| Attached are the final results. Note that they improved slightly from
the results referenced in .0
A web-ified edition of the attached (with appropriate hyperlinks) will
be posted to
http://tlg-www.zko.dec.com/~henning/TidBits.html
SUBJ: Final SPEC CPU95 for Rawhide/466
Summary:
- Leadership has been achieved in 5 out of 8 categories:
I. SPECint95 second place
II. SPECfp95 (4 cpu) leadership
III. SPECfp95 (2 cpu) leadership
IV. SPECfp95 (1 cpu) fourth place
V. SPECint_rate95 (4 cpu) leadership
VI. SPECint_rate95 (2 cpu) leadership
VII. SPECfp_rate95 (4 cpu) second place
VIII. SPECfp_rate95 (2 cpu) leadership
- We have an opportunity to ask HP some embarrassing questions
about their Runway bus.
Here are the current results compared to the competitors' results as posted
to the SPEC web site, http://www.specbench.org, through February 10, 1997.
I. Top 10 SPECint95
Rank System SPECint_base95 SPECint95
---- ------------------------- -------------- ---------
1 AlphaStation 500/500 12.6 15.0
..new...AlphaServer 4100 5/466...........12.1............14.1.....
..new...AlphaServer 4000 5/466...........12.1............14.1.....
2 AlphaServer 8200 5/440 11.2 13.6
3 AlphaServer 8400 5/440 11.2 13.6
4 AlphaStation 500/400 9.77 12.3
5 AlphaServer 4000 5/400 10.1 12.1
6 AlphaServer 4100 5/400 10.1 12.1
7 HP 9000 Model K460EG 10.8 11.8
8 HP 9000 Model K460 10.8 11.8
9 HP 9000 Model C180-XP 10.8 11.8
10 AlphaServer 1000A 5/400 9.61 11.5
The AlphaServer 4x00 5/466 has a solid lead over HP, 19% for SPECint95
and 12% for SPECint_base95.
II. Top 10 SPECfp95 (4 CPU)
SPECfp_
Rank System CPUs base95 SPECfp95
---- ------------------------ ---- ------- --------
.new..AlphaServer 4100 5/466...........4.......27.8...........36.1.....
1 AlphaServer 4100 5/400 4 26.6 33.4
2 AlphaServer 8200 5/350 4 21.9 27.9
3 AlphaServer 2100A 5/375 4 20.3 27.3
4 AlphaServer 8400 5/300 4 11.7 25.9
5 AlphaServer 4100 5/300 4 18.9 21.8
6 AlphaServer 2100A 5/300 4 14.9 19.4
7 AlphaServer 2100 5/300 4 9.00 19.2
8 AlphaServer 4100 5/300E 4 17.7 18.4
9 AlphaServer 2100 5/250 4 14.0 17.7
10 origin200 4 14.5 15.6
III. Top 10 SPECfp95 (2 CPU)
SPECfp_
Rank System CPUs base95 SPECfp95
---- ------------------------ ---- ------- --------
.new..AlphaServer 4100 5/466...........2.......22.2...........26.5.....
.new..AlphaServer 4000 5/466...........2.......22.2...........26.5.....
1 AlphaServer 4000 5/400 2 20.7 24.2
2 AlphaServer 4100 5/400 2 20.7 24.2
3 AlphaServer 2100A 5/375 2 16.8 20.7
4 AlphaServer 2000 5/375 2 16.8 20.7
5 AlphaServer 8200 5/350 2 17.0 20.3
6 AlphaServer 8400 5/300 2 11.9 18.1
7 AlphaServer 4100 5/300 2 15.2 16.7
8 AlphaServer 4100 5/300E 2 14.2 15.1
9 Ultra 2 Series 2200 2 13.3 14.7
10 AlphaServer 2100 5/300 2 9.02 14.0
At both 2 and 4 CPUs, the AlphaServer 4x00 5/466 has broken the SMP SPECfp
records established by the AlphaServer 4x00 5/400, and with these new
results 10 out of the top 10 systems will be Alpha systems.
