[IBM Power] Any plans to support new IBM CPUs?

[GPU timpster]

Hello,

It seems the IBM Power 7+, and 8 CPUs would be a great way to get more power into FAH. I don't understand how they work, but wikipedia basically said they have the same floating point performance as a new Intel i7 chip. That's absurd, these systems have 96 cores so a statement like that just sounds wrong. I'm wondering if I could snag one for a couple grand, run linux on it, and see what happens when I run F@H on the CPU. Does anyone have a guess if it would work at all?

[Joe_H]

My educated guess is the answer is no. F@H is targeted towards widely available consumer and pro-sumer hardware that includes AMD and Intel CPU's and GPU's from nVidia and AMD. IBM currently only uses the POWER line of CPU's in mid to high range servers, so they are not widely available. PG would likely only see a few contributors from POWER based systems, so the effort to support them would not justify porting to that hardware.

Theoretically an AMD64 system could be run in emulation on a POWER based system and the existing client fold on that. But that would be up to the operator of that system. Or if POWER based systems became commonly available, an unlikely event at this point in time, PG could possibly leverage their past development of PowerPC code they did for Mac's about 10 years ago to create a native folding client. But in my opinion that is very unlikely to occur.

[gwildperson]

The name POWER has been around a long, long time. Is this an upgraded version of their server architecture from 20+ years ago? They have a history of selling solutions which involve proprietary hardware and proprietary software. They call anything they build a standard but get out of the market if it becomes a commodity that's priced for the mass market.

[7im]

Standard project policy is to watch all new developing technologies. FAH always "plans" to develop clients for hardware when that hardware becomes relevant to the "@home" market. Donors have been asking about folding on smart devices for many years. Same answer applied for many years. And now we actually have a phone client.

It's like picking the low hanging fruit. Most results with the least effort because Folding@home has limited resources. To rephrase what Joe_H said, the development cost has to be worth the results. And where to spend the limited resources to get the best results, improving an existing FAHCore, building a new FAHCore, or development a new client for a very small donor base, even if very powerful.

[Joe_H]

The name POWER has been around a long, long time. Is this an upgraded version of their server architecture from 20+ years ago?

Yes, these are an updated version of the POWER architecture. It has had a number of offshoots, PowerPC chips developed by the Apple, IBM & Motorola consortium, Cell processors that were used in the PS3, and others. Some of the features from the offshoots have been merged back into the main POWER line of CPU's. The chips mentioned in the OP are the most recent releases.

[JimboPalmer]

The Power8 is an impressive design for what it is intended to do. I am not sure it is intended to do scientific math.

"POWER8 is designed to be a massively multithreaded chip, with each of its cores capable of handling eight hardware threads simultaneously, for a total of 96 threads executed simultaneously on a 12-core chip."

This implies to me that it expects a great many 'slow' jobs (where slow means they access main memory, which is so slow that the CPU can productively change which program it is thinking about 8 times before main memory returns data)

It is not clear that there are 8 sets of SIMD registers per core (it reads as if there are only two) or that SIMD workload accesses main memory all that often (Many WUs work on video cards with only 512 meg ram, and the POWER8 chip may have 128 meg of cache before it goes to main memory. Fast enough caches (and there seem to be 4 layers of cache) and small enough datasets would reduce the number of times the 8 threads per CPU would need to switch.

The POWER chips excel at connectivity to multiple peripherals, but again F@H does not need entire banks of hard drives.

One future advantage of POWER is that you can design GPUs that use main memory directly without needing a PCI-E bus. If GPU vendors adopt this protocol, you could do excellent GPU folding.

https://en.wikipedia.org/wiki/POWER8

[gwildperson]

Yes, these are an updated version of the POWER architecture. It has had a number of offshoots, PowerPC chips developed by the Apple, IBM & Motorola consortium, Cell processors that were used in the PS3, and others. Some of the features from the offshoots have been merged back into the main POWER line of CPU's. The chips mentioned in the OP are the most recent releases.

And how many of us are running on those offshoots today? My impression is that the consortium made expensive chips for a limited market, avoiding the potential of becoming the kind of commodity that would make another porting of F@H worthwhile. Look where the Cell-PS3-(nonx86) Apple F@H clients are today

[Molasar]

That's a pity. I have 3 IBM Power 7 servers (from 32 to 128 cores) basically sitting in my home lab. Won't mind contributing some cpu horsepower to this project...