PAPER 11/20: The potential of the cell processor for scientific computing
9 views
Skip to first unread message
Guofeng
Nov 17, 2007, 12:53:38 PM11/17/07
to ASU:CSE520 FALL 07 Advanced Computer Architecture
Samuel Williams, John Shalf, Leonid Oliker
Shoaib Kamil, Parry Husbands, Katherine Yelick
Computational Research Division
Lawrence Berkeley National Laboratory
Berkeley, CA 94720
{swwilliams,jshalf,loliker,sakamil,prjhusbands,kayelick}@lbl.gov
ACM CF'06, May 3-5, 2006, Ischia, Italy.
Shoaib Kamil, Parry Husbands, Katherine Yelick
Computational Research Division
Lawrence Berkeley National Laboratory
Berkeley, CA 94720
{swwilliams,jshalf,loliker,sakamil,prjhusbands,kayelick}@lbl.gov
ACM CF'06, May 3-5, 2006, Ischia, Italy.
Rob Duckles
Nov 24, 2007, 10:22:59 PM11/24/07
to ASU:CSE520 FALL 07 Advanced Computer Architecture
Critique:
Introduction:
The paper discusses the Cell processor in the context of scientific
computing--including Fourier Transforms, matrix multiplication, vector
multiplication, etc.
Strengths:
* Identifies concrete benchmarks, and details the results compared to
common alternatives.
* Takes power-consumption into consideration, detailing usage for
particular applications.
* Has a good overview of Cell architecture, providing a foundation
for further discussion of performance, programming, and enhancements.
Weaknesses:
* The paper is limited in scope. There are many useful benchmarks
that were omitted (e.g., Dhrystone, Whetstone). What about other types
of computing applications? The benchmarks chosen seem to be biased
towards Cell's architecture.
* Apples-to-oranges comparison between Cell and Intel/AMD chips.
Cell programs need to be hand-optimized to use all 8 SPE's, yet they
use more generic code for Intel/AMD. What would the performance be on
an Opteron chip if the code was hand-optimized in AMD64 assembly?
* Lacks an appropriate discussion on the strengths/weaknesses of
Cell. Are there any scientific computing tasks that the Cell is ill-
suited for?
On Nov 17, 10:53 am, Guofeng <guofeng.d...@gmail.com> wrote:
> Samuel Williams, John Shalf, Leonid Oliker
> Shoaib Kamil, Parry Husbands, Katherine Yelick
> Computational Research Division
> Lawrence Berkeley National Laboratory
> Berkeley, CA 94720
> {swwilliams,jshalf,loliker,sakamil,prjhusbands,kayeli...@lbl.gov
Introduction:
The paper discusses the Cell processor in the context of scientific
computing--including Fourier Transforms, matrix multiplication, vector
multiplication, etc.
Strengths:
* Identifies concrete benchmarks, and details the results compared to
common alternatives.
* Takes power-consumption into consideration, detailing usage for
particular applications.
* Has a good overview of Cell architecture, providing a foundation
for further discussion of performance, programming, and enhancements.
Weaknesses:
* The paper is limited in scope. There are many useful benchmarks
that were omitted (e.g., Dhrystone, Whetstone). What about other types
of computing applications? The benchmarks chosen seem to be biased
towards Cell's architecture.
* Apples-to-oranges comparison between Cell and Intel/AMD chips.
Cell programs need to be hand-optimized to use all 8 SPE's, yet they
use more generic code for Intel/AMD. What would the performance be on
an Opteron chip if the code was hand-optimized in AMD64 assembly?
* Lacks an appropriate discussion on the strengths/weaknesses of
Cell. Are there any scientific computing tasks that the Cell is ill-
suited for?
On Nov 17, 10:53 am, Guofeng <guofeng.d...@gmail.com> wrote:
> Samuel Williams, John Shalf, Leonid Oliker
> Shoaib Kamil, Parry Husbands, Katherine Yelick
> Computational Research Division
> Lawrence Berkeley National Laboratory
> Berkeley, CA 94720
Chi-Li Yu
Nov 27, 2007, 1:37:50 AM11/27/07
to asucse520-fall-07-advanc...@googlegroups.com
Dear all,
The attached file is my critique about this paper.
Regards,
Chi-Li
Chi-Li Yu
Nov 27, 2007, 1:55:32 AM11/27/07
to asucse520-fall-07-advanc...@googlegroups.com
Sorry. I forgot attached the file.
Let me send it again.
Let me send it again.
Chi-Li
----- Original Message -----
From: "Guofeng" <guofen...@gmail.com>
To: "ASU:CSE520 FALL 07 Advanced Computer Architecture"
<asucse520-fall-07-advanc...@googlegroups.com>
Sent: Saturday, November 17, 2007 10:53 AM
Subject: PAPER 11/20: The potential of the cell processor for scientific
computing
>
Sugan Vinayagam
Nov 27, 2007, 12:43:57 PM11/27/07
to asucse520-fall-07-advanc...@googlegroups.com
Hi all,
Please find my critique about this paper.
