PowerPC
 

Is anyone actively using MLucas on a PowerPC chip? I see Debian has an outdated package. I'm considering picking up an old PowerPC Mac to do verification runs.

[QUOTE=Mark Rose;475342]Is anyone actively using MLucas on a PowerPC chip? I see Debian has an outdated package. I'm considering picking up an old PowerPC Mac to do verification runs.[/QUOTE]

Remind me - what (if any) SIMD capabilites did the later PowerPC versions have, and did it support FMA in either the scalar-double or SIMD float64 arithmetic? And how would you estimate the float64 throughput (and power consumption) of those PowerPCs with the cheap ARMv8 versions of the Pi3 and Odroid C2? I'm using one of the latter for some slow-but-low-wattage DCs; Luigi (aka ET) is using a 5-times-quadcore version of same for similar. The 5-socket stack should be able to average close to one DC per week @2560K FFT.

[QUOTE=ewmayer;475347]Remind me - what (if any) SIMD capabilites did the later PowerPC versions have, and did it support FMA in either the scalar-double or SIMD float64 arithmetic? And how would you estimate the float64 throughput (and power consumption) of those PowerPCs with the cheap ARMv8 versions of the Pi3 and Odroid C2? I'm using one of the latter for some slow-but-low-wattage DCs; Luigi (aka ET) is using a 5-times-quadcore version of same for similar. The 5-socket stack should be able to average more than one DC per week @2560K FFT.[/QUOTE]

I'd hopefully be looking at a PowerPC 970MP system. Each CPU as two cores, and I'd ideally get a two CPU system. Each core has 1 MB of L2.

One of the AltiVec units per core supports FMA over 4 x 32-bit floating point vectors (with 32 registers) per clock. AltiVec doesn't support 64-bit operations, which have to be run on the main FPU units. Each of the two FPUs can execute a single 64-bit FMA per clock.

I understand the Pi3 give a single 64-bit FMA per core per clock, so 4 x 1.2 GHz or 4.8 G-fma-ops. At 2.5 GHz per PowerPC core, four cores would give 20 G-fma-flops (or more if there were a way to leverage 32-bit ops in parallel with the 64-bit ops).

The dual channel DDR2-400 (6.4 GBps theoretical) would probably be a bottleneck, but that is slightly more than a Pi3. I'm not sure how effective the 1MB of L2 would be. How memory starved is a Pi3?

The G5 series of CPUs were famous for being power hogs, so the Pi3 and C2 would handily win from an efficiency perspective.

Wikipedia:
[QUOTE]The 970MP is a dual-core derivative of the 970FX with clock speeds between 1.2 and 2.5 GHz, and a maximum power usage of 75 W at 1.8 GHz and [U]100 W at 2.0 GHz.[/U] Each core has 1 MB of [URL="https://en.wikipedia.org/wiki/L2_cache"]L2 cache[/URL], twice that of the 970FX. Like the 970FX, this chip was produced at the 90 nm process. When one of the cores is idle, it will enter a "doze" state and shut down.[URL="https://en.wikipedia.org/wiki/PowerPC_970#cite_note-6"][6][/URL] The 970MP also includes partitioning and virtualization features.[URL="https://en.wikipedia.org/wiki/PowerPC_970#cite_note-7"][7][/URL][URL="https://en.wikipedia.org/wiki/PowerPC_970#cite_note-8"][8][/URL][/QUOTE]
What speed range are you considering? What sort of cooling would the units you might consider have?

Is there some particular attraction to this architecture, or the OS it uses? I worked with Power PC Macs early in the digital imaging part of my photolab career. The last Mac Pro I used on a job had dual Xeons, though I don't remember core counts. For their times they were powerful workstations. I remember being blown away by installing [U][B]64 MB![/B][/U] of RAM in a newly arrived Power Mac.

[QUOTE=kladner;475357]Wikipedia:

What speed range are you considering? What sort of cooling would the units you might consider have?

Is there some particular attraction to this architecture, or the OS it uses? I worked with Power PC Macs early in the digital imaging part of my photolab career. The last Mac Pro I used on a job had dual Xeons, though I don't remember core counts. For their times they were powerful workstations. I remember being blown away by installing [U][B]64 MB![/B][/U] of RAM in a newly arrived Power Mac.[/QUOTE]

I was thinking a PowerPC Quad, the one that came with two 970MPs, the last, and most powerful PowerPC Mac made. I'm feeling the itch to have something different, but I loathe having equipment without an actual use haha.

The Xeons would be kind of pointless since I have much faster x86-64 hardware already.

They would be truly antique Xeons, besides.

I do understand wanting some variety in your projects. Hardware and OS change the scenery and the "customs." Kind of like voyaging to another time or place.

Well, that sounds like an intersting code-build miniproject, Mark - let us know how your "smart space heater" performs once you get it. (And also how much you paid - I would hope it would be quite cheap at this point.)

[QUOTE=ewmayer;475489]Well, that sounds like an intersting code-build miniproject, Mark - let us know how your "smart space heater" performs once you get it. (And also how much you paid - I would hope it would be quite cheap at this point.)[/QUOTE]

Still debating on it, and it would have to be for a cheap price.

[QUOTE=Mark Rose;475499]Still debating on it, and it would have to be for a cheap price.[/QUOTE]

I see several G5/970 systems [url=https://www.ebay.com/sch/items/?_nkw=powerpc+970&_sacat=&_ex_kw=&_mPrRngCbx=1&_udlo=&_udhi=&_sop=12&_fpos=&_fspt=1&_sadis=&LH_CAds=&rmvSB=true]for under $100 on eBay[/url], though the shipping price tends to overwhelm the system price here, since the G5 is large and heavy. eBay allows sellers to include a local-pickup option in their listings, if you can find a suitable 4sale system within easy driving distance from you, that would make the most sense.