# New Silicon Insider Article

Article: Escape From the Planet of x86
By: mas (mas769.delete@this.hotmail.com), June 18, 2003 4:11 pm
Nate Begeman (nbegeman@umich.edu) on 6/18/03 wrote:
---------------------------
>Bill Todd (billtodd@metrocast.net) on 6/18/03 wrote:
>---------------------------
>>However, that interesting .pdf (thanks for the pointer)

yeah, great link alberto. Also shows more than just a shrink with double the L3 set associativity,bandwidth and actual relayout.

>>does support my understanding
>>that Madison indeed consumes 130 W at 1.3v (and 1.5 GHz). So how the power consumption
>>could be halved by reducing the voltage only 15% and the frequency only 13% (to
>>1.1v and 1.3 GHz, as you suggest below) remains unclear to me (though I don't have
>>any problem believing that a 1 GHz part - possibly at even lower voltage - might halve the power requirement).
>
>Yeah, half does seem a little optimistic. However, we know the 1.3GHz part has
>half the L3 cache, and cache leakage power is 7% of the >130W by itself. Going from

I took the 2nd pie chart on page 19 to show the final power distributions after the L3 power reduction scheme install.
Using this as a starting point. Take away 40% of the IO power for halving the cache, 2% and say another 1% out of the 5% static leakage for the same reason gives you a starting point of 130 * 0.97 = 126.1 W which corresponds to
121W dynamic/io and 5.1 W leakage

Static leakage power loss, unlike dynamic/io power varies linearly with voltage. Therefore estimated power of 1.1V 1.3 Ghz 3Mb Deerfield = (121 * (1.1/1.3)^2 * 1.3/1.5) + (5.1 * 1.1/1.3) = 80W which is definitely P4/Athlon/Opteron territory.

Also the schmoo defintely backs up the 1.67 Ghz figure I have seen quoted for the 9Mb version which also implies a bus frequency rise from 200 to 333 Mhz (multiplier 5 *333 = 1.67 Ghz) leading to a bandwidth rise from 6.4 to 10.67 Gb/s.

