IBM Offers POWER Technology for Licensing, Forms OpenPOWER Consortium

29 Comments

View All Comments

• alpha754293 - Thursday, August 8, 2013 - link

  Well, one of the things that POWER is known for is their compute ability. A decent POWER5 processor can probably still do donuts around ANY x86 processor in Monte Carlo simulations because the chip itself at the macro level is pretty much a Monte Carlo processor. Having said that however, it was also the trademark that they were quite woefully and semi-ironically power inefficient; which in today's world of rising energy costs, is just one of the many nails in the coffin.
  
  By opening up the architecture and development of the hardware, perhaps you can see a lower power version of it, but we will have to see how that goes.
  
  The other benefit that POWER brings is their mini-POWER processors (the PowerPC, which powered Macs and a bunch of offshoots), and their baby-POWER processors in the form of the Bluegene/L/P/Q series. Those are largely used in MPP installations where the lack of core speed is compensated for by just having a TON of them, with the advantage being VERY power efficient (the opposite end of the spectrum).
  
  And BTW, bank mainframes still use IBM POWER systems a LOT because at 4 FLOPs/Hz, nothing has been able to beat that. And IA-64 resembles a copy of the POWER architecture, with poorer execution and implementation (plus general lack of support from hardware and software vendors alike).

• Kevin G - Wednesday, August 7, 2013 - link

  Intel does not offer custom x86 chips. If you want an x86 SoC you can either go to AMD who will gladly craft a custom chip for you if pay them or get what every SoC Intel is offerig everyone else because that is what Intel thinks should be on an SoC.
  
  ARM's advantage has been customization. You could license a CPU core from them, go to another vendor for the memory controller, USB controller etc., add a bit of your own logic and then choose where you'd like for it to be manufactured. Each step comes at a price for licensing but it is fairly open.
  
  The odd thing is that most people are seeing this announcement from IBM as attempting to emulate ARM's business model. This is incorrect as IBM already offers up embedded PowerPC cores designs in a similar fashion. The news worthy change is that IBM is opening up their ultra high end POWER cores in this model.

• dgingeri - Wednesday, August 7, 2013 - link

  It looks like nobody has bothered to actually answer the question. POWER has a few advantages over x86, but most are only useful in the server realm.
  
  1. the architecture is designed around accuracy. While the SSE instructions in X86 have reduced the accuracy of PC calculations today, POWER was designed to calculate them out to many more significant digits. This is useful for many types of scientific research and banking, not so much for running Excel or games.
  
  2. POWER has many ECC routines built into every stage of the architecture, including the instruction set. Because of this, data corruption and system crashing happens far less often.
  
  3. POWER was built from the ground up for virtualization, instead of being an addon like X86. This makes it much more efficient when running virtual machines.
  
  These things could be useful in a PC environment, but not really game changers.

• Kevin G - Wednesday, August 7, 2013 - link

  1. SSE and the binary floating point units inside POWER both adhere to the IEEE 754 spec. POWER6 and POWER7 do offer a binary decimal unit which do offer greater precision but require applications to be rewritten to use the new instructions and data types.
  
  2. It is true that every data path in IBM's power lines have ECC support, the instruction set itself inherently is not ECC protected. Thus the instructions moving around inside a modern POWER chip are ECC protect, that can end once they move off chip.
  
  Data corruption and system crashing related to errors encountered inside the CPU is surprisingly rare when systems are run within spec. Memory chips are where most of the corruption originates. End to end ECC protection does offer better detection as integrity is maintain throughout the design.
  
  3. Actually virtualization is also an add-on to the POWER architecture. Hardware virtualization capabilities didn't arrive until the POWER4 and really didn't hit its prime until POWER5. A lot of POWER's virtualization capabilities were ported or adapted from IBM's mainframe efforts which have had decades of polish.

• Brutalizer - Wednesday, August 7, 2013 - link

  The POWER7 was a good cpu when it was released some years ago, but now there are faster cpus. x86 is faster in some benchmarks, at a much lower price. Oracle SPARC T5 is way faster and holds several world records, but it is a newer cpu so that is to be expected.
  
  The main advantage of Unix cpus such as IBM POWER7 is the RAS. What costs is not performance, but reliability. RAS is very difficult to get, and very expensive. Intel Xeon is getting some RAS features now, but they can not compete with the RAS from traditional Unix cpus such as IBM POWER, Oracle SPARC and Itanium. This is the reason a IBM Mainframe is very expensive, RAS costs. Speed doesnt. Once a server is fast enough, you want it to be reliable in the Enterprise market. Downtime must be avoided.

• DanNeely - Wednesday, August 7, 2013 - link

  In their highest end Xeon variant, IIRC Intel has added most of the RAS features that were previously only available with the Itanium. Although they continue to insist otherwise, I expect the remainder are planned to be added soon at which point the Itanic will be allowed to sink.

