mersenneforum.org  

Go Back   mersenneforum.org > Great Internet Mersenne Prime Search > Hardware

Reply
 
Thread Tools
Old 2005-10-20, 00:45   #1
Peter Nelson
 
Peter Nelson's Avatar
 
Oct 2004

232 Posts
Default Dualcore mac G5 machines from Apple

Don't know if anyone noticed this too,

but Apple have launched some G5 machines using dualcores.

This means they have a QUAD core G5 system available.

Also it supports the modern PCI Express interface standard.

Not suitable for prime95 but could be good for similar priming or math programs.

The dual processor G5 was always pretty good at FP math for clustering etc. Now it seems they have twice that power in a box. They claim a linpack benchmark of 21 GFlops.

I know the plan is to switch to Intel so was quite surprised by this launch, but they look quite nice machines.

Applestore dualcore pricing:
Dualcore 2GHz $1999
Dualcore 2.3GHz $2499
Quad (two dualcore) $3299

These machines use DDR2 and PCI-E and come with Nvidia graphics.
Peter Nelson is offline   Reply With Quote
Old 2005-10-20, 03:03   #2
rogue
 
rogue's Avatar
 
"Mark"
Apr 2003
Between here and the

11000100001012 Posts
Default

They would be great candidates for the PIES project. They could also run a multi-threaded build of Glucas.
rogue is offline   Reply With Quote
Old 2005-10-20, 08:44   #3
Peter Nelson
 
Peter Nelson's Avatar
 
Oct 2004

232 Posts
Default

Yes, LL testing using Glucas multithreaded was exactly the kind of thing I thought people here might like one for.

Peter Nelson is offline   Reply With Quote
Old 2005-10-20, 10:27   #4
Greenbank
 
Greenbank's Avatar
 
Jul 2005

1100000102 Posts
Default

Oh my.

Must resist temptation.
Must resist temptation.

*looks at bank balance*

Temptation resisted.

I am waiting, however, for a motherboard with Dual S939 sockets. I like the idea of a quad Athlon X2 4400+. Mmmm.
Greenbank is offline   Reply With Quote
Old 2005-10-20, 13:44   #5
fatphil
 
fatphil's Avatar
 
May 2003

3×7×11 Posts
Default

Quote:
Originally Posted by rogue
They would be great candidates for the PIES project. They could also run a multi-threaded build of Glucas.
<nod/>

Although since last night (yawn, couldn't stop once I'd started) the Athlons are homing in on the G5's speed-wise, but you still have 15-20% edge. I've been told not to come to work tomorrow, so I'll look at the RISC-oriented transforms then, and over the weekend. Of course, I have no way of testing them for speed.

Does anyone know if dual-cores can just be slotted into the place of the old single-core CPUs? If I were to become attached to this G5, I might want something with more grunt than just 2*2GHz. I'm not willing to spend money on the whole darn box.

Phil
fatphil is offline   Reply With Quote
Old 2005-10-20, 18:34   #6
Peter Nelson
 
Peter Nelson's Avatar
 
Oct 2004

232 Posts
Default

Fatphil:

I dont think these dualcore processors can be dropped into a single core machine because as I understand it their interface to memory is different (although I could be wrong I don't think you will get hold of these chips easily). Also the old firmware is unlikely to support them as-is, even though the OSX is designed to support multicores/multithreading.

Greenbank:

It is not in AMD's plans to produce motherboards with more than one S939. The chips in this socket are NOT designed for SMP only one processor (single or dualcore).

For dual chips on a board you need Opteron 2xx (now available in single or dualcore) which are currently on S940.

In 2006 all AMD sockets are changing, but the Opteron 2xx and equivalents will be on a NEW socket F with more pins (1207 pins). The new (incompatible) 940 pin socket (M2) will be for single chips (ie AMD64 and Opteron 1xx equivalent) only according to my information.

Opteron is a great architecture, but I am holding off until they bring out the nextgen sockets (which will be DDR2 NOT current DDR).

The only way to tell if dualcore Xeon/P4 or dualcore AMD64 or dualcore Opteron or dualcore G5 is faster for a given application would be to benchmark it although we can make some rough guesses. I suspect Opteron and G5 might be similar but the winner for a given app depends on how the app is coded and the nature of the calcs required in it (how well they fit the architecture or are optimised for).
Peter Nelson is offline   Reply With Quote
Old 2005-10-20, 21:07   #7
fatphil
 
fatphil's Avatar
 
May 2003

3×7×11 Posts
Default

Quote:
Originally Posted by Peter Nelson
I dont think these dualcore processors can be dropped into a single core machine ...
Sloblock.

