• We should now be fully online following an overnight outage. Apologies for any inconvenience, we do not expect there to be any further issues.

[] IBM unveils Power8 and OpenPower pincer attack on Intel’s x86 server monopoly

Page 5 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.
Toggle sidebar Toggle sidebar
T

thunng8

Member
Jan 8, 2013
167
72
101
norseamd said:
so these are for workstations?

ddr4?

how many dimms?
Click to expand...

No, currently only for servers, but I see no reason why it couldn't be used for workstations later on. Nvidia is an openPOWER memeber.

Currently 1600mhz DDR3 and announced servers can have up to 512GB per socket using 8 64GB DIMMS. 2 socket servers can have up to 1TB.
 
E

Enigmoid

Platinum Member
Sep 27, 2012
2,907
31
91
thunng8 said:
Do you mean i7? i3 does not do hyperthreading. Do you have any hard data?



As per my previous post, IBM quotes 100% scaling from single thread to 8 threads.
Click to expand...

I3 has hyperthreading, both on mobile and desktop.

Your link looks like a 100% increase in performance, not 100% scaling. 100% scaling means 8x the ST performance when using 8 threads. I'm assuming that those numbers are for 1-8 threads on the same core.
 
Last edited:
T

thunng8

Member
Jan 8, 2013
167
72
101
Enigmoid said:
I3 has hyperthreading, both on mobile and desktop.

Your link looks like a 100% increase in performance, not 100% scaling. 100% scaling means 8x the ST performance when using 8 threads. I'm assuming that those numbers are for 1-8 threads on the same core.
Click to expand...

I stand corrected on i3. It is just some i5 does not have hyper threading which makes it confusing.

Yes, I meant 100% increase. Link is for 24 cores running a thread per core I.e. 24 threads vs 24 cores running 8 thread per core I.e. Total of 192 threads.
 
I

Idontcare

Elite Member
Oct 10, 1999
21,110
64
91
thunng8 said:
I stand corrected on i3. It is just some i5 does not have hyper threading which makes it confusing.

Yes, I meant 100% increase. Link is for 24 cores running a thread per core I.e. 24 threads vs 24 cores running 8 thread per core I.e. Total of 192 threads.
Click to expand...

If this is true and IBM sees zero performance penalty per thread in loading the core with 1 thread versus 8 threads then that would imply they have redundantly built-in 8x the hardware resources per core while at the same time not sharing the resources across threads within a core in such a way as to boost single-threaded IPC (no clustering) when operating in low thread-count conditions.

Which seems really odd, they may as well have just made eight smaller cores instead of a larger 8-thread core. Yield would certainly have been higher, and the design costs for the core would certainly have been less.
 
J

jpiniero

Lifer
Oct 1, 2010
16,840
7,284
136
I'm not sure if it's even a good idea these days to build a server processor like this without having some sort of consumer version to spread out the costs.
 
R

rtsurfer

Senior member
Oct 14, 2013
733
15
76
I am curious if anyone has any Ballpark numbers about how much fasrer IBM processors are from Intel & how much more power they consume.
 
T

thunng8

Member
Jan 8, 2013
167
72
101
rtsurfer said:
I am curious if anyone has any Ballpark numbers about how much fasrer IBM processors are from Intel & how much more power they consume.
Click to expand...

~2X faster per core

190W (3.52ghz 12 core POWER8) vs 130W (2.7Ghz 12 Core Xeon 2697v2)
 
T

thunng8

Member
Jan 8, 2013
167
72
101
rtsurfer said:
Do you have any data you can conclude this from. ?
Some benchmarks or test results.
Click to expand...

see previous 3 pages

specJEnterprise: 2x
SAP SD 2 Tier: 2.2x
Specint_rate: 1.88x
specfp_rate: 2.1x

compairson is 12 core 3.52Ghz POWER8 vs 12 core 2.7Ghz Xeon 2697v2
 
Last edited:
T

thunng8

Member
Jan 8, 2013
167
72
101
Idontcare said:
If this is true and IBM sees zero performance penalty per thread in loading the core with 1 thread versus 8 threads then that would imply they have redundantly built-in 8x the hardware resources per core while at the same time not sharing the resources across threads within a core in such a way as to boost single-threaded IPC (no clustering) when operating in low thread-count conditions.

Which seems really odd, they may as well have just made eight smaller cores instead of a larger 8-thread core. Yield would certainly have been higher, and the design costs for the core would certainly have been less.
Click to expand...

