Power7 v Power6 - it's all about the cache IBM is launching the first of its Power7-based systems today, and the company thinks that the innovations inside the Power7 processor are going to give it a leg-up on the competition in terms of capacity, throughput, and energy-efficiency. But how do those Power7 processors stack up to the existing Power6 and Power6+ processors …
At 1.2 billion transistors on a 45nm process, this chip will be significantly smaller than the 2 billion transistor, 65nm, IA64 beast coming out today.
With 8 cores at 3GHz+ and 4 way SMT it surely will also be significantly faster.
Unless Intel bring something amazing to the table, but getting Itanium to 2GHz would be a major achievement.
On 40nm Cypress is 340mm^2 for 2b transistors, so making a leap of numeric assumptions I guess that Power7 is well under 300mm^2. Yeah, you know that's good to go for a PlayStation4...
Timothy Prickett Morgan writes, "Sparc T 64-threaded T2 and T2+... quad-core, eight-threaded Tukwilas... the Power7 chip has 32 threads"'
It is nice to see the trail which first generation OpenSPARC T1 had blazed with 32 threads is being followed by IBM and Intel, both applying different technology to compete with Sun's second and third generation 64 threaded processors.
Timothy Prickett Morgan writes, "The effect of this eDRAM on the Power7 design, and its performance, is two-fold. First, by adding the L3 cache onto the chip..."
The use of embedded DRAM, to reduce transistors, squeeze more cores, and reduce latency was a great idea, even with the refresh logic added onto the chip!
Every benefit comes with a drawback.
The use of Static RAM has been traditionally beneficial to the chip manufacturers, since they could get fast and regular access to the memory cells, without having to wait for a slow refresh signal to propagate across the RAM. It is interesting that no one (and I mean NO ONE) is talking about the impact of performance for the CPU cores needing to wait for refresh on the eDRAM.
I wonder what the ratio of performance hit to reduction in latency was in moving to eDRAM?
Multi-Ported Static RAM allows for fast (simultaneous) access from multiple cores into cache. With multi-process heavy workloads, where data in the cache may not be simultaneously accessed from different cores or hardware strands, eDRAM may be a good fit. With software multi-threaded heavy workloads, where the data in the cache will be accessed simultaneously by multiple cores and hardware strands, eDRAM may suffer in comparison to multi-ported SDRAM due to excessive inefficient re-loads from main memory and inefficient sharing.
I wonder what the ratio of benefit to performance hit in throughput for moving to eDRAM was in comparison under various real-world workloads where multi-threaded applications need to share the instructions & data in the cache?
I wonder if the performance of eDRAM will be as linear as SDRAM, as the processors get loaded up? (This reminds me of the Intel 50MHz 80486 vs Intel 66Mhz (33MHz bus) 80486 tradeoff from years past...)
The move to eDRAM is very interesting, almost as interesting as OpenSPARC moving to highly threaded octal cores many years ago... will other vendors emulate IBM in the move to eDRAM L3 cache, the same way IBM, Intel, and AMD are moving to 64 bit octal-core as OpenSPARC did years ago?
I wonder if SPARC is dead. I hear SPARC64 VIII has been canceled.
I wonder what would have happened if the T2 would have had a full function core, vs a SPARC II core.
I wonder what the T2+ would be like if i ran more than 1 thread at a time with round robin of 8 threads.
The power7 SAP and Java benchmarks put the T2+ to shame.
What i don't wonder about is power superiority vs anything Oracle has.
Cheers David
Exciting technology in Power7 and an interesting story in how they continue to be the leader in systems.
Will be intertesting to see how it compares to Tukwila which the only thing I have seen is going to four cores still with only two threads.
We have written off Sparc and have Itanium on the divest list.
It's down to three vendors. Intel or AMD for x86 or Power. Pick the right horse for the right race and track condition.
And the right system for your business.
As far as competition is concerned, eDram is implemented on IBM's SOI technology. Yet Intel is choosing to use the bulk CMOS and HKMG for its Chips, because of higher cost and difficulties of implementing. That put Intel at disadvantages should AMD licenses IBM's technology, note that AMD is already a SOI shop.
Dont forget that we have not seen the rest of the next gen CPUs yet. IBM is first out with it's POWER7. Soon we will see Intel Nehalem-EX (which is said to be faster than POWER7), AMD 12-core Bulldozer, Fujitsu eight core SPARC64 Venus, Intel Itanium Tukwila, Sunacle Niagara T3 (which is said to be as fast as one T5440. Sunacle will release an eight socket T3 server, with 2048 threads, it will kick butt, just as Niagara T2 kicked POWER6 butt).
So IBM is first out. Let us see the rest of the CPUs, ok?
Nehalem-EX will have 8 cores but less performance per core than EP.
Bulldozer will stuffer from the same more core less performance but core problem.
SPARC64 Venus aka VIII was canceled a year ago and Oracle has no interest in it.
