover, IBM and Tyan are donating the specifications of the reference system to the OpenPower Foundation and members will have access to it and be able to use it to make customizations of their own.

In parallel with the discussion of the OpenPower Foundation and the participation of its key members, IBM is also releasing its first systems based on the Power8 chip, which the company first started talking about last summer. As we told you it would, IBM is starting the Power8 roll out with entry systems and is focusing on scale-out workloads rather than on scale-up NUMA systems. The Power8 chip has nearly 4 billion transistors and is manufactured using IBM's 22 nanometer processes at its plant in East Fishkill, New York. The chip has twelve cores on a single die, just like the top-end "Ivy Bridge-EP" Xeon E5-2600 v2 chip from Intel, but it sports a lot more memory bandwidth and I/O bandwidth. Each socket on a Power8 system can deliver 230 GB/sec of sustained memory bandwidth (more than three times a Xeon E5 chip) and 96 GB/sec of peak I/O bandwidth. Each chip has 96 MB of L3 cach, and on large NUMA configurations the servers are equipped with a maximum of 128 MB of L4 cache that sits between memory cards and the processors to stage data. The Power8 chips have two memory controllers that support either current DDR3 or future DDR4 memory and two PCI-Express 3.0 controllers for peripherals.

Interestingly, the PCI controllers have an overlay called the Coherent Accelerator Processor Interface, or CAPI, which creates a virtual memory space combining the main memory from the CPU main memory and any local memory used GPU, FPGA, DSP, and other accelerators plugged into the system. This should simplify hybrid system computing and speed up the performance of accelerated systems. Right now, data has to be moved back and forth between accelerators and processors (be they X86, Power, or otherwise) and this is a hassle.

The Power8 chips can run at between 2.5 GHz and 5.5 GHz, based on tests that IBM has done, and early benchmark results that pit the Power8 against its predecessors, the eight-core Power7 and Power7+ chips, were run at a 4 GHz baseline clock speed in the labs. The lab environment is not, of course, a production environment and as is the case with all processor manufacturers, the initial Power8 chips are running a little slower and many of the variants that will start shipping soon in IBM's own Power Systems rack machines do not have anywhere near their full complement of cores activated. It is early in the yield curve for the Power8 chips and as IBM gains experience with its 22 nanometer processes it will get chips that run at higher clock speeds and with more cores – and very likely at a lower cost as well, thus increasing its profits on the chips and systems it sells.

Right now, IBM is the only supplier of Power8 chips but Suzhou PowerCore, an offshoot of Chinese chip maker China Core that was established back in January to make its own Power8 variants for servers, storage, and other datacenter gear, has licensed the Power8 chip designs through the OpenPower Foundation and is working on its own variants. Motherboard maker Tyan is creating its own system boards, which other server makers will be able to buy to create Power8-based machines, and it is likely that a few more will do the same if the use of Power machinery expands in the datacenter as IBM hopes it will. If a few hyperscale datacenter operators adopt Power8 machines for a select bunch of workloads (any job where high memory or I/O bandwidth is key) then the Power server business could be radically expanded. IBM will also be touting the per-core performance of the machines, too. Steve Sibley, director of worldwide product management for the Power Systems line at IBM, tells EnterpriseTech that a twelve-core Power8 chip can do roughly twice the work as a Xeon E5-2600 v2 with a dozen cores.

Up until now, IBM has sold billions of dollars per year in Power Systems machines, mostly low-end machines for customers using its proprietary IBM i operating system and high-end NUMA machines running AIX underneath databases and enterprise applications. IBM has had some limited success peddling Linux-based machines, too, after offering X86-style pricing on servers that can only run Linux.

IBM is not providing detailed specifications and pricing for its new Power Systems S-Class servers, as the machines are called, until next week. But the company did give EnterpriseTech a sneak peek at some of the salient characteristics of the machines.

The two most important machines for extreme-scale customers that IBM is previewing today are the Power S812L and the Power S822L. Both of these machines have a BIOS setting that only lets them run a Linux operating system and they cannot run IBM's own AIX or IBM i variants. Specifically, these systems can run Red Hat Enterprise Linux 6.5, SUSE Linux Enterprise Server 11 SP3, and Canonical Ubuntu Server 14.04 LTS. The addition of Ubuntu Server is new with the Power8 servers and it was added to the support matrix for a number of reasons, the most obvious being that Ubuntu Server is getting uptake among large enterprises and cloud operators, particularly for those setting up OpenStack-based clouds. These two Linux-only Power8 machines have two processor options: a ten-core chip running at 3.42 GHz and a twelve-core chip running at 3.02 GHz. The processor cards used in the machines support 512 GB per socket (one processor is on one card, just like in the Power7 and Power7+ systems). The machines have room for a dozen disk or flash drives in 2.5-inch form factors and the two socket Power S822L machine allows for up to an additional six 1.8-inch solid state drives to be added to the machine. The base processor card has six PCI-Express 3.0 slots and the second processor card adds another three PCI-Express slots for a total of nine.

Tom Rosamilia, senior vice president in charge of IBM's System and Technology Group, which makes chips, servers, and storage, and Doug Balog, general manager of the Power Systems division, unveiled a rack of Power S-Class machines:

openpower-ibm-servers

Interestingly, Rosamilia said that the founding members of the OpenPower Foundation had input into the clock speeds and energy levels that were made available in the machines, which had already been largely designed by the time the effort was started last year. Going forward, Rosamilia said that OpenPower Foundation members would have even more input into future Power chip and system designs.

The dual-socket Power S822L will be available on June 10, while the single-socket Power S812L will ship in August. Both machines will not only support IBM's own PowerVM server virtualization hypervisor but also its PowerKVM variant. The other three machines in the Power Systems S-Class line will not be able to run PowerKVM or Ubuntu Server.

The Power S822 machine might also interest some scale-out cluster builders. This machine has Power8 processors with fewer cores that run at higher clock speeds. This includes a six-core Power8 running at 3.89 GHz or an eight-core chip spinning at 3.42 GHz. As with the Linux-only machines, the Power S822 has up to 512 GB per socket for memory capacity and offer the same storage and peripheral expansion. This box can run RHEL and SLES as well as IBM's own AIX but it cannot run IBM i.

The two remaining machines run AIX, RHEL or SLES, or IBM i and they come in much larger 4U enclosures. The Power S814 comes with a single processor, which can be a six-core chip running at 3.02 GHz or an eight-core chip running at 3.72 GHz. This box has up to 512 GB of memory per socket, seven PCI-Express 3.0 slots, and room for a dozen disk or flash dives. The Power S824 has two processors including a six-core chip running at 3.89 GHz, an eight-core Power8 running at 4.15 GHz, and a twelve-core chip running at 3.52 GHz.

EnterpriseAI