From ks4kk at cs.virginia.edu Sun May 4 19:01:01 2003 From: ks4kk at cs.virginia.edu (Karthik Sankaranarayanan) Date: Wed Mar 22 16:04:07 2006 Subject: [Hotspot] Re: Hotspot In-Reply-To: <3EB58F34.19D72101@cse.psu.edu> References: <3EB58F34.19D72101@cse.psu.edu> Message-ID: Two possible problems come to my mind - 1. Is the 'temp' array initialized to proper 'initial temperature' values? compute_temp only 'advances' those initial values. 2. What calling interval are you using (i.e., the elapsed_time variable)? When the elapsed_time is too large (like 1 second etc..), the step size for Runge-Kutta solver will be high resulting in large computation errors. Typically, the calling interval is in the order of micro/milli seconds. This issue is fixed now and the tool automatically manages the accuracy of the Runge-Kutta solver. This feature will be available in future releases of HotSpot. Hope this helps - let me know if this solves your problem. Thanks, -karthik On Sun, 4 May 2003, Ananth Hegde Ankadi wrote: > Hi, > I am trying to call compute_temp() in the hot spot tool from a > simulator. > compute_temp(power, temp, flp_adj->n_units, elapsed_time); > After call it if I print temp using > printf("%s\t%.1f\n", flp_adj->units[i]->name, temp[i]); > It is giving me vey huge and irrelavent values. > Something like this. > L2_left > -3213614823702844426850394201710407249553809693250406279120558735537209344.0 > > I am sure the adjceny matri x is created properly because it creates the > floor plan. Also if I print the temperatures after I sim_init() in > sim-foo.c that you have provided, it works well. > I am assigniing the power values properly as well.Like this > power[get_blk_index(flp_adj, "ITB")] = 0.248; > > Could you tell me what could be the problem. I havent changed anything > much. > > A quick reply would be appreciated. > Thanks > Ananth > > > > -- > ---------------------- > Ananth Hegde > Grad Student > Penn State University > ---------------------- > "There are 10 types of people in the world. Those that understand binary > and those that dont!!!!!!!" > From ks4kk at cs.virginia.edu Tue May 6 17:38:30 2003 From: ks4kk at cs.virginia.edu (Karthik Sankaranarayanan) Date: Wed Mar 22 16:04:07 2006 Subject: [Hotspot] Re: technology? In-Reply-To: <3EB808D9.14B57891@cse.psu.edu> References: <3EB808D9.14B57891@cse.psu.edu> Message-ID: Ananth, HotSpot takes power numbers from other simulators and hence is not directly dependent on tech parameters. It is the power simulator that needs to take care of the tech parameters. However, as technology changes, chip, heatsink and spreader specifications (eg - chip area, thickness etc.) might change - correspondingly, proper values need to be set in sim-foo.c and the floorplan file has to be updated with the correct areas. -karthik On Tue, 6 May 2003, Ananth Hegde Ankadi wrote: > Karthik, > Is hotspot dependent on the technology parameter? > If yes which one did you use? and how can it be scaleD? > thanks > Ananth > > ---------------------- > Ananth Hegde > Grad Student > Penn State University > ---------------------- > "There are 10 types of people in the world. Those that understand binary > and those that dont!!!!!!!" > From lshang at EE.Princeton.EDU Thu May 8 22:12:57 2003 From: lshang at EE.Princeton.EDU (Li Shang) Date: Wed Mar 22 16:04:07 2006 Subject: [Hotspot] a question Message-ID: Hello, I have a question about the lateral thermal resistance in heat speader. I read the source code. It seems that the central part of heat spreader is modeled with one vertical thermal resistor. Do you think that is accurate enough to model the thermal profile distribution? Thank you very much. cheers, Li _________________________________ *Li Shang EE Dept. CE group * *Tel (O)609-258-4108 * *Email lshang@ee.princeton.edu * * lshang@princeton.edu * _________________________________ From wh6p at cms.mail.virginia.edu Fri