T3 will be 12 months from now and will still be the slow cores but 16 on a die. It is not until yellowstone falls that Sun finally has a single thread performance chip. That is when Sun goes back to a four core chip with 8 alternating threads. Crazy as it sounds but I have seen the roadmap. They also hope to get to 3GHz...yippeee. The problem is Larry wont have the patience or desire to invest in such a low ROI which has a good chance of being another chip which turns out to be a ROCK.
Check out the SAP and Java benchmarks for the T2+ and Power7.
Time to be open to new technology.
kit posts, "Yet Intel is choosing to use the bulk CMOS and HKMG for its Chips, because of higher cost and difficulties of implementing."
That is a very odd statement - Intel is choosing to do something because it costs more and harder to implement?
Intel is at the top of various price, performance, and volume curves... There is a very good reason for Intel doing something - they are not dummies and should not be so easily discounted.
IBM published that its 3.1GHz Power 7 8-way is 8% faster than 5.0GHz Power 6+ 8-way. The 4.1GHz Power 7 is 47% faster than the 5.0GHz Power 6+. No reason for anyone to make-up anything else. IBM's numbers can certainly be checked, though.
As to Sun somehow inventing the high core-count server, this reeks of 'Howard Stern invented the radio' thinking. INTC showed 100 'cores' on a chip in the 90s, it turned out to be an engineering adventure, though. The point is to establish technology and push it to profitability and acceptance in the market. T1 was mainly 'accepted' by folks that have pictures of McNealy stapled to the ceiling over their bed.
I don't get how IBM misses a deliverable because someone with a website announces what they think 'should be coming out'. I agree that IBM soft-peddled 2009 because the market was sooo soft. INTC did the same. It was just one of those crappy years for technology advancement.
i wonder if power would be faster if ibm designed a faster core - i hear the power 7 thread is about half as fast as a power 6 thread (power7 looks more like sun's opensparc all the time)
> I wonder if SPARC is dead. I hear...
sparc vendors have come and gone - sparc is an open specification, anyone can implement, and there have been multiple sparc cpu fabricators over the years
if power gets killed by IBM, the fools who placed themselves in a single supplier situation of a proprietary system would be in a heap of trouble - os/2 anyone?
> I wonder what would have happened if the T2 would have had a full function core
it would be 2-4x faster than a power socket, instead of only faster than power6
> I wonder what the T2+ would be like if i ran more than 1 thread at a time with round robin of 8 threads
it would be 2-4x faster than a power6 socket, instead of only faster than power6
> The power7 SAP and Java benchmarks put the T2+ to shame
let's hope itty-bitty-machine's brand new power7 is a little faster than the nearly 2 year old t2+!!!
a price/performance ratio comparison on a per-socket basis is needed before any shame will stick
pretty funny how power6 was slower than t2 and ibm modeled power7 after the sparc t processors
"Soon we will see Intel Nehalem-EX (which is said to be faster than POWER7)"
There's not a chance that the Nehalem-EX will be faster than the POWER7.
- Neither will Tukwila (Intel is only claiming 2X montvale)
- Bulldozer will be less than 2x of the 6 core Magny cors (frquency decrease)
- Nehalem estimate is below. It is less than 2x Nehalem EP due to the loss in cpu frequency when going to 8 core. Top bin is only 2.26Ghz)
- Venus is still a long way away and if just look at the pitiful Sparc64-VII result. It will have to be magical increase in performance to even go to 1/2 the performance of POWER7.
- Niagara T3 - You've got to be kidding, right?
Some comparisons here:
64 core / 8 chip 3.86Ghz POWER7 - 2530/2240
64 core / 32 chip 5Ghz POWER6 - 2155/2184
32 core / 8 chip 4.14Ghz POWER7 - 1460/1300
64 core / 32 chip 1.6Ghz Itanium - 797/727 (HP)
48 core / 8 chip 2.6Ghz Opteron - 730/470
64 core / 16 chip 2.88Ghz SPACR64-VII - 706/666
24 core / 4 chip Opteron 2.6Ghz - 400/276
8 core / 1 chip 3.86Ghz POWER7 - 326/293 (scaled from 16 core result)
8 core / 1 chip 2.26Ghz Nehalem-EX - 200/150 (est)
8 core / 2 chip Opteron 2.9Ghz - 155/112
4 core / 1 chip 4.14Ghz POWER7 - 183/163 (scaled from 32 core result)
4 core / 1 chip 3.33Ghz Nehalem - 140/108
8 core / 1 chip 1.6Ghz T2+ - 97/69
You know, there are analysts that think that Nehalem-EX will be faster than POWER7. It is nothing I wish.
Regarding your benchmarks, so what? We have to wait until Nehalem-EX, AMD Bulldozer, Fujitsu SPARC64 Venus and Niagara T3 arrives. As I said before, IBM is first on to the arena. Wait until the rest arrives, before you draw your conclusions?
.
And even if Nehalem-EX is slightly slower than POWER7, so what? Nehalem-EX and POWER7 will be roughly equivalent in speed. But one CPU will be far cheaper. Why pay $400.000 for an IBM P570 server, when you can buy several Nehalem-EX servers for the same price? Then your argument about speed is worthless, because P570 can not outmatch four 8-socket Nehalem-EX servers for the same price. And they provide greater redundancy too. So, which gives most bang for the buck?
