From hitoshi_mizunuma at tet.toshiba.co.jp Tue Nov 2 10:02:07 2004 From: hitoshi_mizunuma at tet.toshiba.co.jp (hitoshi_mizunuma@tet.toshiba.co.jp) Date: Wed Mar 22 16:04:14 2006 Subject: [Hotspot] Is sub-block really hotter than original? Message-ID: Hello, I have a question. If you split a block into two, the temperature of each sub-block tends to be getting higher than that of original one. You can easily make sure it by splitting a cache block and giving a uniform power to them. I guess it is because "splitting" itself increases resistance value of the block in total, causing to exaggerate self-heating. Do you think it makes sense? If it does, each bank for cache or register file might be possibly hotter than monolithic one even though it burns less power... (of course, I assume uniform power distribution here) Thanks for your help. Hitoshi From wh6p at cms.mail.virginia.edu Tue Nov 2 11:13:14 2004 From: wh6p at cms.mail.virginia.edu (Wei Huang) Date: Wed Mar 22 16:04:14 2006 Subject: [Hotspot] Is sub-block really hotter than original? In-Reply-To: References: Message-ID: <6194515.1099393994@hilopc3.ee.Virginia.EDU> Hi Hitoshi, Could you tell us how much hotter you have observed? Is it abnormally hotter? I am assuming you were applying the same power density before and after splitting the cache block, am I right? HotSpot is now actually modeling the center temperature of each block. So if you split one block into two, you are modeling the temperatures of two new locations--centers of the two sub-block, not the original center. Whether these two sub-block center temperatures would be hotter or cooler than the original center, that depends on their neighbor blocks temperatures, it is hard to say. Have you always observed higher temperatures no matter which block is splitted? My guess is that if you split one of the hottest blocks, you should observer lower new temperatures. As to which temperatures can best represent the temperature of the whole block, it is not clear. But if you have only a rough estimate of the block's power, I would say the center temperature of the entire block is a good estimate. Please let us know if you have further questions. Thanks! Wei --On Tuesday, November 02, 2004 11:02 PM +0800 hitoshi_mizunuma@tet.toshiba.co.jp wrote: > Hello, > > I have a question. > If you split a block into two, the temperature of each > sub-block tends to be getting higher than that of > original one. You can easily make sure it by splitting a > cache block and giving a uniform power to them. I guess > it is because "splitting" itself increases resistance > value of the block in total, causing to exaggerate > self-heating. > > Do you think it makes sense? If it does, each bank for > cache or register file might be possibly hotter than > monolithic one even though it burns less power... (of > course, I assume uniform power distribution here) > > Thanks for your help. > Hitoshi > > > > > _______________________________________________ > HotSpot mailing list > HotSpot@cs.virginia.edu > http://www.cs.virginia.edu/mailman/listinfo/hotspot From hitoshi_mizunuma at tet.toshiba.co.jp Tue Nov 2 20:05:34 2004 From: hitoshi_mizunuma at tet.toshiba.co.jp (hitoshi_mizunuma@tet.toshiba.co.jp) Date: Wed Mar 22 16:04:14 2006 Subject: [Hotspot] Is sub-block really hotter than original? Message-ID: Dear all; >My guess is that if you split one of the hottest >blocks, you should observer lower new temperatures. Unfortunately, all of sub-blocks get higher than original. This is true even when you disabled lateral effect. >If this is indeed true then it must be a recent bug in the code It might not be a bug, I guess. Splitting the block always makes each sub-block thermally more isolated because an inter-sub-block heat moving over the wire gets minor. I am just curious about whether you've validated this model for a banked structure (kind of loosely coupled sub-block) or not... Thanks Hitoshi "Mircea R. Stan" hotspot@cs.virginia.edu cc: 2004/11/03 05:14 Subject: Re: [Hotspot] Is sub-block really hotter than original? On Tue, 2 Nov 2004 23:02:07 +0800 hitoshi_mizunuma@tet.toshiba.co.jp wrote: >Hello, > >I have a question. >If you split a block into two, the temperature of each >sub-block tends to be getting higher than that of >original one. >You can easily make sure it by splitting a cache block >and giving a uniform power to them. >I guess it is because "splitting" itself increases >resistance value of the block in total, causing to >exaggerate self-heating. > >Do you think it makes sense? If this is indeed true then it must be a recent bug in the code, I'm pretty sure that we extensivelly checked for issues like this when we were first developing the tool. We will try to replicate this and let you know, we may need to see what exactly you are doing if we cannot replicate it. Mircea >If it does, each bank for >cache or register file might be possibly hotter than >monolithic one even though it burns less power... >(of course, I assume uniform power distribution here) > >Thanks for your help. >Hitoshi > > > > >_______________________________________________ >HotSpot mailing list >HotSpot@cs.virginia.edu >http://www.cs.virginia.edu/mailman/listinfo/hotspot Mircea Stan Associate Professor Charles L. Brown ECE Department University of Virginia From wh6p at cms.mail.virginia.edu Wed Nov 3 10:47:57 2004 From: wh6p at cms.mail.virginia.edu (Wei Huang) Date: Wed Mar 22 16:04:14 2006 Subject: [Hotspot] Is sub-block really hotter than original? In-Reply-To: References: Message-ID: <3305171.1099478877@hilopc3.ee.Virginia.EDU> Hi Hitoshi, I've tried to replicate your observation by splitting "IntReg" into two parts. But I didn't find what you observed. In contrast, for the two sub-blocks of "IntReg", I observed one is hotter and the other is cooler than the original one. The same happened when I splitted "Icache". Could you double-check whether we have the same version of HotSpot and the same splitting process? I was using HotSpot-2.0 (newest version on hotspot website), and I changed in the floorplan file (ev6.flp) by halving the width (or height) of one block, and change conrresponding left-x (or bottom-y) data. In addition, I also changed the power assignment in "simtemplate.c" by removing the old block power assignment and adding two new sub-blocks. The power to each new sub-block is half of the original. (I didn't read the power trace file to assign powers in "sim-template.c". Instead, I assigned constant power to each block and only calculated the steady-state temperatures.) Thanks! -Wei --On Wednesday, November 03, 2004 9:05 AM +0800 hitoshi_mizunuma@tet.toshiba.co.jp wrote: > > Dear all; > >> My guess is that if you split one of the hottest >> blocks, you should observer lower new temperatures. > > Unfortunately, all of sub-blocks get higher than original. > This is true even when you disabled lateral effect. > >> If this is indeed true then it must be a recent bug in > the code > > It might not be a bug, I guess. Splitting the block > always makes each sub-block thermally more isolated > because an inter-sub-block heat moving over the wire gets > minor. > > I am just curious about whether you've validated this > model for a banked structure (kind of loosely coupled > sub-block) or not... > > Thanks > Hitoshi > > > > > > "Mircea R. Stan" > To: hitoshi_mizunuma@tet.toshiba.co.jp, > rginia.edu> > hotspot@cs.virginia.edu > > cc: > 2004/11/03 05:14 > Subject: Re: [Hotspot] Is sub-block really hotter than > > original? > > > > > > > On Tue, 2 Nov 2004 23:02:07 +0800 > hitoshi_mizunuma@tet.toshiba.co.jp wrote: >> Hello, >> >> I have a question. >> If you split a block into two, the temperature of each >> sub-block tends to be getting higher than that of >> original one. >> You can easily make sure it by splitting a cache block >> and giving a uniform power to them. >> I guess it is because "splitting" itself increases >> resistance value of the block in total, causing to >> exaggerate self-heating. >> >> Do you think it makes sense? > > If this is indeed true then it must be a recent bug in > the code, I'm pretty sure that we extensivelly > checked for issues like this when we were first > developing the tool. We will try to replicate > this and let you know, we may need to see what > exactly you are doing if we cannot replicate it. > > Mircea > >> If it does, each bank for >> cache or register file might be possibly hotter than >> monolithic one even though it burns less power... >> (of course, I assume uniform power distribution here) >> >> Thanks for your help. >> Hitoshi >> >> >> >> >> _______________________________________________ >> HotSpot mailing list >> HotSpot@cs.virginia.edu >> http://www.cs.virginia.edu/mailman/listinfo/hotspot > > Mircea Stan > Associate Professor > Charles L. Brown ECE Department > University of Virginia > > > > >