[ars] IBM tries ARM-style licensing to reverse decline of Power CPUs

Quote:

Originally Posted by DaveLT

Nice info. But the best and fastest PPC procs always cost a bomb. Better they are but cost they aren't. In the end it's down to software support ... If their arch always completely changes and in comes instruction set changes, devs will soon drop them off

Click to expand...

Let's take a look at cost. Let's say that the cost to design your chip is $100,000,000 and the physical manufacturing cost is $100. If you make 100,000 chips (around what IBM probably makes), then each chip carries with it a cost $1100. Now, if you make 10,000,000 chips using that design, then your cost drops to only $110 per chip. This is why Intel and ARM have such an advantage in pricing, but IBM can eat a lot of development costs if they want to make the architecture popular again.

PowerPC is a very stable ISA and has changed less than ARM or x86 over the past 20+ years.

The largest argument is the software problem, but PowerPC is well supported by Linux. Additionally, ARM programs don't require many optimizations since the architectures have a lot of similarities (both being RISC).

The final point of interest here is that the POWER consortium already has a couple dozen big corporations on it's board. As much as I like the MIPS ISA, I think that POWER offers the best shot yet of providing a global, unified architecture on this basis alone.

edit: just a note, but POWER has been license-able for years. What's changed here is that IBM is licensing their proprietary chip design for use. Build your own is hard (ask Qualcomm, Nvidia, or Apple how much harder building is than buying from ARM). The ability to buy your way in is huge. You don't have the risk that you put millions into development and come up short. If you like the processor designs, you buy them and your only risk is that you can't market your final product (a risk that can't be eliminated anyway).

Quote:

Originally Posted by Alienware69

calling BS on that sorry lol. More powerful than AMD maybe but not Intel.

Click to expand...

http://www.anandtech.com/print/4285

btw, that benchmark was dated on the IBM side (and done on a 750 which uses a slower processor). My understanding is that the real numbers for that benchmark favor POWER7 by 25+%. In addition, POWER7 doesn't suffer from scaling issues like Xeon when you begin looking at systems with many more sockets. Also noteworthy that the POWER7 design is a much larger process and the design is a couple of years older.

A couple other interesting sources of performance data (benches of POWER compared to Intel are hard to find though the number of people in the industry who recognize that POWER7 is faster are not in similarly short supply)

http://rhaas.blogspot.com/2012/03/performance-and-scalability-on-ibm.html

http://eprints.dcs.warwick.ac.uk/637/1/PMBS.pdf (warning: pdf -- is a whitepaper comparing several server chips, only accounts for westmere and prior. Note also that it not only won in performance, but also in performance per watt)

Here's some specs on the original POWER7

567mm2 @ 45nm lithography
Transistors: 1.2 B

clockspeed: 4.2GHz
Eight processor cores
12 execution units per core -- 2 Fixed point, 2 Load/store, 4 DP floating point, 1 Vector, 1 Branch, 1 Condition Register , 1 Decimal

6-wide in-order dispatch

8-wide out-of-order issue

4 Way SMT per core
32 Threads per chip
256 KB L2 per core 32MB on chip eDRAM shared L3
Dual DDR3 Memory Controllers with 100 GB/s Memory bandwidth per chip
Scalability up to 32 Sockets
360 GB/s SMP bandwidth/chip
20,000 coherent operations in flight
Advanced pre-fetching Data and Instruction

energy-star rated

I don't have any SPARC benchmarks, but the new chips are even faster than the POWER7. It's telling that 25% of the fastest 100 supercomputers run POWER. Intel has to make the same design work for consumers and HPC while IBM only targets one market (and then there's the fact that IBM doesn't have old x86 baggage to deal with).

edit:

@crust_cheese

that's kind of true, but it's a little strange. I've read IBM's docs some and PowerPC isn't a strict subset. Sometimes the same instruction does slightly different things or they use different instructions to do the same thing.

@cdoublejj

IBM, Motorola, and Apple were the first companies to get together in the POWER consortium and worked jointly to create PowerPC as a more consumer-oriented (read: stripped down) POWER ISA. Motorola was supposed to make the G5 chip design for Apple. When they finished, the processor was blazing fast, but has a physical bug and the workaround blew the performance (like how the original Phenoms slowed way down when the TLB workaround was used). IBM's version of the G5 wasn't tuned for what Apple needed and Apple (probably Steve Jobs) thought IBM should be the company to pay for the R&D to which IBM said no as they wouldn't recover their money.