"You know, there are analysts that think that Nehalem-EX will be faster than POWER7. It is nothing I wish."
"Regarding your benchmarks, so what? We have to wait until Nehalem-EX, AMD Bulldozer, Fujitsu SPARC64 Venus and Niagara T3 arrives. As I said before, IBM is first on to the arena. Wait until the rest arrives, before you draw your conclusions?"
Try to do some research before you start with your ramblings.
Let's try to compare not the Power 770 or 780 (the power 570 replacement), which is a monster, but the power 550 replacement the POWER 750. It absolutely
Now benchmark time:
SAP STD Benchmark 2Tier:
http://www.sap.com/solutions/benchmark/sd2tier.epx?num=100
15600 users by a
Power System 750, 4Proc/32Cores/128 Threads, POWER7@3.55 Ghz
4720 users by a:
T5440, 4Proc/32 Cores/256 Threads, T2+@1.6GHz
4370 users by a:
HP ProLiant DL585 G6, 4Proc/24Core/24Thread, Opteron@2.8 Ghz
3800 users by a:
Sun Fire X4270, 2 Proc/8Core/16Thread, Xeon X5570@2.93 Ghz
And the smallest system to beat the POWER 770 at this SAP level (unicode v6 EHP4) is this one:
17430 users by a:
SPARC Enterprise Server M9000, 32 Proc/128Core/256Thread SPARC64 VII@2.88 Ghz
That was on SAP2Tier and the power 750 is a factor 3+ faster than the all the other .
Specint_rate2006:
1,070 specint_rate2006
Power System 750, 4Proc/32Cores/128 Threads, POWER7@3.55 Ghz
360 specint_rate2006
T5440, 4Proc/32 Cores/256 Threads, T2+@1.6GHz
That was on specint_rate2006 and the power 750 is a factor 3 faster than the T5440.
SPECJBB2005:
2.478.929 SPECJBB2005
Power System 750, 4Proc/32Cores/128 Threads, POWER7@3.55 Ghz
841.380 SPECJBB2005
T5440, 4Proc/32 Cores/256 Threads, T2+@1.6GHz
That was on SpecJBB2005 and the power 750 is a factor 3 faster than the T5440.
And if you look at the Nehalem SUN SAP benchmark then it is more than a factor of 4 slower than the POWER7. Now a 4 socket 32 Core Nehalem EX will only clock at 2.26 GHz and it doesn't scale as well as POWER7 so it will not be able to pull a factor of 4.11 more performance by going to x2 more cores that clock 20% slower and x2 chips. Get real.
"And even if Nehalem-EX is slightly slower than POWER7, so what? Nehalem-EX and POWER7 will be roughly equivalent in speed. But one CPU will be far cheaper. Why pay $400.000 for an IBM P570 server, when you can buy several Nehalem-EX servers for the same price? Then your argument about speed is worthless, because P570 can not outmatch four 8-socket Nehalem-EX servers for the same price. And they provide greater redundancy too. So, which gives most bang for the buck?"
And price.. why don't you mention that a T5440 costs 116KUSD 128GB RAM 4 CHIPS@1.6GHz, and a POWER 750 costs 102KUSD with 128 GB RAM 4 CHIPS@3.0GHz.
Now that 14KUSD less for a machine that is 3 times as fast.
And your argument with pitching p570 (You know that that is a POWER5 machine right ?. Power 570 is the POWER6 machine and the POWER 770/780 is the POWER7), against non existing 8 socket Nehalem-EX system is pathetic.
8 Socket Nehalem-EX systems will most likely compete with high'er clocked 4 socket POWER 750 systems.
And 8 socket Nehalem-EX systems won't be cheap, try to figure out what a 8 socket IBM x3950 costs
http://www.tpc.org/results/individual_results/IBM/IBM-x3950M2-Linux-DB2-ES.pdf
rather than just trying to preach some
With regards to Buldozer try to look at Mangy-Cours first.
Try to go to HP's homepage and configure a DL785 and see what it costs. And try to match that to a POWER 750 which will stil most likely be 50% faster.
Venus ? What product ? Where is that ? I haven't seen it ? It's a HPC chip, it is still only foilware for commercial servers. The reality is that 4 socket entry level POWER7 servers are matching big refigerator sized M8000 where just a processor card is x3 more expensive than the POWER7 750 server. And the M9000 is gonna feel the heat from a midrange box like the POWER 770.
The facts are that IBM have kicked the rest of the Server business in the nuts, they have released a series of machines where the low'est bin servers are a factor of four faster than the highest bin of their previous generation. And at a cheaper price you get x4 the Cores and x8 the memory.
So there is just one thing to do for the competition that is pick up the gauntlet and lets have some good old fashioned price war. Good for us Server Buyers in this economy.
// Jesper
Analysis ? FUD is more the word. That guy is not even funny.
