From shangupt at umich.edu  Sun Nov 11 19:23:56 2007
From: shangupt at umich.edu (Shantanu Gupta)
Date: Sun, 11 Nov 2007 22:23:56 -0500
Subject: [Hotspot] HotSpot for other floorplans..
Message-ID: <1194837836.2685.62.camel@aslan.eecs.umich.edu>

Apart from the Alpha EV6 core at 130nm technology node, has HotSpot been
validated for any other floorplan ? Preferably, for a more recent
technology node.

In my attempt to get temperature numbers at a 65nm technology node, I
did linear scaling of the floorplan area . But with this, I am getting
really high peak temperatures. For e.g. -- while running 'mgrid',
floating point units go up as high as 148 C. Has anyone else seen such a
behavior in HotSpot ? 

I'll be glad to get some feedback/suggestion on this. Thanks,

Shantanu



From shangupt at umich.edu  Sun Nov 11 20:49:52 2007
From: shangupt at umich.edu (Shantanu Gupta)
Date: Sun, 11 Nov 2007 23:49:52 -0500
Subject: [Hotspot] HotSpot for other floorplans..
In-Reply-To: <4737D22C.1020907@mail.cs.virginia.edu>
References: <1194837836.2685.62.camel@aslan.eecs.umich.edu>
	<4737D22C.1020907@mail.cs.virginia.edu>
Message-ID: <1194842992.2685.84.camel@aslan.eecs.umich.edu>

On Sun, 2007-11-11 at 20:10 -0800, Kevin Skadron wrote:
> Shantanu,
> 
> How did you scale the power from 130 to 65?  Also, how did you scale the 
> packaging parameters?  The defaults in HotSpot v3 for TIM thickness, for 
> example, are too high for 65nm.
> 

I am using the sim-alpha/Wattch simulator to generate my power numbers.
I know it is quite outdated, but that's the only architecture level tool
available for public (read academic) use. I changed some technology
dependent parameters inside the Wattch/CACTI code to get an
approximation for power dissipation by a 65nm version of the EV6 chip
running specINT/FP benchmarks. I kept the frequency at 3 GHz.
        
I have just changed the spreader/sink dimensions to appropriately match
the 65nm floorplan dimensions. But, I did not have any source to modify
the other parameters. By the way, I am using HotSpot v4.

> HotSpot has been validated for several chips, including the POWER5 and 
> Alpha EV6, and extensively against various ANSYS simulations (see our 
> WDDD 2007 paper).  You might also want to refer to the ISCA'07 paper 
> from Renau's group.
> 

Thanks for the references, I have read the WDDD paper. The ISCA 07 paper
looks very interesting, I'll take a look.
        
Thanks,
        
Shantanu

> --Kevin
> 
> Shantanu Gupta wrote:
> > Apart from the Alpha EV6 core at 130nm technology node, has HotSpot been
> > validated for any other floorplan ? Preferably, for a more recent
> > technology node.
> > 
> > In my attempt to get temperature numbers at a 65nm technology node, I
> > did linear scaling of the floorplan area . But with this, I am getting
> > really high peak temperatures. For e.g. -- while running 'mgrid',
> > floating point units go up as high as 148 C. Has anyone else seen such a
> > behavior in HotSpot ? 
> > 
> > I'll be glad to get some feedback/suggestion on this. Thanks,
> > 
> > Shantanu
> > 
> > 
> > _______________________________________________
> > HotSpot mailing list
> > HotSpot at mail.cs.virginia.edu
> > http://www.cs.virginia.edu/mailman/listinfo/hotspot
> 
> 


From mrs8n at cms.mail.virginia.edu  Mon Nov 12 07:56:27 2007
From: mrs8n at cms.mail.virginia.edu (Mircea R. Stan)
Date: Mon, 12 Nov 2007 10:56:27 -0500
Subject: [Hotspot] HotSpot for other floorplans..
In-Reply-To: <1194837836.2685.62.camel@aslan.eecs.umich.edu>
References: <1194837836.2685.62.camel@aslan.eecs.umich.edu>
Message-ID: <web-14875310@marlin.mail.virginia.edu>

On Sun, 11 Nov 2007 22:23:56 -0500
  Shantanu Gupta <shangupt at umich.edu> wrote:
> Apart from the Alpha EV6 core at 130nm technology node, has HotSpot been
> validated for any other floorplan ? Preferably, for a more recent
> technology node.
> 
> In my attempt to get temperature numbers at a 65nm technology node, I
> did linear scaling of the floorplan area . But with this, I am getting
> really high peak temperatures. For e.g. -- while running 'mgrid',
> floating point units go up as high as 148 C. Has anyone else seen such a
> behavior in HotSpot ? 