• Henriok - Friday, August 9, 2013 - link

  Power Architecture processors comes in a very wide variety of flavors. From <50 MHz, sans-FPU and MMU parts that consumes milliwatts (Freescale's e200z0 cores), to 18 core monster-SoCs with powerful FPUs and integrated communications fabrics (BlueGene/Q) and everything in between like customized gaming CPUs with integrated GPU (XCGPU in Xbox 360 S), LTE base stations on a chip (QorIQ B4860) or the most powerful server CPU available with massive amounts of embedded DRAM and enormous amounts of IO (POWER7).
  The cores can be tiny single threaded in order things or four way our of order, multithreaded with multiple memory controllers, cryptography accelerators, special media accelerators, DSPs, floating point engines, SIMD-egines, GPUs, multiple levels of cache, integrated USB, Ethernet and other IO, one core or 18 cores, the same core type or different, run a single operating system or many asynchronously.
  All these are binary compatible.
  All these features are specified under the same living specification, governed by a broad spectrum consortium.
  
  Many cores are available in hard and soft macros, freely licensable from a variety of sources like Synopys, LSI as well as Freescale and IBM. These designs have a veriety of cores to chose from for their bespoke SoCs. Like the e200 and e500 from Freescale and 400 family (405, 440, 460, 470) from IBM, Applied Micro, LSI and Synopsys.
  And these designs can be manufactured by a fab of your choice: IBM, Global Foundries, TSMC, UMC or your own.
  
  Freescale offers a guarantee to manufacture certain designs for 20 years or more. That's a prerequisite to sell parts to certain projects with a long expected life span, like military equipment such as the F-35.
  Many designs like the PowerPC 603e, 440, 750, e500, and 7400 are available in radiation hardened versions for use in space or other extreme environments. Manufactured by several companies, like BAE, Space Micro, Broad Reach, Atmel and Honeywell.
  
  AMD and Intel offers _nothing_ like this. Their offerings doesn't even compare, but on just some special cases.
  ARM and MIPS does compare, but there's nothing from either that can perform as powerfully as Power can. The FPUs, SIMD-engines, frequencies, number of cores available to those who design custom processors based on Power outperform anything either ARM or MIPS can up bring. ARM is least 10 years behind when power consumption isn't the primary constraint of the design.
  SPARC is comparable too, but doesn't offer the range, nor the customization options or choice of designer or manufacturer.
  
  And now, IBM offers their future POWER8 as a base for a new breed of ultra high end customized processors. I can see supercomputer-SoCs with 8x eight way multithreaded POWER8 cores, integrated Tesla GPUs, interconnect intra-CPU fabric from Mellanox, and PCIe for direct attached super fast SSDs.. on one chip. That'd be a super dense supercomputer of which we've never seen. 10-50 times as performance dense as the best design today.
  Can't do that on any other platform.
  I can see Google and Facebook making a series of highly customized server processors with only what they need, and nothing else. Many high speed simple cores, integrated 10 Gbit Ethernet on chip, DDR-controllers, nothing else. They have an enormous amount of servers so they can afford customized hardware on this level just to make their modules cheaper to manufacture and run.
  
  For super computers or the likes of Facebook and Google, the ISA doesn't matter. They are not running off the shelf software anyway, so a recompile to another architecture (than x86) doesn't matter.

• wumpus - Thursday, August 15, 2013 - link

  "the ISA doesn't matter". Fine, you just change the arch selection on the makefile and re-compile major enterprise software and deploy it. I'm sure everybody bothered to make sure all the byte accesses would work just as well on big endian machines as they did on the little endian systems they were written fore.
  
  A more accurate statement is that the costs to change ISA are fixed (and relatively small) while the costs to buy hardware keep increasing (and are significant). Also, Intel makes a great deal of money selling to those us who do care about ISA (even if we hate ours), so they include a cost that these companies aren't interested in.
  
  Finally, I can't see this as anything but a surrender. IBM lost control of the PC world when they opened it (and it ate the mainframe's lunch within the decade, nearly killing IBM). IBM isn't opening this up because they want competition. IBM is opening it up because they can no longer afford the infrastructure to keep making POWER. My guess is that they won't share enough of the profit to convince others to pay those costs either (how many licenses has Nvidia bragged about lately?).

• loa_dir - Wednesday, August 14, 2013 - link

  One of the advantages is performance. POWER chips can have about twice the performance per core in throughput-oriented server applications. In HPC and other FPU-intensive applications the advantage can be even higher.
  
  The POWER chips powerefficiency seems to be quite good also, at least if you look on the green500 list (the most powerefficent supercomputers in the world). IBM occupies quite a lot of the top spaces.

• jjj - Wednesday, August 7, 2013 - link

  Wonder what this means for Project Denver ,if anything.