Ok lets have a look at what this guy said:
"I was right with my prediction in July 2009that the Power7 would have less single-thread performance than Power6, thus following the development in other processor franchises:"
The throughput of 85220 (Certificate 2010004) is good value, but when you look at the thread count you will see, that the performance halfed when compared to a Power6 thread. 85220 P7 SAPS divided by 128 Pow
er7 threads is roughly 665 SAPS per thread. 20520 P6 SAPS (Certificate 2009023) divided by 16 Power8 threads results in 1282 SAPS per threads.
The reduced thread performance is consistent throughout many benchmarks. SPECint_rate_2004, SPECjbb_2005 as published by an IBM overview document show pretty much the same behaviour.
The Power7 recites pretty much the complete CMT playbook. More threads, but slower threads. I think they will run in the same class of challenges at customer sites, expecting huge performance leaps for single thread application after viewing at benchmarks in favor of CMT systems. We learned that the hard way. Now IBM will learn the same ;-)"
This guys simply doesn't understand SMT. But well he is perhaps only used to work with statically scheduled coarse grainded multithreading of Niagara. A single thread can get the full use of all execution units it can eat inside a core. It doesn't get 25% of the core as it does on Niagara. You get 100% and then some. And the 'some' you get is more than the 25% you get on a Niagara processor thread.
It's like watching a guy with a musket making fun of a guy with a machinegun cause the machinegun guy hasn't got a bag of gunpowder with him.
"SAPS is a quite cache sensitive benchmark. So the large 32 MByte cache helped for sure. I want to see some additional benchmarks with a working set size larger than the caches."
Ehhh.. when SAP changed the SAP 2Tier standard benchmark to use Unicode +50% memory consumption and limited the response time to 1 Sek, what happened ?
What happened to POWER:
Before:
power 570 8 Core's at 4.7GHz 4010 users (6.0 2005)
After:
power 550 8 Cores at 5.0GHz 3752 users (6.0 Unicode EHP4)
Now that is 6% fewer users with an 6% increase in Clock speed.
Lets compare to the SU^H^HOracle T5440.
Before:
T5440 32 Cores at 1.4GHz 7520 users (6.0 2005)
After:
T5440 32 Cores at 1.6GHz 4720 users (6.0 Unicode EHP4)
Again this is laughable. Also
"The p7 systems like the p770 are still not capable to have an uniform dram speed over the complete range of ram sizes. With the smaller modules enabling them to put 32 GB – 512 GB into the p770 the dram is clocked at 1066 MHz DDR3, with the larger modules the dram is clocked at 800 MHz DDR3."
Yes, so what ?
"It looks like the p7 hits the scaleability limit of the i OS. The benchmarks for CPW were done with partitioning. The p770 performance page stated:
The 64-core system was configured with two 32-core partitions. The 48-core system was confihured [sic!] with two 24-core partitions."
Yes this is iOS, and if he knew something about iOS, he would know that a 32 core partition is more than 99.99999% of all iOS users could ever use.
And btw. a 32 core power 770 system that is roughly a system of the same size as a fully configured M8000. So what is he complaining about.
"IBM talked about the upgrade of p6 to p7 and the ability to keep the serial number. After reading a draft document about the p770 and the p780, it's pretty much the only component you can keep. Okay ... not that bad ... but they swap the CECs (the 4u chassis containing the CPU boards). That's a classic forklift upgrade. (look at section 1.9 of the draft)"
Well, lets see.. going from SUN E6900 forklift upgrade to M8000 and if SUN hadn't foobared Rock it would have been a forklift upgrade to a Rock system. So now I am just stuck with, well two versions of sparc that behaves very differently.
"Albeit there are two P5IOC-2 chips for IO-connectivity, the p770 seems shares the imbalance in the CEC with the p570. Both IOCs are connected to the chip 0 whereas both GX++ ports are connected to the chip
1. At first this is a single point of failure (failure of chip 0 leads to failure of the I/O in the CEC) and it puts additional load on the interconnect between both chips. As it's a NUMA design, the interconnect is shared with the memory transmissions between both CPUs. I'm not sure if this a problem in practice, but time will tell. (page 30)"
The p770 is a midrange box. What do you expect ? Sure it has redundant clock card with is able to failover without the machine going down. Something that even an M9000 cannot do.. Gee. And it's not like the machine would crash it would survive as he also writes. So basically a main chip component fails and the system keeps running with 75% of the IO still intact. Now that is pretty nice for a midrange system.
"For maintainance at the CPU boards you have to unplug a complete CEC and get it out of the rack. You will lose a lot of compute power and memory when doing so.That's not IBM specific. This is a common probl em when more compute power is in a single enclosure. But given the compute power i hoped for a way to change CPU boards from the outside."
Again this is a midrange box, if you want books that go buy a power 780, you can also get faster cores in that one. But the key here is that you can hotplug a CEC. It's not like you can do that one a SUN midrange system. It's not the power 770's fault that it's just as fast as a M8000 or M9000.
