Pop quiz. You're in desperate need of making Java fly and have the coin to prove it. Do you shell out on IBM's new 4.7GHz Power6 eye candy or keep on replacing Unix kit with x86 systems? For the moment, it looks like you will want to saunter into the 4GHz realm if you're a midrange server type. IBM has just released SpecJbb …
interesting indeed, if you read what it says.......
loaded the specjbb2005 results since 1 jan 2007
some interesting stuff indeed.
I added bops/core column, because it tells you about scaling. And hence how much IBM needs to worry. or not.
the power6 at 2 cores does 44045 bops ( 27% faster than the speediest X86, by the way )
the power6 at 8 cores does 43343 bops
the power6 at 16 cores does 43248 bops
pretty much linear performance, indeed.
Now for the shivering......
the closest thing to that power6 is 3 Xeon boxes, 4 cores and 8 cores from Fujitsu ( the same guys that beat up Sun for many years now.... on SPARC ....)
4-way = PRIMERGY BX620 S3, Intel Xeon processor 5160, 3.0 GHz
8-way = PRIMERGY BX620 S3, Intel Xeon processor X5355, 2.66 GHz
16-way = PRIMERGY RX800 S3, Intel Xeon processor 7150N, 3.50 GHz
the xeon at 4 cores does 34587 ( = 100%)
the xeon at 8 cores does 27669 ( = 80%)
the xeon at 16 cores does 21041 ( = 61%)
those power6 chip systems delivers near linear performance, ladies and gentlemen....... the Xeon doesnt, and you can imagine what the numbers of a 32-way Xeon will look like.
Yes we need to worry. About IBM having no competition that can get close.
if you care, you can download a piece of my XLS with the bops/core columns in it. or get it from www.spec.org yourself.
UltraSPARC T1 still the most powerful Java processor
A single-socket 65nm, POWER6 at 4.7 GHz (and 750 million transistors!) scores 88,089 SPECjbb bops. The latest UltraSPARC T1 results for the new T6300 blade with a 90nm, 1.4 GHz, 300 million transistor processor scores 96,523 SPECjbb2005 bops.
Given Java appservers are designed to scale horizontally and a chassis of ten T6300s looks like a killer WebLogic or WebSphere platform.
And Sun has the 65nm Niagara 2 waiting in the wings, which should double performance, and Sun has promised will also be available for the Blade 6000.
What's a processor?
In reply to "Ultrasparc T1 still the most powerful Java processor" the author ignores that the T1 has 8 cores per socket, the POWER6 has two. So, per core the POWER6 does 265% better with only 235% of the clock cycles. That might be where some of the extra transistors go, not to mention that the POWER6 has more than 8Mb cache on board to the T1's 3Mb. Plus built-in virtualisation, error detection and correction... comparing apples and oranges doesn't even come close, does it?
How much power for money?
The amount of cores is rather unimportant, as decision maker I'm more interested about the bang for buck aspect. How much is the Sun T1 in above configuration, and how much is the IBM box? I don't buy cores but performance.
Chip to Chip, UltraSPARC T1 beats Power6 on Java
In regards to the comments that Power6 cores are faster than the UltraSPARC T1's cores, thats like saying my cylinder in my 4-cylinder engine is faster that your cylinder in your 8-cylinder engine (its the engine performance that matters)!
Who cares how fast are the cores or how much cache the chip needs for Java? The bottom line is how fast the CHIP is and for right now, the UltraSPARC T1 blows away the Power6 at its server melting 4.7GHz. How much wattage does the Power6 require compared to the UltraSPARC T1's <80Watts? How much is it going to cost you to run? No one knows-IBM sure doesn't want you to know otherwise it will be in all the press materials.
Number of cores does indeed matter
-- if you are the guy who has to write the check for the software license fees. These tend to be based on the number of cores the software executes on. When you compare the software charges for the same solution running on two IBM cores or on eight Sun cores, you *will* ask your technology people to think again about recommending the Sun solution.
Even with licencing costs the T1 still wins.
For example Oracle:
Power 6 - 0.75 licence per core = 1.5. Round up (Oracle say you must) = 2 licences
UltraSPARC T1 - 0.25 licence per core = 2 licences
T1 wins hands down on performance, cost, power consumption.. and the T2 is just a few months away!
Performance of the CHIP is irrelevant
It's the performance of the SERVER that counts. The same is true of the power draw - IBM might not publish the chip power draw but then a chip is just so much semi-precious metal and sand and is useless without the electronics around it. If your look at the power rating for the whole server (an 8-core T2000 is max 400W for 74,365 BOPS = 185BOPS/Watt, an 8-core p570 is max 1400W for 346,742 BOPS = 247BOPS/Watt) there's a 33% POWER6 advantage. Of course, it's still an apples and oranges comparison at present because you're comparing a 4-socket midrange server with a single-socket entry server.
But it still comes back to that software licensing. If you pay per-core licence on your software the POWER6 will likely be streets ahead on TCO as the hardware and environmental cost pales into insignificance next to software. If you don't it's up to you whether a theoretical 33% more performance per watt is worth it. Obviously it's closer for a blade server than my standalone server example but also harder to gauge the watts and an even poorer comparison in class of machine.
Who mentioned Oracle?
This is an article about JAVA benchmarks. Last I looked, Oracle give their java app server away at no extra cost when you pay for the rest of their apps (usually per seat) and that licensing calc only applies to the database. What makes anyone think the performance argument is the same for Oracle database, which scales in an entirely different way?
Don't be too worried, RG
The 16-core score is from a (still current, but) older generation Intel processor codenamed Tulsa based on the Netburst architecture, whereas the 4- and 8-core results are both based on the newer Core microarchitecture. On that basis, you wouldn't expect to see linear scalability per core.
Tigerton will hit the streets next quarter - the Core microarchitecture replacement of Tulsa - then you can do a more realistic "scalability per core" test.
Intel soon switches to a 45nm fab process - boosting performance and lowering power consumption across the range. Then there's the successor to the Core microarchitecture just around that corner...I think Intel is giving IBM a pretty good run for its money.
Paul, the sad part is that IBM is still the only manufacturer actually in keeping with what is commonly understood to be Moore's law (doubling of processing power every 18 months). I think it's funny that you're telling RG that IBM should watch out for Intel's regurgitated Pentium 3. Intel still has some very serious things to put into consideration:
Itanium has been a loss-leader in the CPU market for years, not to mention the worst selling series of cpus currently on the market.
Intel keeps bumping FSB speeds, with negligible gains due to outdated memory architectures.
Intel is still worst-in-class across the board (since they sold off their ARM business) for performance-per-watt.
Intel's Core Duo had some significant issues with XOR operations, causing software raid5 to be mostly unusable.
To date, Intel has invested more in Itanium than it could ever hope to recoup via chip sales, despite horrendous pricing per cpu on ia64 systems.
I personally applaud IBM for strategic partnerships with AMD on manufacturing fabs, and showing the world that Steve Jobs is a moron for going Intel on Apple's products.