Quote:
Originally Posted by Peter Nelson
The only way to tell if dualcore Xeon/P4 or dualcore AMD64 or dualcore Opteron or dualcore G5 is faster for a given application would be to benchmark it although we can make some rough guesses. I suspect Opteron and G5 might be similar but the winner for a given app depends on how the app is coded and the nature of the calcs required in it (how well they fit the architecture or are optimised for).
My money's on:
Big bignums - Ath64. More bits of multiplication per unit time with SSE2, and with a fair amount of bandwidth to actually keep the cache full of the data you want to use next.
Small bignums - G5. Sod Karatsuba - lets just hammer muladds out at an *insane* rate. I don't expect Altivec to play a part in this - too narrow.

Alpha always used to be "the architecture of choice for number theory" (EM, and probably RobH, and myself too - refering to small bignums), and I suspect that G5 is the thing most likely to supplant it, as much of the design philosophy is the same.
fatphil is offline   Reply With Quote
Old 2005-10-20, 22:31   #8
rogue
 
rogue's Avatar
 
"Mark"
Apr 2003
Between here and the

142058 Posts
Default

Quote:
Originally Posted by fatphil
Alpha always used to be "the architecture of choice for number theory" (EM, and probably RobH, and myself too - refering to small bignums), and I suspect that G5 is the thing most likely to supplant it, as much of the design philosophy is the same.
I don't know about that. I would like to see how fast the new G5s are compared to the pre-existing crop. The ones I have bench around 5000 (with GMP-Bench) as I've coded a faster addmul.

The problem with PPC is that you need two 64x64 multiplies to get a 128-bit product and as for division, there is no 128-bit/64-bit division. IIUC, Alpha has a single instruction to get the 128-bit product. I don't know about the division, but I would suspect it has a 128-bit/64-bit division instruction.
rogue is offline   Reply With Quote
Old 2005-10-20, 23:00   #9
moo
 
moo's Avatar
 
Jul 2004
Nowhere

32916 Posts
Default

**goes to bestbuy with cd with testing program finds a new duel duel core mac, insersts disk, runs stress test and stores on thumbdrive**
moo is offline   Reply With Quote
Old 2005-10-20, 23:05   #10
ewmayer
2ω=0
 
ewmayer's Avatar
 
Sep 2002
Rep├║blica de California

3·53·31 Posts
Default

Quote:
Originally Posted by rogue
The problem with PPC is that you need two 64x64 multiplies to get a 128-bit product and as for division, there is no 128-bit/64-bit division. IIUC, Alpha has a single instruction to get the 128-bit product. I don't know about the division, but I would suspect it has a 128-bit/64-bit division instruction.
Alpha, Itanium and G5 all have a similar 64-bit multiply model in that they have 2 separate instructions to obtain the lower and upper halves of a 128-bit product of 64-bit integer inputs. On Alpha (ev6 and later) and Itanium these are fully pipelined, i.e. one MUL instructions can start every cycle (and thus we need 2 cycles to obtain a 128-bit product in pipelined fashion); on the G5 two 64-bit muls can start every 6 clocks. AFAIK Alpha has no hardware integer divide, as this can be emulated via simpler operations.
ewmayer is offline   Reply With Quote
Old 2005-10-20, 23:25   #11
rogue
 
rogue's Avatar
 
"Mark"
Apr 2003
Between here and the

142058 Posts
Default

Quote:
Originally Posted by ewmayer
Alpha, Itanium and G5 all have a similar 64-bit multiply model in that they have 2 separate instructions to obtain the lower and upper halves of a 128-bit product of 64-bit integer inputs. On Alpha (ev6 and later) and Itanium these are fully pipelined, i.e. one MUL instructions can start every cycle (and thus we need 2 cycles to obtain a 128-bit product in pipelined fashion); on the G5 two 64-bit muls can start every 6 clocks. AFAIK Alpha has no hardware integer divide, as this can be emulated via simpler operations.
Many thanks for that clarification on Alpha. I just looked in the PPC manual. You are correct on the multiply. The divide is a killer, 1 per 67 cycles.
rogue is offline   Reply With Quote
Reply

Thread Tools


Similar Threads
Thread Thread Starter Forum Replies Last Post
Apple Moving to ARM CPUs? ewmayer Mlucas 13 2018-04-05 21:16
Could Apple become the new Windows? jasong jasong 2 2012-12-07 05:57
New to Apple; New to GIMPS Unregistered Information & Answers 4 2009-03-16 13:10
Prime95/LLR on dualcore CPU - strange thing? Cruelty Software 1 2006-07-27 04:53
64-bit GMP-ECM on Apple G5/OS X v10.4 PBMcL GMP-ECM 5 2005-06-04 06:12

All times are UTC. The time now is 04:18.

Tue Apr 13 04:18:30 UTC 2021 up 4 days, 22:59, 1 user, load averages: 1.91, 1.95, 1.88

Powered by vBulletin® Version 3.8.11
Copyright ©2000 - 2021, Jelsoft Enterprises Ltd.

This forum has received and complied with 0 (zero) government requests for information.

Permission is granted to copy, distribute and/or modify this document under the terms of the GNU Free Documentation License, Version 1.2 or any later version published by the Free Software Foundation.
A copy of the license is included in the FAQ.