>the chart, and using the "chip power scales linearly with clock and as the square of voltage" approximations...
>
>Take out ~5W for half the cache
>125W * 1.3/1.5 = 108W
>108W * 1.1^2/1.3^2 = 77W
>
>We aren't at 65W (half power), but it's not all that far away, assuming the 1.3GHz
>chip uses a 1.1V source and the 1.5GHz chip uses the 1.3V source. I'm not sure
>that's a safe assumption to make, but 77W isn't too horrendous for the amount of
>performance that a 1.3GHz Itanium2 is likely to deliver. The only way that you
>could get those last 12W back is if somehow going to 1.1V from 1.3V *significantly*
>reduced leakage in the chip, on the order of what, about two thirds?
TopicPosted ByDate
New Silicon Insider ArticleDavid Kanter2003/06/17 03:39 PM
Srockholm Syndromeanonymous2003/06/17 03:50 PM
Srockholm SyndromeNate Begeman2003/06/17 04:32 PM
Srockholm Syndromeanonymous2003/06/18 02:23 PM
Srockholm SyndromeScott Robinson2003/06/20 08:25 AM
New Silicon Insider ArticleBill Todd2003/06/17 09:51 PM
New Silicon Insider ArticleAlberto2003/06/18 07:29 AM
New Silicon Insider ArticleJosé Javier Zarate2003/06/18 10:16 AM
New Silicon Insider ArticleBill Todd2003/06/18 03:10 PM
New Silicon Insider ArticleNate Begeman2003/06/18 03:25 PM
New Silicon Insider ArticleTvar'2003/06/18 03:41 PM
New Silicon Insider ArticleAlberto2003/06/18 03:58 PM
New Silicon Insider ArticleTvar'2003/06/18 04:04 PM
New Silicon Insider ArticleAlberto2003/06/18 04:24 PM
New Silicon Insider ArticleTvar'2003/06/18 04:32 PM
New Silicon Insider ArticlePaul DeMone2003/06/18 04:13 PM
New Silicon Insider ArticleTvar'2003/06/18 04:23 PM
New Silicon Insider Articlemas2003/06/18 04:11 PM
New Silicon Insider ArticleAlberto2003/06/18 03:45 PM
New Silicon Insider ArticleBill Todd2003/06/18 11:46 PM
New Silicon Insider ArticleDavid Wang2003/06/19 12:13 AM
New Silicon Insider ArticleBill Todd2003/06/19 01:14 AM
New Silicon Insider ArticleDavid Wang2003/06/19 10:52 AM
New Silicon Insider ArticlePaul DeMone2003/06/18 04:04 PM
New Silicon Insider ArticleBill Todd2003/06/18 11:28 PM
New Silicon Insider ArticlePaul DeMone2003/06/19 12:43 AM
New Silicon Insider ArticleRob Young2003/06/19 10:23 AM
New Silicon Insider ArticleBill Todd2003/06/19 04:53 PM
New Silicon Insider ArticleDavid Wang2003/06/18 11:29 PM
New Silicon Insider ArticleBill Todd2003/06/19 12:03 AM
New Silicon Insider ArticleJosé Javier Zarate2003/06/19 05:33 AM
New Silicon Insider Articlemas2003/06/19 06:37 AM
New Silicon Insider ArticleBill Todd2003/06/19 04:40 PM
New Silicon Insider ArticleDavid Wang2003/06/19 05:25 PM
New Silicon Insider ArticleBill Todd2003/06/19 06:00 PM
New Silicon Insider ArticleAlberto2003/06/19 06:29 PM
New Silicon Insider ArticleSpeedy2003/06/19 06:48 PM
New Silicon Insider ArticleAlberto2003/06/20 04:57 AM
New Silicon Insider ArticleDavid Wang2003/06/19 06:52 PM
New Silicon Insider ArticleBill Todd2003/06/19 09:00 PM
New Silicon Insider ArticleAnonymous2003/06/20 02:20 AM
New Silicon Insider ArticlePaul DeMone2003/06/20 09:11 AM
New Silicon Insider ArticleAnonymous2003/06/22 04:48 PM
New Silicon Insider ArticlePaul DeMone2003/06/22 05:49 PM
New Silicon Insider ArticleVincent Diepeveen2003/06/22 06:25 PM
New Silicon Insider ArticleJosé Javier Zarate2003/06/22 07:55 PM
New Silicon Insider ArticleAnonymous2003/06/23 09:59 AM
New Silicon Insider ArticlePaul DeMone2003/06/19 07:53 PM
New Silicon Insider ArticleBill Todd2003/06/19 08:53 PM
New Silicon Insider ArticleDavid Wang2003/06/19 09:08 PM
New Silicon Insider ArticleBill Todd2003/06/20 02:28 AM
New Silicon Insider ArticleDavid Wang2003/06/20 11:35 AM
New Silicon Insider ArticlePaul DeMone2003/06/20 12:29 PM
New Silicon Insider ArticleBill Todd2003/06/20 07:10 PM
New Silicon Insider ArticleMarc M.2003/06/21 06:06 AM
New Silicon Insider ArticleBill Todd2003/06/21 12:07 PM
New Silicon Insider ArticleBill Todd2003/06/20 07:01 PM
New Silicon Insider ArticleDavid Wang2003/06/20 07:52 PM
New Silicon Insider ArticleBill Todd2003/06/20 08:53 PM
New Silicon Insider ArticleDavid Wang2003/06/20 09:14 PM
New Silicon Insider ArticleVincent Diepeveen2003/06/20 09:52 PM
New Silicon Insider ArticleMarc M.2003/06/21 08:16 AM
New Silicon Insider ArticleVincent Diepeveen2003/06/22 05:24 PM
New Silicon Insider ArticleSingh, S.R.2003/06/21 04:39 AM
New Silicon Insider ArticleDavid Wang2003/06/21 09:10 AM
IPF CompilersNate Begeman2003/06/21 10:10 AM
IPF CompilersPaul DeMone2003/06/21 10:45 AM
Use ILP to extract more ILPPaul DeMone2003/06/20 11:48 PM
New Silicon Insider ArticlePaul DeMone2003/06/20 09:06 AM
New Silicon Insider ArticleSingh, S.R.2003/06/20 10:41 AM
New Silicon Insider ArticleDavid Kanter2003/06/21 04:34 PM
New Silicon Insider ArticlePaul DeMone2003/06/22 03:22 PM
New Silicon Insider ArticleBill Todd2003/06/20 06:52 PM
New Silicon Insider ArticleMarc M.2003/06/21 08:54 AM
New Silicon Insider ArticleDaniel Gustafsson2003/06/19 12:12 PM
New Silicon Insider ArticlePaul DeMone2003/06/20 03:20 PM
New Silicon Insider ArticleBryan Gregory2003/06/20 02:14 PM
New Silicon Insider Articlemas2003/06/20 02:43 PM
New Silicon Insider ArticlePaul DeMone2003/06/25 11:29 AM
New Silicon Insider ArticleJosé Javier Zarate2003/06/25 11:43 AM
New Silicon Insider ArticlePaul DeMone2003/06/25 11:52 AM
lol, amazing coincidence :-) (NT)mas2003/06/25 04:15 PM
New Silicon Insider ArticleYoav2015/04/01 04:43 AM