Latest server results

[Viditor]

I am reposting work from someone on another site, but all can be checked here...
For those of you who don't know or care anything about servers or workstations, please ignore this thread as it has absolutely nothing to do with desktop performance.


SPECint_rate:System....................Configuration...............................Base.......Peak
DL580 G3 (3.66GHz, Cranford)......2 cores, 2 chips, 1 core/chip*...30.5.......30.8
DL580 G3 (3.333GHz, Potomac)....2 cores, 2 chips, 1 core/chip*...36.5.......36.6
DL385 (Opteron 280).......................2 cores, 1 chip, 2 cores/chip....35.7.......37.7
ML350 G4p (3.6GHz, Irwindale)......2 cores, 2 chips, 1 core/chip*...40.4.......40.5
DL385 (Opteron 254).......................2 cores, 2 chips, 1 core/chip....39.2.......43.7
ML570 G3 (3.0GHz, Paxville MP)....4 cores, 2 chips, 2 cores/chip*..56.9......57.2
DL580 G3 (3.66GHz, Cranford)......4 cores, 4 chips, 1 core/chip*...57.1......57.4
RX300 S2, 2.80 GHz Paxville DP....4 cores, 2 chips, 2 cores/chip*..59.2......59.2
DL385 (Opteron 280).......................4 cores, 2 chips, 2 cores/chip....71.0.....75.0
DL585 (Opteron 880).......................4 cores, 2 chips, 2 cores/chip....71.1.....75.1
DL580 G3 (3.333GHz, Potomac).....4 cores, 4 chips, 1 core/chip*...72.6.....72.8
DL585 (Opteron 854).......................4 cores, 4 chips, 1 core/chip....80.9......85.5
ML570 G3 (3.0GHz, Paxville MP).....8 cores, 4 chips, 2 cores/chip*..104.....105
DL585 (Opteron 880).......................8 cores, 4 chips, 2 cores/chip...136.....144

SPECfp_rate:System..........................Configuration.........................Base......Peak
DL580 G3 (3.66GHz, Cranford)......2 cores, 2 chips, 1 core/chip*...22.6......23.1
ML570 G3 (3.0GHz, Paxville MP).....2 cores, 1 chip, 2 cores/chip*...23.9......24.5**
DL580 G3 (3.333GHz, Potomac)....2 cores, 2 chips, 1 core/chip*...28.7......29.4
ML350 G4p (3.6GHz, Irwindale)......2 cores, 2 chips, 1 core/chip*...31.5......32.3
DL385 (Opteron 280).......................2 cores, 1 chip, 2 cores/chip....30.8......33.7
DL385 (Opteron 254).......................2 cores, 2 chips, 1 core/chip....41.7......45.7
ML570 G3 (3.0GHz, Paxville MP).....4 cores, 2 chips, 2 cores/chip*.31.4......32.1
DL580 G3 (3.66GHz, Cranford)......4 cores, 4 chips, 1 core/chip*...34.9......35.6
RX300 S2, 2.80 GHz Paxville DP....4 cores, 2 chips, 2 cores/chip*..42.5......42.5
DL580 G3 (3.333GHz, Potomac)....4 cores, 4 chips, 1 core/chip*...52.7......53.9
DL385 (Opteron 280)......................4 cores, 2 chips, 2 cores/chip...60.8......66.7
DL585 (Opteron 880)......................4 cores, 2 chips, 2 cores/chip...61.6......67.1
DL585 (Opteron 854)......................4 cores, 4 chips, 1 core/chip....78.7......86.0
ML570 G3 (3.0GHz, Paxville MP)....8 cores, 4 chips, 2 cores/chip*..47.2.....48.2
DL585 (Opteron 880)......................8 cores, 4 chips, 2 cores/chip...118.....128
*Hyperthreading disabled.
*HP did not post a two-core, single-chip Paxville MP SPECint_rate score.

 

[Markfw]

OK, If I read that correctly, the 280 kicks a$$, correct ??

 

[Viditor]

Originally posted by: Markfw900
OK, If I read that correctly, the 280 kicks a$$, correct ??

True...but the real key is how much better the Opterons get as you scale the number of chips. This is where HT truly shines!

Edit: Let me add to that...
Many of the Intel pundits are quick to point out that Intel still has the majority of server shipments. This is absolutely correct! The reason is that at the low end, Xeon is still quite competitive (not better, but still in the ballpark). However, where AMD is really making massive inroads is the high-end/high-profit servers. It is a strange twist of fate that at 4P and higher, Intel isn't really worth a look anymore. Even with Itanium, Opteron is more than competitive and it costs less to purchase and maintain (not to mention the price of software conversion for EPIC).
This is not to say that AMD has no competitor...IBM's Power 5+ is more than competitive!

System.......................................Configuration................................Base......Peak
DL585 (Opteron 880)...................4 cores, 2 chips, 2 cores/chip.........66.5......73.9
IBM p5 550 (1900 MHz, 4 CPU)...4 cores, 2 chips, 2 cores/chip*..........77.1......78.5
DL585 (Opteron 854)...................4 cores, 4 chips, 1 core/chip..........74.5......83.9
IBM p5 550Q (1500 MHz, 8 CPU).8 cores, 4 chips, 2 cores/chip*.........122......124
DL585 (Opteron 880)...................8 cores, 4 chips, 2 cores/chip........125......140

SPECfp_rate:System....................Configuration................................Base......Peak
DL585 (Opteron 880)...................4 cores, 2 chips, 2 cores/chip.........69.4.......75.0
DL585 (Opteron 854)...................4 cores, 4 chips, 1 core/chip...........89.6.......98.5
IBM p5 550 (1900 MHz, 4 CPU)...4 cores, 2 chips, 2 cores/chip*..........129........133
DL585 (Opteron 880)...................8 cores, 4 chips, 2 cores/chip.........131........144
IBM p5 550Q (1500 MHz, 8 CPU).8 cores, 4 chips, 2 cores/chip*..........174........178
*SMT on

 

[Markfw]

Thats what I thought ! (and actually expected based om known facts and results)

Intel is hurting big time !

