Power9: Google gives Intel a chip-flip migraine, IBM tries to lures big biz IBM's Power9 processor, due to arrive in the second half of next year, will have 24 cores, double that of today's Power8 chips, it emerged today. Meanwhile, Google has gone public with its Power work – confirming it has ported many of its big-name web services to the architecture, and that rebuilding its stack for non-Intel …
When IBM told Apple to sod off because Apple represented less than 1% of its chip sales/production, Apple went with Intel and became the largest OEM using Intel's CPUs. Now, what kind of a niche does IBM really have when they can do that? I remember IBM boasting that they could host the entire WWW on just one of their mainframes. Maybe they still can.
So IBM is a niche player in the commodity data center market. Now, where is it that they are the big player? They are selling CPUs hand-over-fist someplace, but where?
>Apple told IBM to shove its power somewhere else
The power chips are for when you need massive compute and have a devil-may-care attitude to power consumption. That isn't most of today's desktops, not today's laptops and certainly not today's tablets, which is where Apple wanted and needed to be. You don't put Intel's latest and greatest xeons in consumer PCs either and it doesn't make sense to have laptops running intel and mac pro's running power chips.
The key playing space of the Power infrastructure is virtualisation. When IBM went Linux, it also went to town on the virtualisation that it had plenty of expertise on already on mainframes.
As a consequences of IBM's ability to virtualise everything and even divvy up the CPU in segments you end up with jaw dropping resilience. You pay for that, of course, but if you want things to stay online you could do worse than kit you can add and remove while the platform stays live. Add to that ZFS support so you can virtualise where everything lives and you have quite a flexible infrastructure.
The only problem is that you will need to make sure you have the expertise to run it. There are lots of variables to play with, which we know to translate as lots of opportunities to screw up :).
As others have noted I seem to recall it was more Apple telling IBM to sod off as POWER chips just weren't really efficient enough for desktop use - hence why the latest and greatest Power PC machines had to be water cooled.
Even if IBM did tell Apple to sod off - now speaking hypothetically - if that was the case I suspect the reason behind it would have been that the development costs for the desktop range of POWER chips simply wouldn't have been worth the income. The development and production costs for processors are insane and, despite Apple flogging about 5-6 million Macs a year, I doubt that was truly enough to build a business case to continue the processor line. Given that we're going a way back and Apple hadn't reached fever pitch I doubt Apple actually sold 5-6 million POWER-powered Macs at the time either.
I disagree with IBM's approach on many things, but business is not as simple as "IBM threw away X amount of business so they're mental" - if that business was not profitable for IBM, and could not be made profitable within a reasonable period (if at all) then binning it was absolutely the correct decision.
Also don't be daft, Apple are nowhere near Intel's largest OEM partner. Looking at the number of PCs shipped they're 5th in the world. Ok so yes many of the other companies also supply AMD processors in their systems, but they are in the minority compared to the number of Intel boxes shipped and even if you say something extreme like "40% of all Lenovo/Dell/HP PCs are shipped with AMD CPUs" (which is just not true) they're all still considerably larger than Apple in terms of Intel sales.
http://www.gartner.com/newsroom/id/3185224
The figures also don't include servers, storage controllers, embedded systems etc that many of the other vendors manufacture. Half a million server units might not add much to the units sold but in terms of value to Intel: servers have more expensive chipsets, (typically) more and (typically) higher value processors and Intel also have a significant share of the 1Gb and 10Gb server-grade NIC market.
http://www.gartner.com/newsroom/id/3174018
Of course Intel must worry: datacenter processors is the fastest growing segment and responsible for big profits. With Google seemingly ready to increase its laughingly small number of GCP regions, it's bound to purchase processors by the hundreds of thousand.
Unless Intel someday makes strides in IoT / Wearables / the next hyped market, it will be highly dependent on the datacenter unit to bring sales and profits.
"Either the US government hasn't thought to outlaw the export of CPU blueprints, or Big Blue's technology in foreign hands isn't seen as a strategic threat to national security."
So I presume that the TLAs are hoping that the Chinese will not find the back doors that are cleverly hidden in the hardware designs.
On the bright side: All your data are now belong to both Uncle Sam and Uncle Ping.
Lends a new meaning to the game of PingPong ....
El Reg - we need an Unobtainium alloy reinforced tinfoil hat icon ...
Today the POWER8 is slower than the previous generation E5v3. Intel has just released the E5v4 which is faster, so POWER8 is lagging behind more and more. Here we see several benchmarks where POWER8 is slower than x86
https://blogs.oracle.com/BestPerf/entry/201510_specpu2006_t7_1
I doubt POWER9 will be faster than x86, because POWER8 is slower than x86. Why would POWER9 be faster?
Even if POWER9 is twice as fast as POWER8, it will still not be enough to outperform SPARC M7. SPARC M7 is typically 2-3x faster than POWER8 and the fastest x86 cpu. SPARC M7 is all the way up to 11x faster. Oracle SPARC servers are always >2x fast as the previous generation. In four years Oracle have released five generations SPARC, each generation doubled performance. Oracle does not rest, soon M8 will be relased, again doubling performance. POWER is slowest, and most expensive. Why do not IBM switch over to x86 completely and give up the hardware market, as IBM has done with almost everything else? IBM has already sold off all the hardware divisions except POWER and Mainframe.
