Spectre speed: Solaris/Sparc vs Linux PC

[Tim Roy]

We ran an evaluation between Monte Carlo run on Spectre on a Solaris 8 Sun
Blade 1000 900MHz vs Linux PC 2.8GHz and found there wasn't much of an
improvement in the simulation time (3min 15sec vs 2min 56sec).

The simulation was a Vbe sweep of a single NPN transistor, 100mv in 1mv
steps, for 1000 runs

Has anyone else tried this with similar results?

 

[fogh]

Tim Roy wrote:

We ran an evaluation between Monte Carlo run on Spectre on a Solaris 8 Sun
Blade 1000 900MHz vs Linux PC 2.8GHz and found there wasn't much of an
improvement in the simulation time (3min 15sec vs 2min 56sec).

The simulation was a Vbe sweep of a single NPN transistor, 100mv in 1mv
steps, for 1000 runs

Has anyone else tried this with similar results?


I believe the improvement comes from the intel C-compiler used on the

spectre binary, and it might not show up if you have such a small circuit.

 

[tritue]

If you you have a large simulation 1000+ transistors PC will be 2=6 time
faster and 10 time cheaper.
The IO speed is about the same in SUN / PC so if you try to save a lot of
signals, the run time will be almost the same due to IO.
ttt

"Tim Roy" <troy1@irf.com> wrote in message
news:S8SNb.3624$dd6.87@lakeread02...

We ran an evaluation between Monte Carlo run on Spectre on a Solaris 8
Sun
Blade 1000 900MHz vs Linux PC 2.8GHz and found there wasn't much of an
improvement in the simulation time (3min 15sec vs 2min 56sec).

The simulation was a Vbe sweep of a single NPN transistor, 100mv in 1mv
steps, for 1000 runs

Has anyone else tried this with similar results?

 

[Jay Lessert]

"Tim Roy" <troy1@irf.com> wrote in message news:<S8SNb.3624$dd6.87@lakeread02>...

We ran an evaluation between Monte Carlo run on Spectre on a Solaris 8 Sun
Blade 1000 900MHz vs Linux PC 2.8GHz and found there wasn't much of an
improvement in the simulation time (3min 15sec vs 2min 56sec).

The simulation was a Vbe sweep of a single NPN transistor, 100mv in 1mv
steps, for 1000 runs

Has anyone else tried this with similar results?

You're probably seeing cache effects; the tiny sim you're
running (1 transistor) probably fits entirely in the SB1000's
8MB cache, and doesn't fit entirely in the P4's 512KB cache.

US-III has a nice FPU as well.

Some year-old spectre benchmark results:

Sim: spectre, 70ns transient 2-run monte, pll_vco, 380 transistors
Wallclock CPU
Host CPU/Clock Time(s) Time(s) Factor
---- ------------- --------- ------- ------
compute5 USIII 750 893 872 1.00
tux2 Ath1800+ 1545 898 885 1.01
concorde P4 2400 583 574 0.65

Sim: spectre 600ns transient sim, dac, 750 transistors
Wallclock CPU
Host CPU/Clock Time(s) Time(s) Factor
---- --------- --------- ------- ------
compute3 USIII 750 192 187 1.00
tux2 Ath1800+ 1545 171 167 0.89
concorde P4 2400 128 102 0.66

Needless to say, we're not planning on buying any more
Sun Blades...

I *WISH* I could run Solaris x86 instead of Linux, though.

-Jay-

 

[Kholdoun TORKI]

3.2 CPU/Clock factor gaining 30% CPU time !!?

What about other issues, like multi-jobs launched,
multi-users, reliability ...?

=================
Kholdoun TORKI
http://cmp.imag.fr
=================

Jay Lessert wrote:

You're probably seeing cache effects; the tiny sim you're
running (1 transistor) probably fits entirely in the SB1000's
8MB cache, and doesn't fit entirely in the P4's 512KB cache.

US-III has a nice FPU as well.

Some year-old spectre benchmark results:

Sim: spectre, 70ns transient 2-run monte, pll_vco, 380 transistors
Wallclock CPU
Host CPU/Clock Time(s) Time(s) Factor
---- ------------- --------- ------- ------
compute5 USIII 750 893 872 1.00
tux2 Ath1800+ 1545 898 885 1.01
concorde P4 2400 583 574 0.65

Sim: spectre 600ns transient sim, dac, 750 transistors
Wallclock CPU
Host CPU/Clock Time(s) Time(s) Factor
---- --------- --------- ------- ------
compute3 USIII 750 192 187 1.00
tux2 Ath1800+ 1545 171 167 0.89
concorde P4 2400 128 102 0.66

Needless to say, we're not planning on buying any more
Sun Blades...

I *WISH* I could run Solaris x86 instead of Linux, though.

-Jay-



"Tim Roy" <troy1@irf.com> wrote in message
news:<S8SNb.3624$dd6.87@lakeread02>...

We ran an evaluation between Monte Carlo run on Spectre on a Solaris
8 Sun
Blade 1000 900MHz vs Linux PC 2.8GHz and found there wasn't much of an
improvement in the simulation time (3min 15sec vs 2min 56sec).