IV. Top 10 SPECfp95 (1 cpu)
SPECfp_
Rank System CPUs base95 SPECfp95
---- ------------------------ ---- ------- --------
1 AlphaStation 500/500 1 18.3 20.4
2 HP 9000 Model K460 1 18.3 20.2
3 HP 9000 Model K460EG 1 18.3 20.2
..new...AlphaServer 4000 5/466.........1........17.8...........19.2.....
..new...AlphaServer 4100 5/466.........1........17.8...........19.2.....
4 HP 9000 Model K450 1 17.0 18.8
5 HP 9000 Model C180-XP 1 17.2 18.7
6 RISC System/6000 595 1 15.4 17.6
7 AlphaServer 4000 5/400 1 16.0 17.2
8 AlphaServer 4100 5/400 1 16.0 17.2
9 AlphaServer 8200 5/440 1 16.0 17.0
10 AlphaServer 8400 5/440 1 16.0 17.0
With the new AlphaServer 4x00 5/466 results, Alpha gains 3 out of the
top 5 single-CPU SPECfp95 results, displacing two HP systems in the top
5. The AlphaServer 4x00 5/466 trails the HP K460 by 5% on single-CPU
SPECfp95 (but leads it by 19% on SPECint95 - see table I.).
NOTE: The above table was produced by doing a query for
#cpu=1 at the SPEC website; therefore it leaves out the
SGI origin2000, which is listed as:
SPECfp_
Number of CPU(s) base95 SPECfp95
----------------- ------- --------
origin2000 "32 (1 CPU used)" 17.4 19.0
If one were to include the orgin2000 in table IV, it would
appear just below the new AlphaServer 4x00 results.
V. Top 10 SPECint_rate95 (4 cpu)
SPECint_ SPECint_
Rank System CPUs rate_base95 rate95
---- ------------------------ ---- ---------- --------
.new..AlphaServer 4100 5/466...........4.......419............485.....
1 AlphaServer 4100 5/400 4 349 422
2 HP 9000 Model K460 4-CPU 4 380 418
3 AlphaServer 2100A 5/375 4 336 397
4 RM400 Model C90 4 375 375
5 RM400 Model C90 4 375 375
6 HP 9000 Model K450 4-CPU 4 342 375
7 origin200 4 278 304
8 superserver J654i 4 292 292
9 AlphaServer 4100 5/300 4 255 278
10 AlphaServer 8400 5/300 4 261 261
VI. Top 10 SPECint_rate95 (2 cpu)
SPECint_ SPECint_
Rank System CPUs rate_base95 rate95
---- ------------------------ ---- -------- -----------
..new.AlphaServer 4100 5/466...........2.......210............250......
..new.AlphaServer 4000 5/466...........2.......210............250......
1 HP 9000 Model K460 2-CPU 2 192 212
2 AlphaServer 4000 5/400 2 180 211
3 AlphaServer 4100 5/400 2 180 211
4 AlphaServer 2000 5/375 2 170 204
5 AlphaServer 2100A 5/375 2 170 204
6 HP 9000 Model K450 2-CPU 2 172 189
7 HP 9000 Model D370 2-CPU 2 168 184
8 HP 9000 Model D270 2-CPU 2 168 184
9 superserver J652i 2 151 151
10 AlphaServer 4100 5/300 2 129 144
Both 2 and 4cpu SPECintrate_95 are handily won by the AlphaServer 4100
5/466.
VII. Top 10 SPECfp_rate95 (4 cpu)
SPECfp_ SPECfp_
Rank System CPUs rate_base95 rate95
---- ------------------------ ---- ----------- -------
1 Power Challenge 4 449 471
.new..AlphaServer 4100 5/466...........4.......437............466.....