"But there is a comment in the documentation, i found rather disturbing. On page 141 IBM states as requirements to do Hot-Node Add, Hot-Node Repair and Memory Upgrade:
Move business applications to another server using the PowerVM Live Partition Mobility feature or quiesce them. This means that all critical applications must be halted or moved to another system before the operation begins. Noncritical applications can be left running. The partitions may be left running at the operating system command prompt."
Wonder where he found that ? Most likely in an old manual.
Why not try a new one: http://publib.boulder.ibm.com/infocenter/powersys/v3r1m5/index.jsp?topic=/ared3/ared3kickoff.htm
Lets look in this: ftp://ftp.software.ibm.com/common/ssi/sa/wh/n/pow03023usen/POW03023USEN.PDF which is what the manual advices yo to read.
it says:
Memory and processor capacity are the easiest types of resources to plan for since PowerVM can transparently change processor and memory assignments to accommodate the removal of a node. The only requirement that the system administrator must plan for is that there be enough free memory and processor capacity available in other nodes on the system to accommodate processor and memory resources that are in use in the node being evacuated.
Well that more or less says it all. And lets not mention the fact that he talks about old POWER6 systems in an blog about POWER7....
// jesper
I agree that POWER7 has the record - AS FOR NOW. But wait, all guests have not arrived to the party yet.
Maybe I should instead do like some person I "talked" to: claim these are cherrypicked benchmarks, claim that Niagara is still the fastest (because if you look at how many Hz Niagara uses, it gets more work done than a POWER7 which must use a high Hz), or other weird excuses.
But no, I am not like that kind of guy. I am not like him. I admit that POWER7 is fastest as of now (before every one else releases their new next gen CPUs). I have no problems of admitting that, why would I? The benchmarks say something, and I trust them. I am not like you (cherrypicked benchmarks, Sun crafted benchmarks, more efficient Hz, or whatever strange excuses you earlier have presented to refuse the benchmarks. I like the one best where you explain that although you need four 5GHz POWER6 to match two Nehalem 2.93GHz in official TPC-C benchmarks, the POWER6 is still faster!). But no, I am not that kind of guy.
But I also expect Nehalem-EX and other CPUs to play in the same league as the POWER7 at a much lower price. Let us say that Nehalem-EX is 90% as fast as a POWER7, which one should you buy then? For the price of one small POWER7 server, I can probably buy several 8-socket Nehalem-EX servers. Let us say I can buy three Nehalem-EX servers. Then I get 8x3=24 Nehalem-EX cpus, and that nice cluster will easily outperform eight POWER7, dont you think? And they provide more redundancy.
Or if we consider the top P595 which were used for the old TPC-C record. It costed $35 million list price. I bet the top POWER7 server will cost as much, or even more? If a 8-socket Nehalem-EX server costs $35.000, then I can buy 1000 intel servers. That intel cluster with 8.000 Nehalem-EX CPUs will surely best 64 of the POWER7 in the IBM top offering server p795(?). Dont you think? And that cluster will provide greater redundancy, yes? So I fail to see why I should not buy several Intel Nehalem-EX servers for the price of one IBM POWER7 server?
With that said, let us wait for the other guests before we draw any final conclusions? I understand that you do not want that, but I want to know about the other chip manufacturers offers before I draw a final conclusion. I want to see the whole picture, even if you dont.
I love it when people compare Intel servers, built with commodity hardware against enterprise-class Intel or IBM servers. NO ONE will take stock parts and come anywhere near the performance, uptime or maintainability of the Power 7 (or a similar Intel server).
These machines have special inter-processor communications, special memory configurations, special I/O designs and special north/south bus configurations to give extraordinary throughput.
If you compare a four-core blade server against a six-core Power 7, you could gloss over the speed gap with a cost ratio of 10 or 20 to One. The Intel box wins all of those sales. If the machine needs to stay up 350 days in a row, Intel doesn't win those sales. The first market is MUCH larger, the per-server profit is MUCH larger for the latter.
"Maybe I should instead do like some person I "talked" to: claim these are cherrypicked benchmarks, claim that Niagara is still the fastest (because if you look at how many Hz Niagara uses, it gets more work done than a POWER7 which must use a high Hz), or other weird excuses."
You still don't get the distinction between fast and throughput.
Now Niagara is not fast, it has a exceptionally good throughput. This means that it can get parallel jobs with many threads done in a short time (fast). If you want to define this as fast, no problem you do that, I have never denied that.
"But no, I am not like that kind of guy. I am not like him."
Yes that is a line taken from the Eminem song "Stan", just watch where you are driving :)
<ramblings deleted>
"But I also expect Nehalem-EX and other CPUs to play in the same league as the POWER7 at a much lower price. Let us say that Nehalem-EX is 90% as fast as a POWER7, which one should you buy then? For the price of one small POWER7 server, I can probably buy several 8-socket Nehalem-EX servers. Let us say I can buy three Nehalem-EX servers. Then I get 8x3=24 Nehalem-EX cpus, and that nice cluster will easily outperform eight POWER7, dont you think? And they provide more redundancy."