Thanks,
Sugan.
Please find my critique about this paper.
Thanks,
Sugan.
> {swwilliams,jshalf,loliker,sakamil,prjhusbands, kayelick}@lbl.gov
Pradnyesh Gudadhe
Nov 27, 2007, 3:59:26 PM11/27/07
to asucse520-fall-07-advanc...@googlegroups.com
Please find my critique attached.
Thanks.
Regards,
Pradnyesh--
Please visit my UPDATED website here:
http://www.geocities.com/paddyinpilani
Thanks.
Regards,
Pradnyesh
Please visit my UPDATED website here:
http://www.geocities.com/paddyinpilani
jun shen
Nov 27, 2007, 7:07:41 PM11/27/07
to ASU:CSE520 FALL 07 Advanced Computer Architecture
summary:
Summary of "The Potential of the Cell Processor
for Scientific Computing"
Jun Shen 993992089 Nov 20th, 2007
1. Paper Outline
The paper examines the performance of cell-based computer
architecture. The paper defines performance model for evaluation,
which includes dense matrix multiply, sparse matrix vector multiply,
1D/2D FFT. And the paper validates the accuracy of their model with
published hardware result and performs a further comparison between
cell architecture and existing processors such as AMD operon and Intel
Itanium2. Finally the paper proposes modest micro-architectural
modification to increase the efficiency of double-precision
calculation.
Overview of critique
2. Strength:
I. The paper presents comparison between cell architecture with other
alternatives and gives quantitative comparison result.
II. The used benchmark covers most important algorithms including
matrix computation, stencil computation and FFT (fast Fourier
transforms). Besides, the paper also addresses double precision and
single precision computation.
III. The paper also presents cell+ architecture to improve the double-
precision computation performance of cell.
IV. The comparison also takes power consumption into consideration.
3. Weakness:
I. The performance comparison is not fair. Here are the reasons.
First, the code running on cell is hand optimized. Second, the
selected computations are most frequently used in game (except double
precision computation) and therefore are strength of cell
architecture.
II. Cell is incompetent in SpMV, a kind of computation which lacks
unaligned load support.
III. The paper shall give more detailed explanation of cell+
architecture.
IV. Only 6 of 8 SPEs are available for use in reality, so the
performance data of cell with 8 working SPEs is dubious.
Summary of "The Potential of the Cell Processor
for Scientific Computing"
Jun Shen 993992089 Nov 20th, 2007
1. Paper Outline
The paper examines the performance of cell-based computer
architecture. The paper defines performance model for evaluation,
which includes dense matrix multiply, sparse matrix vector multiply,
1D/2D FFT. And the paper validates the accuracy of their model with
published hardware result and performs a further comparison between
cell architecture and existing processors such as AMD operon and Intel
Itanium2. Finally the paper proposes modest micro-architectural
modification to increase the efficiency of double-precision
calculation.
Overview of critique
2. Strength:
I. The paper presents comparison between cell architecture with other
alternatives and gives quantitative comparison result.
II. The used benchmark covers most important algorithms including
matrix computation, stencil computation and FFT (fast Fourier
transforms). Besides, the paper also addresses double precision and
single precision computation.
III. The paper also presents cell+ architecture to improve the double-
precision computation performance of cell.
IV. The comparison also takes power consumption into consideration.
3. Weakness:
I. The performance comparison is not fair. Here are the reasons.
First, the code running on cell is hand optimized. Second, the
selected computations are most frequently used in game (except double
precision computation) and therefore are strength of cell
architecture.
II. Cell is incompetent in SpMV, a kind of computation which lacks
unaligned load support.
III. The paper shall give more detailed explanation of cell+
architecture.
IV. Only 6 of 8 SPEs are available for use in reality, so the
performance data of cell with 8 working SPEs is dubious.
Aarul Jain
Nov 27, 2007, 11:44:28 PM11/27/07
to asucse520-fall-07-advanc...@googlegroups.com
Please see my comment regarding the critique
The paper is limited in scope. There are many useful benchmarks
that were omitted (e.g., Dhrystone, Whetstone). What about other types
of computing applications? The benchmarks chosen seem to be biased
towards Cell's architecture.
that were omitted (e.g., Dhrystone, Whetstone). What about other types
of computing applications? The benchmarks chosen seem to be biased
towards Cell's architecture.
-> The benchmarks you are mentioning are for general purpose computing. Moreover as far as I know, Dhrystone does not have floating point operations.
Further sparse matrix multiplication is a scientific computing kernel that might not be suited for cell architecture. Thus I think authors are unbiased towards choosing benchmarks kernels.
Thanks
-Aarul
Arizona State University, Tempe, US
Ph: 480-278-9230
0 new messages