2 origin200 4 405 446
3 AlphaServer 4100 5/400 4 420 439
4 AlphaServer 2100A 5/375 4 328 438
5 AlphaServer 8200 5/300 4 400 420
6 AlphaServer 8400 5/300 4 400 420
7 HP 9000 Model K460 4-CPU 4 368 398
8 HP 9000 Model K450 4-CPU 4 356 383
9 AlphaServer 4100 5/300 4 294 303
10 AlphaServer 2100 5/300 4 246 251
The AlphaServer 4100 5/466 shows some key strengths in 4-CPU SPECfp_rate95.
It handily beats SGI's much hyped origin 200 system, and almost catches
SGI's Power Challenge with 195 MHz R10000 and 8-way memory interleaving.
But if the term "mid-range" is defined to mean systems that hold up to
4-CPUs, the latter is clearly not a mid-range system. It should also be
noted that various additional high-end servers would appear at or near the
top of this list if they were to be measured with only 4 cpus, including
both the origin 2000 and the AlphaServer 8400 5/440.
More importantly, this result tops the HP K460 by 17%. One might ask,
"What's wrong with the K460, HP? It seems your CPU might be sort
of OK, since EV56 and PA-8000 each take one round in the single-CPU
SPECint and SPECfp race (see tables I and II). But what happens
with scaling? Why does the K460 drop off so much compared to the
AlphaServer 4100? Is it possible that your Runway bus with its
alleged theoretical 960 MB/sec peak can't deliver more than a small
fraction of that bandwidth in real applications? And by the way
why haven't you ever published a McCalpin STREAMS bandwidth number
for a K460?"
VIII. Top 10 SPECfp_rate95 (2 cpu)
SPECfp_ SPECfp_
Rank System CPUs rate_base95 rate95
---- ------------------------ ---- ----------- -------
..new..AlphaServer 4100 5/466...........2.......278............301.....
..new..AlphaServer 4000 5/466...........2.......278............301.....
1 HP 9000 Model K460 2-CPU 2 267 297
2 HP 9000 Model K450 2-CPU 2 252 279
3 AlphaServer 4000 5/400 2 257 275
4 AlphaServer 4100 5/400 2 257 275
5 HP 9000 Model K420 4-CPU 2 264 275
6 AlphaServer 2100A 5/375 2 244 258
7 AlphaServer 2000 5/375 2 244 258
8 Power Challenge 2 229 241
9 AlphaServer 8200 5/300 2 205 215
10 AlphaServer 8400 5/300 2 205 215
Even at 2 CPUs, the AlphaServer 4100 5/466 shows better scaling than the
HP K460.
Finally, comparing Rawhide/400 vs. Rawhide466:
Metric #cpus Rawhide/400 Rawhide/466 Ratio
------------------- ----- ----------- ----------- -----
SPECint95 1 12.1 14.1 1.17
SPECint_base95 1 10.1 12.1 1.20
SPECfp95 4 33.4 36.1 1.08
SPECfp95 2 24.2 26.5 1.10
SPECfp95 1 17.2 19.2 1.12
SPECfp_base95 4 26.6 27.8 1.05
SPECfp_base95 2 20.7 22.2 1.07
SPECfp_base95 1 16.0 17.8 1.11
SPECint_rate95 4 422 485 1.15
SPECint_rate95 2 211 250 1.18
SPECint_rate95 1 109 127 1.17
SPECint_rate_base95 4 349 419 1.20
SPECint_rate_base95 2 180 210 1.17
SPECint_rate_base95 1 90.6 109 1.20
SPECfp_rate95 4 439 466 1.06
SPECfp_rate95 2 275 301 1.09
SPECfp_rate95 1 155 173 1.12
SPECfp_rate_base95 4 420 437 1.04
SPECfp_rate_base95 2 257 278 1.08
SPECfp_rate_base95 1 144 160 1.11
Note that a 17% improvement in clock speed (466 vs. 400) has translated
into a similar improvement for most integer metrics, and sometimes even
more than 17%, due to compiler and tuning improvements.
For floating point benchmarks the improvement is much less than clock
speed, especially as the number of CPUs increases, because SPECfp95
depends heavily on main memory bandwidth.
|