Home

Processor Watch

Submitted by punkytse on December 5, 2002 - 12:25.


Seeking of Faster Processor

My daily work related to develope a system that distributes financial market data in real time.
This requires a system be capable to deliver messages as fast as possible. In the first system we developed,
we found that the performance result in not we expect. After weeks of software tuning and code refactoring, the performance improves,
but it could be further improved if some of the codes are rewritten. Anyway, we are better off now. But I always had
an argument that performance of today does not mean for tommorrow, as we can deploy a system with a
faster machine in tomorrow, as long as Moore's law hold true.

Our system is current developed and tested on Sun Entreprise 4500. This E4500 had 10 450MHz UltraSPARC II CPUs.
Since the hardward had been bought for 3-4 years, I doubted that the UltraSPARC is underperformed when compared
to the current advanced processors.


SPECint2000 Benchmark and Today's Processor

I recalled there is an organization that publishes CPU benchmark. It is SPEC. (See Resources
)
I searched the SPECint2000 CPU benchmark and it does provide awful lots of benchmark of including famous and infamous processors.
I downloaded the CVS Dump result, put it into the Excel spreadsheet and selected some CPUs to compare.
The benchmark I selected is related to integer performance. There is another SPECfp2000 benchmark that measures
floating point performance, but the integer performance is good enough to show some indications.

The selected CPUs include the following classes: SUN's UltraSPARC II & III, IBM's new POWER4, HP PA-RISC,
Alpha processor (previous by DEC, them Compaq), and Intel's Itanium. Each of these processors could be run under
UNIX (Solaris, AIX, HP-UX and Tru64) and each of these vendor also provides mid-range to high-end servers that our
system would likely be deployed to. Moreover, the IA-32 processors from Intel and AMD are also
included in my compare list, since they are known having higher processor clock speed, but mainly available for
PC and small to mid-range servers (1-4 CPUs). They provide very good indicator on how processor speed contributes to
performance. In contrast, it will provide strong evidence on Megahertz Myth. (See Resources

)

Please find below chart for the SPECint2000 result:

Please note that the above result is the base score provided by SPECint2000 benchmark result. The peak score is not considered.
SPECint2000 result provides more benchmark scores than I listed above. The above result is just a subset from SPECint2000 provided.
In the following sections, I will explain my findings for each class of processors by vendor.


SUN UltraSPARC

The SPECint2000 benchmark of UltraSPARC III processor is only provided by Sun Blade and Fire Series V880. Unfortunately,
there is no E10K, E12K and E15K benchmark available from SPEC. Even the benchmark for mid-range servers such as
Sun Fire 3800/4800/6800 is also missing. However, Sun web site provides information
about which processors are used for their servers (See Resources
).
It was found that The UltraSPARC III Cu 900 and 1050MHz processors are commonly used by Sun Fire Servers.

The most powerful UltraSPARC found from the benchmark result is UltraSPARC III Cu 1050 MHz (score 537).
However, this score lies in the middle performance level when comparing with others.
The performance of fastest UltraSPARC is somewhat disappointing.

I also included the result of Sun E4500 server, which runs 400 MHz UltraSPARC II CPU.
It is the server I work with everyday, and I would like to see the performance differences when comparing to some new and
advanced processors. Indeed, the result shows that it only obtains 198 scores, which is the worst performance from my list.
I search through the whole list of SPECint2000 result and it is only comparable to Pentium III 450 MHz (score 211) and
POWER3-II 333 MHz (score 196) found in IBM RS/6000 44P-170.


IBM POWER4

Last year, IBM released a successor of POWER3 processor, POWER4. POWER4 could now be found in the every
pSeries servers: high-end p690, mid-range p670 and entry-level p630.
From the benchmark result, 1.3HGHz POWER4 processor in high-end p690 Regatta (score 804) gives stunning
performance over UNIX servers of the same class such as HP's Superdome and SUN's E15K.
Even the less powerful p630 running 1GHz POWER4 scores 624.
It easily outperformes other entry-level UNIX servers such as Sun's V880 and HP's rp2470/5470.

In the comparison result, I also include p610 running the 333MHz POWER3-II processor. It obtains 313 scores in the
benchmark. This explains that POWER4 processor has a vast performance advancement over POWER3.

One more findings, From the System Performcen Report from IBM (See Resources
),
it shows that the new 1.45GHz POWER4+ in p650 scores 909, which surpass current POWER4 in p690 performance.
The change of fabrication process to 0.13 micron sees the performance boost without changing architecture in the
core that pays dividend to the performance while consumes less power. See Resources

for the
press release about POWER4+ processor from IBM.


HP PA-RISC and Alpha

HP provides two classes of UNIX servers, PA-RISC based HP-UX and Alpha based Tru64 UNIX.
The most recent and commonly available PA-RISC based servers are rp74xx running PA-8700 750MHz.
The HP-UX server selected in the listed result is the mid-range rp7400 that could host up to 8 CPUs.
It scores 520, which is just slightly slower that the fastest UltraSPARC 1050MHz.
The more powerful PA-8700+ is also included in the test result.
The chips running at 875 MHz could also be installed on the high-end Superdome that could host up to 64 CPUs.
Although there is no Superdome benchmark published from SPEC,
I found the c3750 workstation that runs the same chip. It shows scores 642.
This score exceeds the fastest UltraSPARC chips, and its performance lies between 1.0GHz and 1.1GHz POWER4 processor.