May 9 09:59:17 2003 From: wh6p at cms.mail.virginia.edu (Wei Huang) Date: Wed Mar 22 16:04:07 2006 Subject: [Hotspot] a question In-Reply-To: References: Message-ID: <1209781.1052474357@Lonely_Wolves.Virginia.EDU> Hi Li, Actually, we had thought about that. The temperature gradient of chip bottom/heatspreader is negligible compared to the average temperature of that part. In one of our simulations, the gradient is about 4 degrees but the chip bottom average is 107 degrees above ambient. This senario considers almost the worst case: 0.1mm chip thickness, huge power density difference accross the chip (4.4W/mm^2 max, 0.2W/mm^2 min). The error is about 4%. This small error won't vary steady-state temperatures much, it can be neglected so that the model is less complicated. But for the transient part, because the thermal time constant is very small for the chip, in order to get precise transient response, we have to consider the fact that the chip bottom/spreader is not isothermal when calculating thermal capacitance. The actual heat-up thermal mass is bigger than the chip itself. This is why there is a factor 2.0 to the chip block's thermal capacitance. We found that this factor is related to chip/spreader geometry sizes and material. You will find more details in our tech report soon. Hope this helps. Wei Huang --On 2003??5??8?? 22:12 -0400 Li Shang wrote: > Hello, > > I have a question about the lateral thermal resistance in > heat speader. I read the source code. It seems that the > central part of heat spreader is modeled with one > vertical thermal resistor. Do you think that is accurate > enough to model the thermal profile distribution? Thank > you very much. > > cheers, > > Li > > _________________________________ > *Li Shang EE Dept. CE group * > *Tel (O)609-258-4108 * > *Email lshang@ee.princeton.edu * > * lshang@princeton.edu * > _________________________________ > > _______________________________________________ > HotSpot mailing list > HotSpot@cs.virginia.edu > http://www.cs.virginia.edu/mailman/listinfo/hotspot From ankadi at psu.edu Sun May 25 11:09:28 2003 From: ankadi at psu.edu (ANANTH ) Date: Wed Mar 22 16:04:07 2006 Subject: [Hotspot] Calling compute temp Message-ID: <200305251509.LAA10189@webmail9.cac.psu.edu> Hi, I still have questions over caling compute temp. Your tool says /* set the per cycle power values as returned by Wattch/power simulator If I were to call compute_temp every 100000 cycles, Should these power values be an aggregate(which I think should be since you are measuring the average later but the comment indicates the other way) over 100000 cycles or the average over 100000 cycles. Assuming it to be an aggregate, Say I asign the following power[get_blk_index(flp_adj, "L2_left")] = 0.480; power[get_blk_index(flp_adj, "L2_right")] = 0.452; power[get_blk_index(flp_adj, "L2_bottom")] = 0.980; power[get_blk_index(flp_adj, "Icache")] = 0.5243; power[get_blk_index(flp_adj, "Dcache")] = 0.9035; power[get_blk_index(flp_adj, "Bpred")] = 5.993; power[get_blk_index(flp_adj, "DTB")] = 0.113; power[get_blk_index(flp_adj, "FPAdd")] = 0.378; power[get_blk_index(flp_adj, "FPReg")] = 0.238; power[get_blk_index(flp_adj, "FPMul")] = 0.404; power[get_blk_index(flp_adj, "FPMap")] = 0.017; power[get_blk_index(flp_adj, "IntMap")] = 2.666; power[get_blk_index(flp_adj, "IntQ")] = 2.233; power[get_blk_index(flp_adj, "IntReg")] = 2.336; power[get_blk_index(flp_adj, "IntExec")] = 2.454; power[get_blk_index(flp_adj, "FPQ")] = 3.411; power[get_blk_index(flp_adj, "LdStQ")] = 2.035; power[get_blk_index(flp_adj, "ITB")] = 0.248; And my elapsed time was 100000*PERIOD where PERIOD is 1e-9. I get huge temperatures of the order of 10000s. I also have tried setting the power no.s to average over 100000 sycles, eve then the temperature after one call becomes say for BPRED around 700K. Could you please tell me where am I going wrong? Thanks Ananth