"Or if we consider the top P595 which were used for the old TPC-C record. It costed $35 million list price."
Again with the math. That was the List price of the whole solution just cause you keep mentioning it doesn't make it more the truth.
That is like me saying that 12 T5440's cost 18MUSD, and it says so right here:
http://www.tpc.org/results/individual_results/Sun/Sun_T5440_TPC-C_Cluster_ES_122309_v2.pdf
Now that is 1.5MUSD a piece compared to the price of 101KUSD of a power 750.. Wee Wee Wee...
"I bet the top POWER7 server will cost as much, or even more? If a 8-socket Nehalem-EX server costs $35.000, then I can buy 1000 intel servers."
Well the CPU's might. You obvious have no idea of what a server cost. 350KUSD for a 8 socket Nehalem-EX server sound about right. x3950 is around that price I guess.
"That intel cluster with 8.000 Nehalem-EX CPUs will surely best 64 of the POWER7 in the IBM top offering server p795(?). Dont you think? And that cluster will provide greater redundancy, yes? So I fail to see why I should not buy several Intel Nehalem-EX servers for the price of one IBM POWER7 server?"
You can do what you like. But perhaps you want to get an accountant to help you.
You are a factor of 10 wrong in your Nehalem server price.
You are a factor of 6 wrong in your cost on the p595
So you are basically down to having 1000/60 = 16 nehalem Servers. Now clustering these servers together with your favorite Oracle RAC software will at best give you 80% scalability
Hence 10.5 times the throughput of one server.
So you will have 10.5 times the throughput of a single server. So basically it's 84 sockets of nehalem versus 64 POWER7's.
That is most likely pretty evenly matched. And then you can try to do the math on the RAC software............ which will cost you many times that of the hardware.
With that said, let us wait for the other guests before we draw any final conclusions? I understand that you do not want that, but I want to know about the other chip manufacturers offers before I draw a final conclusion. I want to see the whole picture, even if you dont..
This line of argumentation means that you can never ever compare anything.... *grin*
// jesper
Maybe you didnt read what I wrote? I can write it again.
If you can buy several 8-socket Nehalem-EX servers for the price of ONE power server, which choice provides the most redudancy and performance you think?
It does not matter if Nehalem-EX reaches 90% of POWER7 speed, you will still get a cluster with 3-4 Intel servers for the same price.
One cluster with 24-32 Nehalem-EX cpus or one POWER7 machine with 8 cpus, which is fastest and more reliable?
.
Jesper Frimann,
Why dont you repost on the guys blog again? You got owned the first time, maybe you will get owned again? ;o)
I didn't get owned. You didn't read the part:
"Sorry about the Power6/Power6+ statement ... i thought about something different unrelated to SAP at that moment. Sorry, for that"
Gotta give him some credit for admitting he was wrong. And the guy when you give him some resistance is actually quite knowledgeable, and sensible when he wants to be.
And you keep your Nehalem-EX dreams alive don't you... YOU CANNOT BUY SUCH A SERVER CASUSE THE DON'T EXIST.
And they don't cost 36KUSD but more like x10. BLEH.
// Jesper
Regarding Niagara being fast or having high throughput. So what does this mean for a cpu to be "fastest"? It can get the work done in the shortest amount of time, yes? So if you have high throughput, thus can finish all work in a shorter time - the cpu is fast, yes? Or do you mean that "fast" is when the CPU has low latency? Or can add two numbers fastest, or what? What is "fast"? How do you define "fast"? My definition is wrong?
.
There are speculations from my side, and there are speculations from your side. I ask questions, "if this is true, then...". You, sort of, declare things "this is how it is, the Nehalem WILL cost 350kUSD, the top POWER7 machine WILL cost roughly equal" etc.
As I said, I am not as sure as you, and I want to know more facts about the other parties before I draw a conclusion. I also understand that you do not want more facts, but are happily declaring your guesses as facts, this is in line with your earlier reasonings, so this is nothing to be surprised of. When I say I want to see all facts to be able to draw a conclusion, you dont want that, you write: "This line of argumentation means that you can never ever compare anything"
Anyway, I am NOT willing to blindly accept your statements about future Nehalem-EX (price and performance) as facts. You can declare me ignorant how much you want. But I will NOT believe you until I see the facts. I hope you can accept this.
.
If the POWER7 is clearly faster than the rest of the cpus in it's generation: SPARC Venus, Nehalem-EX, Niagara T3, AMD Bulldozer, etc - then I will have no problems to admit that. I know you have problems admitting that e.g. Nehalem is faster than POWER6 in TPC-C, but that is totally in line with your earlier reasoning and nothing to be surprised of.
I can say that again: if POWER7 is fastest, I will admit it. Simple as that. I will not deny it categorically as you did. And if some other CPU is almost as fast, but that server cost a fraction of the price, I can not see the point of buying POWER7? But if the future servers to come, are all roughly equal in price just as you vigorously claim (without knowing anything) then I see no hindrance to buy POWER7.