AlphaServer DS, ES and GS series for Tru64 UNIX provides a full-range of systems from entry-level to high-end.
Even with this wide coverage of systems, the performance of Alpha 21264C CPUs is also astonishing.
The slowest 1000MHz chips (scores 618) is comparable to 1GHz POWER4 (scores 624),
while the fastest 1250MHz chip (scores 845) runs faster than 1.3GHz POWER4 (scores 804). Among the fastest chips,
Alpha outperforms PA-RISC by about 30%.

After the merger with Compaq, HP continues supporting the AlphaServer previously provided by DEC and Compaq.
HP has been publishing the transition plan for Alpha to Itanium-based system (See Resources
).
It is believed that Alpha will go obsolete and be replaced by Itanium very soon.


IA64 Architecture - Itanium

HP is one of the vendor to provide Itanium-based UNIX system. As Itanium is co-developed with Intel, HP is the
first vendor to adopt Itanium in their server line. From the result, it is comfirmed that the performance of the
first version of Itanium (800 MHz) is worse (scores 379), the same fact told by the press media.

However, Itanium 2 is another world.
It shows some dramatic improvement in performance. From just increasing 25% clock speed, it contribute the performance
over 110%. The larger and faster on-chip memory cache and wider memory bandwidth in Itanium 2 makes a lot of difference
to its predecessor, Itanium (a.k.a. "Merced") (See Resources
).


IA32 Architecture - Pentium 4 and AMD Athlon XP

Speaking in single processor performance, the IA32 processors (Pentium 4 and AMD Athlon) still leading its way.
Comparing with Itanium 2, the fastest P4 2.8GHz runs 25% faster while AMD Athlon XP 2800+ (2250Mhz) has more
than 10% performance lead. However, IA32 processors are traditionally lacking of faster memory bandwidth,
larger memory support, poor SMP scalabilty. This makes it harder for baking IA32 processor into
highly reliable and scalable systems for high performance computing.

By looking at the clock speed, it was also known that Pentium 4 relies heavily on clock frequency to boost
performance. From the full list of result provided by SPEC, Pentium 4 2.2GHz processor (scores 808) gives
the same performance as Itanium 2 1000MHz. That means Pentium 4 is not a very efficient processor, it runs
lesser instructions per cycle than other processors. It also proved a strong evidence to Megahertz Myth.
On the other side, AMD Athlon is not far better than Pentium 4. Athlon XP 2600+ 2133MHz (scores 813) is closed
to Itanium 2 in performance.


Other processors

I have also included SGI R14000 and Fujitsu SPARC64 CPUs in the benchmark result. A little bit surprise from the
result, the fastest SGI R14k is not better than Sun UltraSPARC 1050MHz. And an interesting findings from
Fujutsu's PRIMEPOWER, which uses SPARC-branded CPU. Its SPARC64 CPU uses SPARC instruction
set and could run Solaris. From the information provided by SPEC, I found that the soon-to-be released PRIMEPOWER 900 runs a new
SPARC64 V processor at 1350Mhz. It provides the fastest performance (scores 747) among others servers made by Sun.

At the Microprocessor Forum held in October, IBM announced PowerPC 970. This chip is rumored to be used by Apple
for its next generation PowerMac and PowerBook. PowerPC 970 is the first 64-bit processor in PowerPC family and
is derived from POWER4 processor. The added SIMD/Vector Engine in its core gives performance boost for processing
graphics and multimedia data. IBM published a presentation of PowerPC 970 (See Resources

),
it shows the estimated performance of 937 scores in SPECint2000 in its 1.8GHz processor. Such performance hinted that
Apple could likely adopt IBM PowerPC 970 for its next generation products.

Another future processor to watch is AMD Opteron. AMD called x86-64 is 8th-Generation Processor Architecture.
At the Microprocessor Forum, AMD estimated that a server running a 2GHz Opteron achieved a score of 1202
(See Resources
). From technical point of view, Opteron is a nice
processor that is compatible to the existing x86-32 instruction set while provides relatively good performance.
If AMD makes Opteron sales well, this processor could likely be a processor for PC in future.


Conclusion

From SPECint2000 result, IBM POWER4 is the greatest RISC processor among others. Indeed, Alpha is a very good
processor, but considered that it will be phrased out and replaced by Itanium line, its future is dull.
Itanium 2, on the other hand, is a new finding in terms of its speed from the benchmark result.
Although, HP still does not produced mainframe class server like Superdome that based on Itanium,
I strongly believed that this young and premature processor using EPIC architecture will start drawing attention to
the public and is becoming a very competitive contender challenging RISC-based processors.
Dell's adoption of Itanium 2 chip in their server line is a very good sign.

But again, processor speed does not mean anything at all. A good system composes of fast CPU, board memory banwidth,
wide system bus, and efficent data transfer, but ultimately, it relies on a decent OS to run the target software.


Resources

» login to post comments