 

[Viditor]

Originally posted by: Markfw900
Thats what I thought ! (and actually expected based om known facts and results)

Intel is hurting big time !

Yes and no...remember that the low end is by far the vast the majority of server marketshare, it is also the most price sensitive. If Intel can get their power/performance down sufficiently (which many of us expect is possible with Woodcrest), they can be quite competitive in this market. However, it won't help them in the high-end/enterprise area. The pushing back of their CSI platform to 2009 was far more important than most realise...keep in mind that the same thing happened to AMD with the first Hammer chips.
AMD will (IMHO) share the high-end with IBM for the next 4 years, and Intel will not really be a contender until at least then...

 

[Accord99]

Spec_rates are useful but not particularly similar to most types of apps used in the enterprise server market. TPC-C is a more relevant benchmark in the 4P+ market. In this the 4P IBM x366 with Xeon Cranfords outscores the 4P DL585 and the 8P x460 outscores the 4 socket, dual-core DL585. Unlike Opteron systems, the IBM x4 series can scale up effectively to 32 sockets. Opterons have trouble past 4 sockets, they need the Horus chipset. In these systems, the chipset is as or even more important than the processors.

http://tpc.org/tpcc/results/tpcc_result...nt=false&orderby=submitted&sortby=desc

 

[Viditor]

Originally posted by: Accord99
Spec_rates are useful but not particularly similar to most types of apps used in the enterprise server market. TPC-C is a more relevant benchmark in the 4P+ market. In this the 4P IBM x366 with Xeon Cranfords outscores the 4P DL585 and the 8P x460 outscores the 4 socket, dual-core DL585. Unlike Opteron systems, the IBM x4 series can scale up effectively to 32 sockets. Opterons have trouble past 4 sockets, they need the Horus chipset. In these systems, the chipset is as or even more important than the processors.

http://tpc.org/tpcc/results/tpcc_result...nt=false&orderby=submitted&sortby=desc

You need to read through the full disclosures...
The HP DL585-G1/2.4 GHz is from last December, sells for $360,470, and has TPC-C of 123,027
The IBM 366 c/s is from this April, sells for $993,179, and has a TPC-C of 141,504

A 13% increase in TPC-C for triple the price isn't really a step up...

As far as scaling to a 32-way system, Horus is expected to be appearing by the end of this year (32-way server chipset for Opteron). This is even more assured because Phil Hester (IBM fellow and former head of it's PowerPC division, also co-founder of Newisys) is now the CTO at AMD...

 

[Accord99]

Originally posted by: Viditor
You need to read through the full disclosures...
The HP DL585-G1/2.4 GHz is from last December, sells for $360,470, and has TPC-C of 123,027
The IBM 366 c/s is from this April, sells for $993,179, and has a TPC-C of 141,504

A 13% increase in TPC-C for triple the price isn't really a step up...

Most of the price difference has to do with the storage systems, IBM uses fibre channel which is much more expensive, performs similar to SCSI but are better for manageability. The x366 server barebones is cheaper.

 

[Viditor]

Originally posted by: Accord99

Originally posted by: Viditor
You need to read through the full disclosures...
The HP DL585-G1/2.4 GHz is from last December, sells for $360,470, and has TPC-C of 123,027
The IBM 366 c/s is from this April, sells for $993,179, and has a TPC-C of 141,504

A 13% increase in TPC-C for triple the price isn't really a step up...

Most of the price difference has to do with the storage systems, IBM uses fibre channel which is much more expensive, performs similar to SCSI but are better for manageability. The x366 server barebones is cheaper.

Fair enough, but there are no submissions with that configuration. TPC-C is a total system measurement (# of transactions) based on real-world design, so it is (as you say) a better measurement of how a complete system will perform...but not the CPU or even the platform. SPEC is a much better guage at that level...

 

[Accord99]

For a single CPU, at higher processor counts, SpecFP_rate becomes only a memory bandwidth test. Many Spec_Int components can run entirely in the cache memory of modern processors. The rate tests just run multiple instances of the standard Spec tests, so it's the poster child for NUMA.

 

[Viditor]

Originally posted by: Accord99
For a single CPU, at higher processor counts, SpecFP_rate becomes only a memory bandwidth test. Many Spec_Int components can run entirely in the cache memory of modern processors. The rate tests just run multiple instances of the standard Spec tests, so it's the poster child for NUMA.

Fair enough (I didn't say perfect, only better...).

Try this though...go back to the TPC-C results you posted...sort by price/tpmC...then look at the first system that exceeds 100k tpmC. Notice how far ahead that DL585 is over the rest of the competition? That's exactly what most companies will be looking at...

 

[Viditor]

BTW...I don't know if anyone noticed, but the HP DL385 (with a single dual core Opteron) performed just as well as the ML570 (with 2 dual core Paxvilles) in Spec_FP...

 

[gwag]

wonder how a quad Power mac would do....