The short answer is that you cannot just scale down the floorplan
but keep everything else the same and expect a normal range for
temperatures. The high temperatures are due to the higher power
densities when you scale down the floorplan. You should look at the
power numbers (both total power and per unit power). If the total
power goes up it means that the higher peak temperatures are due
to all of them going up (say because you are not scaling the voltage but
the frequency goes up) then it means that the heatsink thermal resistance
needs to be scaled as well (basically this means you need a more
"expensive" packaging). If the total power doesn't go up it means that
the higher peak temperatures are only due to the local power densities
going up, which means that you need to improve the heat spreading,
either by a thinner die and/or by a better spreader. Hope this helps.
BTW, I would like to stress that this is *not* a problem with HotSpot,
but rather with your choice of parameters. We have validated HotSpot
for several floorplans and various technology nodes.

Mircea

> 
> I'll be glad to get some feedback/suggestion on this. Thanks,
> 
> Shantanu
> 
> 
> _______________________________________________
> HotSpot mailing list
> HotSpot at mail.cs.virginia.edu
> http://www.cs.virginia.edu/mailman/listinfo/hotspot

Mircea Stan

The views expressed in this email represent the personal
views of the sender and do not represent the official
position of the University of Virginia

From mrs8n at cms.mail.virginia.edu  Mon Nov 12 08:07:18 2007
From: mrs8n at cms.mail.virginia.edu (Mircea R. Stan)
Date: Mon, 12 Nov 2007 11:07:18 -0500
Subject: [Hotspot] HotSpot for other floorplans..
In-Reply-To: <1194842992.2685.84.camel@aslan.eecs.umich.edu>
References: <1194837836.2685.62.camel@aslan.eecs.umich.edu>
	<4737D22C.1020907@mail.cs.virginia.edu>
	<1194842992.2685.84.camel@aslan.eecs.umich.edu>
Message-ID: <web-14875872@marlin.mail.virginia.edu>

> I have just changed the spreader/sink dimensions to appropriately match
> the 65nm floorplan dimensions. But, I did not have any source to modify
> the other parameters. By the way, I am using HotSpot v4.

Aha, this is the problem! You should *not* scale the spreader and heatsink
when scaling down the technology, your choice of spreader and heatsink
should be dictated by the actual power numbers. Look for the
total power from Wattch for the two cases and adjust the heatsink
and spreader sizes appropriately (i.e. the watts/unit area be approx.
the same), this should take care of the average temperatures,
as for the local hotspots you may need to make the die and/or the
TIM thinner as suggested by Kevin. Hope this helps!

Mircea

> 
>> HotSpot has been validated for several chips, including the POWER5 and 
>> Alpha EV6, and extensively against various ANSYS simulations (see our 
>> WDDD 2007 paper).  You might also want to refer to the ISCA'07 paper 
>> from Renau's group.
>> 
> 
> Thanks for the references, I have read the WDDD paper. The ISCA 07 paper
> looks very interesting, I'll take a look.
>        
> Thanks,
>        
> Shantanu
> 
>> --Kevin
>> 
>> Shantanu Gupta wrote:
>> > Apart from the Alpha EV6 core at 130nm technology node, has HotSpot been
>> > validated for any other floorplan ? Preferably, for a more recent
>> > technology node.
>> > 
>> > In my attempt to get temperature numbers at a 65nm technology node, I
>> > did linear scaling of the floorplan area . But with this, I am getting
>> > really high peak temperatures. For e.g. -- while running 'mgrid',
>> > floating point units go up as high as 148 C. Has anyone else seen such a
>> > behavior in HotSpot ? 
>> > 
>> > I'll be glad to get some feedback/suggestion on this. Thanks,
>> > 
>> > Shantanu
>> > 
>> > 
>> > _______________________________________________
>> > HotSpot mailing list
>> > HotSpot at mail.cs.virginia.edu
>> > http://www.cs.virginia.edu/mailman/listinfo/hotspot
>> 
>> 
> 
> _______________________________________________
> HotSpot mailing list
> HotSpot at mail.cs.virginia.edu
> http://www.cs.virginia.edu/mailman/listinfo/hotspot