The simulation was a Vbe sweep of a single NPN transistor, 100mv in 1mv
steps, for 1000 runs

Has anyone else tried this with similar results?

 

[gennari]

CPU speed isn't the only thing to look at. Ultrasparcs can do more work per
CPU cycle because I assume the CPU core datapath is longer and more complex.
For example, an Ultrasparc likely takes fewer actual CPU cycles to perform a
floating-point division than a Pentium4. Factors such as on-chip cache size
and architecture bit depth (32 vs. 64) also are important.

If you want to see the real benchmarking results then take a look at
http://www.specbench.org/cpu2000/results/cpu2000.html

The Itanium2 is even faster/MHz than the Ultrasparc.

Frank

"Kholdoun TORKI" <Kholdoun.Torki@imag.fr> wrote in message
news:4009610C.5020509@imag.fr...

3.2 CPU/Clock factor gaining 30% CPU time !!?

What about other issues, like multi-jobs launched,
multi-users, reliability ...?

=================
Kholdoun TORKI
http://cmp.imag.fr
=================

Jay Lessert wrote:

You're probably seeing cache effects; the tiny sim you're
running (1 transistor) probably fits entirely in the SB1000's
8MB cache, and doesn't fit entirely in the P4's 512KB cache.

US-III has a nice FPU as well.

Some year-old spectre benchmark results:

Sim: spectre, 70ns transient 2-run monte, pll_vco, 380 transistors
Wallclock CPU
Host CPU/Clock Time(s) Time(s) Factor
---- ------------- --------- ------- ------
compute5 USIII 750 893 872 1.00
tux2 Ath1800+ 1545 898 885 1.01
concorde P4 2400 583 574 0.65

Sim: spectre 600ns transient sim, dac, 750 transistors
Wallclock CPU
Host CPU/Clock Time(s) Time(s) Factor
---- --------- --------- ------- ------
compute3 USIII 750 192 187 1.00
tux2 Ath1800+ 1545 171 167 0.89
concorde P4 2400 128 102 0.66

Needless to say, we're not planning on buying any more
Sun Blades...

I *WISH* I could run Solaris x86 instead of Linux, though.

-Jay-



"Tim Roy" <troy1@irf.com> wrote in message
news:<S8SNb.3624$dd6.87@lakeread02>...

We ran an evaluation between Monte Carlo run on Spectre on a Solaris
8 Sun
Blade 1000 900MHz vs Linux PC 2.8GHz and found there wasn't much of an
improvement in the simulation time (3min 15sec vs 2min 56sec).

The simulation was a Vbe sweep of a single NPN transistor, 100mv in 1mv
steps, for 1000 runs

Has anyone else tried this with similar results?

 

[Jay Lessert]

Kholdoun TORKI <Kholdoun.Torki@imag.fr> wrote in message news:<4009610C.5020509@imag.fr>...

3.2 CPU/Clock factor gaining 30% CPU time !!?

Yes. Remember that analog circuit simulators play to US-III's
strengths (compared to P4): big cache and good FPU performance.

On cache-friendly applications (DRC/LVS is the classic) the
number goes from 30% to 50-60% (2X to 2.5X faster).

On big Verilog jobs (no FPU, still a cache-buster though) it's
somewhere in between.

What about other issues, like multi-jobs launched,

Why would I run multiple jobs in parallel on a single-CPU
box? Do I *want* to run slower? Regardless of the hardware,
you want 1 job/cpu average, and you use SGE (or something like
it) to manage job queues, etc. Also remember, I can buy
4 x86 boxes for every 1 US-III box.

multi-users

In a multi-user, load average >>1 environment Solaris behaves
more gracefully than any Linux system I've ever run. In an
EDA environment, you want to avoid load average >>1, though.

reliability ...?

On good quality x86 hardware, I've been very satisfied with
with my RH8.0 uptimes. 1-year uptimes are normal, if the
power stays on.

On bad hardware, uptimes are bad. We've been buying
Dell 400SC boxes lately, and are very pleased with those

-Jay-

 

[Harald Neubauer]

Here are my results running a spectre job on different machines..
We are running some Double Xeon 368 machines with Redhat 7.2 and the
uptime is about 1 year (since they were booted). They run multi user
multi jobs.
The double Xeon with 1 GB is about 1500 Eur. Prices for Suns in
equivalent computing area are well know much above.
We presently do not buy any Suns as eda workstations any more...

name Clk OS Proz s simulation
cx013 3000 linux 8.0 2Xeon 276
100%
cx011 2800 linux 8.0 P IV 335 121%
cx008 2400 linux 7.2 2Xeon 383 139%
2400+ linux 8.0 Amd 424 (454 Eur low cost System) 154%
tw038 900 Blade1000 2Sun 432 157%
cx009 2000 linux 7.2 2Xeon 462 167%
ux033 1200 linux 7.2 P III 663 Laptop Socket 240%
ux033 1200 linux 7.2 P III 861 Laptop accu 312%
cw061 450 Ultra 250 2Sun 1140 413%
cw075 500 Blade 100 Sun 1250 453%
cw057 300 Ultra 10 Sun 1880 681%
cw032 170 Ultra 1 Sun 2390 866%


Regards, Harald