Sorry for the misunderstanding it is a 100% (i.e. 2x) increase in throughput going from single thread to 8 thread (not 8x). You can see the number IBM estimates in my previous post. Note going from 4 thread to 8 threads doesn't give a large increase.

Advantage of big core is the single thread performance. POWER8 by my estimate is about 20% faster in single thread than Ivy Bridge Xeon (but because of big gains in multithread, advantage rises to about 2x per core when multithreaded).
 
Last edited:
R

rtsurfer

Senior member
Oct 14, 2013
733
15
76
thunng8 said:
see previous 3 pages

specJEnterprise: 2x
SAP SD 2 Tier: 2.2x
Specint_rate: 1.88x
specfp_rate: 2.1x

compairson is 12 core 3.52Ghz POWER8 vs 12 core 2.7Ghz Xeon 2697v2
Click to expand...

Okay thanks.
I will read the last few pages properly.
 
witeken

witeken

Diamond Member
Dec 25, 2013
3,899
193
106
thunng8 said:
Advantage of big core is the single thread performance. POWER8 by my estimate is about 20% faster in single thread than Ivy Bridge Xeon (but because of big gains in multithread, advantage rises to about 2x per core when multithreaded).
Click to expand...

So Haswell is about as fast as POWER8 (single threaded)?
 
N

nismotigerwvu

Golden Member
May 13, 2004
1,568
33
91
Phynaz said:
IBM has an x86 license.
Click to expand...

But they would far too behind in terms of cross licensing of modern tech between AMD and Intel to be able to come up with a competitive product. That and I believe their license is basically worthless since all the 486-class (last generation they competed in) related patents have all expired.
 
witeken

witeken

Diamond Member
Dec 25, 2013
3,899
193
106
But the comparison is 3.52GHz for POWER8 vs 2.7GHz for Intel, so wouldn't Haswell be ~30% faster per clock?

Edit:

I just found this nice overview of POWER8: Power8 Muscles Up for Servers.

Intel has not updated its high-end Xeon processor since introducing Westmere-EX in 2011. By the time Power8 ships, Intel will be in production with a 22nm product that is likely to add cores and increase CPU speed so as to improve the total throughput by about 40% (see MPR 8/12/13, “Intel Adds 14nm to Server Roadmap”). This upgrade will still leave Xeon behind in performance per socket, but the Intel processor uses far less power than the others. Its 130W TDP rating equates to about 100W typical, making Westmere-EX the most power efficient of the processors in this chart that are currently in production.
Click to expand...
 
Last edited:
I

Idontcare

Elite Member
Oct 10, 1999
21,110
64
91
thunng8 said:
Sorry for the misunderstanding it is a 100% (i.e. 2x) increase in throughput going from single thread to 8 thread (not 8x). You can see the number IBM estimates in my previous post. Note going from 4 thread to 8 threads doesn't give a large increase.

Advantage of big core is the single thread performance. POWER8 by my estimate is about 20% faster in single thread than Ivy Bridge Xeon (but because of big gains in multithread, advantage rises to about 2x per core when multithreaded).
Click to expand...

Ah yes, that makes much more sense now. Thanks for revisiting and explaining the performance scaling numbers to me.
 
T

thunng8

Member
Jan 8, 2013
167
72
101
witeken said:
But the comparison is 3.52GHz for POWER8 vs 2.7GHz for Intel, so wouldn't Haswell be ~30% faster per clock?

Edit:

I just found this nice overview of POWER8: Power8 Muscles Up for Servers.
Click to expand...

Things get a bit complicated with Turbo boost etc. i.e the Xeon 2.7Ghz can turbo boost to 3.5Ghz and the POWER8 to around 3.9Ghz. But overall, I think single threaded performance of Haswell and POWER8 should be within ballpark of one another but POWER8 would have much higher multithread performance. Of course Haswell Xeons are not yet on the market and won't be until late this year/early next year.

The article linked above has some facts wrong. POWER8 is 190W (not 250W). POWER7+ is around 180W (not 250W). POWER7 was 250W for top bin and I think they assumed POWER7+ was the same and it is not.

Updated Xeons (Ivy Bridge) that I did the comparison against was released in late 2013 and early 2014 so that quote about Xeon's not being updated since 2011 is irrelevant.
 
You must log in or register to reply here.

TRENDING THREADS

COMPANY
RESOURCES
FOLLOW
Top Bottom