But, I also want to see the rest of the CPUs, before I draw any conclusions. Ok? You are simply guessing, but dont try to imply that I am ignorant or something, for wanting to see all facts.
.
Regarding the guy who owned you, I guess he didn't own you.
Well done, Jesper - you've been doing a lot of reading and have a darn good understanding what with so little sleep with that new born kicking around! :-)
Kebby - why don't you just take a year off and come back when (if?) T3 sees the light of day. Fast to you obviously means Siebel benchmarks since that's how you establish the latest T-box is faster than POWER6 when the rest of the world knows better.
POWER7 is a nice achievement...my favourite highlights...more cores while actually getting better performance per core. Two memory controllers to keep the beast fed at phenomenal data rates. Lots of on board cache. eDRAM to keep the transistor count down, hence the power requirements sane and help put more cores on board. It's a very nice design. And 4.1GHz premium clock sort with all those cores...nicely done.
The dog and pony show I saw said stay tuned for more astounding benchmarks given the annoucement was supposedly pulled ahead from May to Feb. Back to my BEvERage...
A.C
I didnt know that Venus is canceled. Is it just the normal FUD from IBMers, or is it true? Do you have links so I can read more? I have learned that IBMers here FUD and lie a lot. So please forgive me for not trusting you.
VirtualGreg
I agree that POWER7 is a nice CPU, I have never claimed else? It is fastest right now. I admit it, what is the problem? I do not disagree, just as some IBMers would do. ("No, the POWER6 is faster, despite you need four of them to match two Nehalem in TPC-C"). I have no problems accepting benchmarks and white papers? ("No, all these benches where Niagara win over POWER6 are cherrypicked, or made by Sun, or there is some other problem with them. There must be some problem").
But I am trying to say that I want to see the rest of the CPUs, is that wrong of me? Why are you IBMers so desperate to explain that the rest of CPUs are all canceled, etc? I want to see all CPUs in this next generation, to be able to compare them on equal merit.
Playstation3 was one year later than Xbox360, but all agrees that PS3 and Xbox360 are the same generation and should be compared. Microsoft did not say "no, PS3 is canceled, X360 is better than PS2, PS3 will not arrive, you can not compare PS3 vs X360", etc and all the other weird things you IBMers say? X360 and PS3 are comparable, I want to compare all CPUs in this next gen. I see nothing wrong with that? Could you explain what is so wrong of me wanting to wait for all other CPUs?
Venus:
Well it's not on any Oracle roadmaps:
http://regmedia.co.uk/2009/09/11/sun_sparc_roadmap.jpg
So I guess we are not going to see it anytime here soon.
And with regards to comparing POWER7 to current T2+. Then most of the comparisons have been made with the 1.6GHz T5440. The 1.6GHz version of this machine is from Q3 2009. That it is just a little more than half a years old. That makes it damn current.
You have been waving your Siebel benchmark around now for half a year now. And that benchmark is from May 2007, and so is the POWER6 servers you taunt. So perhaps you shouldn't be so self righteous.
And it is hard right now for the T2+ based T5XXX servers.
The 1/2 socket servers are outperformed by the Nehalem 1/2 socket servers and blade servers like BL280c, which cost far less. And are faster.
And the T5440 is a good deal slower and actually also more expensive than for example the more versatile new power 750.
http://shop.sun.com/is-bin/INTERSHOP.enfinity/WFS/Sun_NorthAmerica-Sun_Store_US-Site/en_US/-/USD/ViewStandardCatalog-Browse;pgid=lkpIezKOHTdSR0kLFbvpraZJ0000ClS4R_t3;sid=mY6DGousqqODd8DbThfLH2QPH7RMirAioIOSa-oMG4WOj10G-s0=?CatalogCategoryID=ZXVIBe.d7kYAAAEZYYsJ0gWj&PriceIt=true&Country_Territory=Sun_Catalogue_US
and
http://www-03.ibm.com/systems/power/hardware/750/browse_aix.html
So for the T5440 to match a system like the power 750 Oracle needs to get T3 out of the door asap and perhaps cut prices on the T5440 also. As the power 750@3GHz would still have a 40% price/performance advantage over the T5440, if the T5440 doubled in performance.
// Jesper
This post has been deleted by its author
Ok, maybe you know something I dont, but when we talk about Venus, you explained to me it is not canceled, and you show it is not on Oracle's roadmap - but I thought Fujitsu manufactured the Venus, and manufactured the Venus servers?
.
"...And with regards to comparing POWER7 to current T2+. Then most of the comparisons have been made with the 1.6GHz T5440. The 1.6GHz version of this machine is from Q3 2009. That it is just a little more than half a years old. That makes it damn current...."
The Niagara T2 had a clock increase, and suddenly it had a complete overhaul and is magically transformed into a "next gen" cpu, on par with Niagara T3?
If Pentium4 had a clock increase and sold even today, you would consider it a "next gen" cpu? Some other IBMer explained that Tukwila and POWER7 were announced the same day and therefore they are in the same generation, but if they are they were announced different dates, they would not be comparable. I dont agree with either of you.