Mircea Stan

The views expressed in this email represent the personal
views of the sender and do not represent the official
position of the University of Virginia

From shangupt at umich.edu  Mon Nov 12 09:55:42 2007
From: shangupt at umich.edu (Shantanu Gupta)
Date: Mon, 12 Nov 2007 12:55:42 -0500
Subject: [Hotspot] HotSpot for other floorplans..
In-Reply-To: <web-14875872@marlin.mail.virginia.edu>
References: <1194837836.2685.62.camel@aslan.eecs.umich.edu>
	<4737D22C.1020907@mail.cs.virginia.edu>
	<1194842992.2685.84.camel@aslan.eecs.umich.edu>
	<web-14875872@marlin.mail.virginia.edu>
Message-ID: <1194890142.2685.90.camel@aslan.eecs.umich.edu>

Hi Mircea,

Thanks a lot for your detailed replies. I completely agree with you, it
is not a HotSpot problem. It is the issue with the config setting I have
been using with HotSpot. 

My intent was just to understand what settings in HotSpot I would need
to change when using a different floorplan. I think now I have a good
grasp on what is going on, and should be able to tune them for my
floorplan.

Thanks again,

Shantanu


On Mon, 2007-11-12 at 11:07 -0500, Mircea R. Stan wrote:
> > I have just changed the spreader/sink dimensions to appropriately match
> > the 65nm floorplan dimensions. But, I did not have any source to modify
> > the other parameters. By the way, I am using HotSpot v4.
> 
> Aha, this is the problem! You should *not* scale the spreader and heatsink
> when scaling down the technology, your choice of spreader and heatsink
> should be dictated by the actual power numbers. Look for the
> total power from Wattch for the two cases and adjust the heatsink
> and spreader sizes appropriately (i.e. the watts/unit area be approx.
> the same), this should take care of the average temperatures,
> as for the local hotspots you may need to make the die and/or the
> TIM thinner as suggested by Kevin. Hope this helps!
> 
> Mircea
> 
> > 
> >> HotSpot has been validated for several chips, including the POWER5 and 
> >> Alpha EV6, and extensively against various ANSYS simulations (see our 
> >> WDDD 2007 paper).  You might also want to refer to the ISCA'07 paper 
> >> from Renau's group.
> >> 
> > 
> > Thanks for the references, I have read the WDDD paper. The ISCA 07 paper
> > looks very interesting, I'll take a look.
> >        
> > Thanks,
> >        
> > Shantanu
> > 
> >> --Kevin
> >> 
> >> Shantanu Gupta wrote:
> >> > Apart from the Alpha EV6 core at 130nm technology node, has HotSpot been
> >> > validated for any other floorplan ? Preferably, for a more recent
> >> > technology node.
> >> > 
> >> > In my attempt to get temperature numbers at a 65nm technology node, I
> >> > did linear scaling of the floorplan area . But with this, I am getting
> >> > really high peak temperatures. For e.g. -- while running 'mgrid',
> >> > floating point units go up as high as 148 C. Has anyone else seen such a
> >> > behavior in HotSpot ? 
> >> > 
> >> > I'll be glad to get some feedback/suggestion on this. Thanks,
> >> > 
> >> > Shantanu
> >> > 
> >> > 
> >> > _______________________________________________
> >> > HotSpot mailing list
> >> > HotSpot at mail.cs.virginia.edu
> >> > http://www.cs.virginia.edu/mailman/listinfo/hotspot
> >> 
> >> 
> > 
> > _______________________________________________
> > HotSpot mailing list
> > HotSpot at mail.cs.virginia.edu
> > http://www.cs.virginia.edu/mailman/listinfo/hotspot
> 
> Mircea Stan
> 
> The views expressed in this email represent the personal
> views of the sender and do not represent the official
> position of the University of Virginia
> 
>