The K20s have arrived and I wish to arrange a
downtime to swap them for
the teslas. When would a convenient time be to take c8000eval down?
From: "christophkreisbeck(a)gmail.com" <christophkreisbeck(a)gmail.com>
Date: Sunday, March 10, 2013 7:46 PM
To: Jerry Lotto <lotto(a)harvard.edu>
Cc: Alan Aspuru-Guzik <alan(a)aspuru.com>om>, hanspeter pfister <
pfister(a)seas.harvard.edu>gt;, "hpfister(a)seas.harvard.edu" <
hpfister(a)seas.harvard.edu>gt;, "lgreenhill(a)cfa.harvard.edu" <
lgreenhill(a)cfa.harvard.edu>
Subject: Re: K20 test
Hi Jerry,
my code was developed on a NVIDIA C2075 and hence is optimized to this
architecture.
When we installed new K20 cards in Regensburg we did a comparison
between the C2075 and K20's. We obtain an increase in performance of the
K20
of about 10%-20%. Therefore, it makes sense that your installed card is a
M2090
and not a K20.
The 10%-20% is indeed not much of performance gain and we are still
investigating the reason for this low factor. There is not much data
transfer between CPU and GPU.
My impression is that our code might run into bandwidth problems when
accessing
GPU memory. But as I mentioned we are still in the early process of
investigations.
Best Christoph
2013/3/10 Lotto, Gerald <lotto(a)harvard.edu>
This is what led me to be concerned about
it:****
[root@c8000eval ~]# lspci | grep NVIDIA****
03:00.0 3D controller: NVIDIA Corporation Tesla M2090 (rev a1)****
81:00.0 3D controller: NVIDIA Corporation Tesla M2090 (rev a1)****
** **
looking at the BOM and it lists****
1 320-3773 NVIDIA M2090 PCIe x16 GPGPU card for C8220X****
1 320-9364 NVIDIA M2090 PCIe x16 GPGPU, 2nd card for C8220X****
** **
*From:* Alan Aspuru-Guzik [mailto:alan@aspuru.com]
*Sent:* Sunday, March 10, 2013 2:12 PM
*To:* Lotto, Gerald
*Cc:* pfister(a)seas.harvard.edu; christophkreisbeck(a)gmail.com;
hpfister(a)seas.harvard.edu; lgreenhill(a)cfa.harvard.edu
*Subject:* Re: K20 test****
** **
How can it be Tesla if Christoph is comParing with a K20!! Christoph can
you query the card to see what model it is? ****
** **
Alan ****
** **
--
Alan Aspuru-Guzik
Associate Professor of Chemistry and Chemical Biology
Harvard University
http://aspuru.chem.harvard.edu****
** **
On Sunday, March 10, 2013 at 2:10 PM, Lotto, Gerald wrote:****
I sent that request a while back. Will try again.
Sent from my Android Galaxy Nexus phone
-----Original Message-----
*From:* Hanspeter Pfister [pfister(a)seas.harvard.edu]
*Received:* Sunday, 10 Mar 2013, 12:27pm
*To:* Lotto, Gerald [lotto(a)harvard.edu]
*CC:* christophkreisbeck(a)gmail.com [christophkreisbeck(a)gmail.com];
hpfister(a)seas.harvard.edu [hpfister(a)seas.harvard.edu];
lgreenhill(a)cfa.harvard.edu [lgreenhill(a)cfa.harvard.edu];
aspuru(a)chemistry.harvard.edu forwards to
aspuru.com [alan(a)aspuru.com]
*Subject:* Re: K20 test****
** **
Hi Jerry,****
** **
Can you please remedy this mistake and tell Dell to send us K20s
instead, and as quickly as possible? We already have two Tesla clusters, so
the current Dell configuration is essentially useless. ****
** **
I am actually surprised it's only 10% difference. I assume the
benchmark application is not GPU bound and either uses the CPU or moves
data a lot. Christoph - can you tell us more about it?****
** **
Thanks!****
** **
- HP****
** **
On Mar 10, 2013, at 12:21 PM, "Lotto, Gerald" <lotto(a)harvard.edu> wrote:
****
****
Looks like the test system dell sent you is actually tesla, not kepler
(k20).
Sent from my Android Galaxy Nexus phone
-----Original Message-----
*From:* Alan Aspuru-Guzik [alan(a)aspuru.com]
*Received:* Sunday, 10 Mar 2013, 11:17am
*To:* Christoph Kreisbeck [christophkreisbeck(a)gmail.com]; Lotto, Gerald
[lotto(a)harvard.edu]; Hanspeter Pfister [hpfister(a)seas.harvard.edu];
Lincoln Greenhill [lgreenhill(a)cfa.harvard.edu]
*Subject:* Re: K20 test****
Dear Christoph ****
10 percent is 10 percent. Can you work with Jerry to see the difference
in specs between Regensburg and the demo setup to see if we can 'regain'
that performance?****
** **
Alan ****
** **
--
Alan Aspuru-Guzik
Associate Professor of Chemistry and Chemical Biology
Harvard University
http://aspuru.chem.harvard.edu****
** **
On Friday, March 8, 2013 at 6:23 PM, Christoph Kreisbeck wrote:****
Dear Alan,
today I did some benchmark tests with the k20eval queue.
Everything worked pretty nice. I got GPU-HEOM running
and did comparisons with Tobias' K20's in Regensburg.
For the tested parameter the wokstation solution in Regensburg
is a little bit faster (about 10%)
Best Christoph
****
** **
** **
** **