"...You have been waving your Siebel benchmark around now for half a year now. And that benchmark is from May 2007, and so is the POWER6 servers you taunt. So perhaps you shouldn't be so self righteous..." Self righteous for what? Most people consider POWER6 and Niagara T2 belonging to the same generation. But that is wrong you think? Most people thinks that PS3 and X360 belongs to the same generation too, but that is also wrong you think?
.
I agree that T5440 is more expensive, and slower on some tasks. But I expect T5440 to be many times faster work similar to Siebel. If you need 28 POWER6's to match four Niagara, and if POWER7 is four times as fast as POWER6, you need seven POWER7 cpus to match four Niagaras.
But back to topic. I tried to show that IBM has expensive gear, and probably Intel Nehalem-EX is cheaper. Yes, I know you explained to me, that Nehalem-EX servers will not be cheaper than POWER7 gear. And to that I am not as sure as you, as the pricing on Nehalem-EX servers are not revealed yet. I suspect Nehalem-EX will be much cheaper. Ok?
Yes, I know you think different, but wait and see I say. No, you say, no waiting. Compare POWER7 to what is on the market now, you say. And to disprove my claim that Nehalem-EX servers will be much cheaper than POWER7, you suddenly start to talk about T5440 pricing?
But I STILL believe (no matter how many times you try to convince me) that Nehalem-EX servers will give you performance in the same league as POWER7 gear, and at a much lower price. I suspect you could buy several Nehalem-EX servers for the same price. I say: wait and we will see facit.
When you claim things, you are rock sure on yourself and you are correct. I am not so sure. Forgive me, but you are just guessing.
You said "I agree that T5440 is more expensive, and slower on some tasks. But I expect T5440 to be many times faster work similar to Siebel. If you need 28 POWER6's to match four Niagara, and if POWER7 is four times as fast as POWER6, you need seven POWER7 cpus to match four Niagaras."
28 vs 4? Where did you come up with this?
Well I thought I pretty much debunked that benchmark in this post:
http://forums.theregister.co.uk/post/674223
But it looks like you found a new one:
http://www.spec.org/osg/jAppServer2004/results/res2007q3/jAppServer2004-20070829-00082.html
http://www.spec.org/osg/jAppServer2004/results/res2009q3/jAppServer2004-20090701-00135.html
Lets start commenting on that one:
1) Java version the T5440 on the benchmark runs Java version 1.6.0_14 the POWER 570 runs 1.5.0. If you have a look at the JBB2005 benchmark you will see that for POWER going from 1.5->1.6 will give you a boost of aprox 25%. Perhaps even more.
2) You keep mentioning two power 570'es. Actually it is one POWER 570 and one p550. The p550 is an oldish POWER5+ machine.
3)The POWER 570 is a 1 CEC version. It's not a fully configured power 570 It could just as well have been a POWER 520.
4) The T5440 machines uses 10Gbit network interfaces, the POWER solution uses 1Gbit hardware.
5) The T5440 database machine uses SDD's, and the POWER uses Hard drives. T5XXX benchmarks also used hard drives back in 2007. The T5440 also has three times the adapters.
6) The T5220 Database server uses 256GB of RAM, the POWER p570 used 8GB. That is 32 times the memory.
7) The T5220 APP server uses 256GB of RAM the POWER 570 uses 16GB of RAM. that is 16 times the memory.
8) The response times on the two systems is very different, on the POWER setup the MAX request time is on average 5.5 times faster than the MAX request time on the T5440's.
Furthermore all the MAX request time on the T5440's are over 20 sek with an average of 30 sek. That is half a minute.
9) If you knew just a little about UNIX you would quickly see that the people who setup the T5440's have done a much better job than the ones setting up the POWER servers. The most basic network and AIX java tuning hasn't been done. It's right in the manual.
10) You keep kebabbing about how not to compare the current hardware with current hardware. The T5440 used in this benchmark, and the benchmark result is from Q3 2009, the power 570 is from Q3 2007. So make up your mind, is it fair to compare or not ?
11) The amount of CPU's you mention is out in the woods.
The IBM solutions used APP:4cores/2Chips POWER6 and DB: 4 cores/2 Chips POWER5 in SUT HW , the SUN submission uses AP:32 cores/4 chips and DB:32 cores/4 chips.
Now that is 4 POWER5+4POWER6 cores versus 64 T2+ cores or 2 Power6 + 2 POWER5 chips versus 8 T2+ chips.
So you have a T5440 solution with 8 times the cores, 24 times the RAM and double the chips. Running on a much newer Java version using SDD's, 10Gbit net etc. etc etc. 2 years between the submissions. Half the cores in the POWER versions being POWER5 cores.
Now I don't want to hear anymore of your compare with next generation etc etc. You are doing it yourself when it suits you.
And if you think that certain workloads with regards to T2+ versus POWER7 why don't you compare the SPECJBB2005 results for T2 versus the POWER7.
// Jesper
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