Lets look at the price of those two systems,
The T7-1 (512GB) -> $58,610
POWER S812LC (256GB) -> $12,999
Taking that to perf/cost, the POWER system is 241% better
Note, the problem with the SPARC system is the amount of memory required to run the huge number of threads. Not to mention the horrible thread performance (
This smells like Oracle marketing, if not Phil Dunn. There is ZERO need to respond to @PlinkerTind's bloviating as they are just trying to be provocative while spreading FUD. I have an outstanding offer to Oracle to accept my "POWER Challenge". You bring your biggest & baddest SPARC server and I'll bring mine. I have a customer who runs Oracle on POWER. This customer will load their data onto both servers. Oracle can tune their OS/virtualization but can't touch the data. I'll tune my OS/virtualization and we will see which server runs the fastest with the fewest # of cores. We will then do a TCA & TCO analysis posting it all online. I will run an additional test which is to import the data into DB2 and run it on POWER. Oracle is welcome to run it if they like...I'll provide the TCA / TCO for this extra credit portion.
By the way, SPARC T7 processor consist of 8 x quad core chiplets that with small L1, L2 & L3 caches with the L1 being the only one dedicated to a core - the others are shared to a 4 core chiplet. 4 Memory Controllers feed the 8 chiplets that Oracle claims only delivers up to 160 GB/s which is 20 GB/s to each chiplet. Compared to the POWER8 SCM which has dual memory controllers delivering up to 230 GB/s to the entire processor/socket. Because they have so far been unable to build strong cores, they resort to building servers with the most cores. With the list price of Oracle Database Enterprise Edition at $47,500 + 22% annual maintenance, the per core performance is VERY important. Oracle's business is to sell software licenses. Instead of Oracle building SPARC 7 to have the same or fewer cores that are stronger so the aggregate performance is more than T5/M5/M6 they instead double the cores & claim 1.9X performance. This is why they have ExaData with so many Intel cores loaded with Oracle software licensed traditional and non-traditionally (storage server licenses hide some of the DB processing shifting some of the TCA/TCO to them making it look better when compared to competitors).
Yes, Oracle is predictable and boring. If they would just shut up and put up we could let the world draw their own conclusions.
First, your "analysis" on SPARC M7 is clearly nonsense and is just spewing with FUD. Talk to your engineering buddies who I'm sure will explain what SPARC M7 is really made of. Your Power Challenge is also full of holes and a complete waste of time. You sound like such a used car salesman!
The bottom line is Oracle just proved *publicly* that a single SPARC M7 processor chip with 32-cores, in a SPARC T7-1, list price of just $58,610 w/512GB RAM, just smoked a Power S824, with 4 x Power8 6-core CPU Chips, 24-cores total, with a list price of $92,716 for 512GB RAM, by 40% using an OPENSOURCE Database testing tool, HammerDB, which can run practically any Database including Oracle DB, DB2, Microsoft SQL Server,, MySQL, MariaDB, PostgreSQL, etc.
SPARC M7 outperforms Power8 ON A PER CORE basis, outperforms PowerS824 system by 40% and is practically half the price. All the disclosures and details are here and feel free to prove Oracle wrong!
https://blogs.oracle.com/BestPerf/entry/20160317_sparc_t7_1_oltp
So, with Oracle Database multiplier of .5x on SPARC vs 1x on Power8, running on SPARC M7, a customer will require less than *HALF* the licenses that they will need running on Power8. Now how is this going to *increase* Oracle license sales? By selling IBM maybe?
So why don't you get one of your Power8 systems, run HammerDB with your favorite Database and publish the results here for everyone to see? The tool is free, you can run any Database you chose and we can compare #'s. Your best to Oracles best, which is already public. www.HammerDB.com.
Even better, put it on your BLOG!!!
And while you are at it, download the free/easy/quick to use STREAM benchmark, to prove your bandwidth claims, at http://www.cs.virginia.edu/stream/
And heres Oracle's SPARC M7 results for all to see. Nothing to hide here.
https://blogs.oracle.com/BestPerf/entry/20151025_stream_sparcm7
and by the way, Oracle tested the PowerS824 since IBM refuses to (something to hide?). Details are here: https://blogs.oracle.com/BestPerf/resource/stream/s824details.pdf
Gee, a SPARC T7-2, 2-sockets, 2 x CPU Chips delivers maximum 302,758 GB/sec Bandwidth (151 GB/sec PER CPU) vs PowerS824 2-socket, 4 x Power8's @ 322,399 GB/sec Bandwidth (81GB/sec PER CPU)-where is the 230 GB/sec you claim??? UNPROVEN MARKETING FLUFF
Feel free to prove Oracle wrong!! Time to PUT UP OR SHUT UP!!
Let's start here. Oracle has 0 (i.e. ZERO) credibility. It isn't just the FUD and benchmark exaggerations but how many times have they been called out for publishing & stating things about their own systems and competitors that were purposefully wrong. So, you can say what you want, write what you want but the fact you work for Oracle, writing at an Oracle blog and even have Oracle testing an IBM POWER8 server is all reason to not believe a damn word you say.
It's laughable for you to say that IBM has something to hide...Right, why are there NO IBM POWER benchmarks running Oracle except for Oracle workloads like EBS? Because Oracle won't allow IBM to. Oracle's end user license agreement gives Oracle the right to approve this and they don't. Please keep your faux outrage as it rings hollow as everything else that you say.
You write in your blog "On a per chip basis, the SPARC T7-1 server demonstrated nearly 5.5 times better performance compared to an IBM Power System S824 server". Ok, let's break it down using "your vernacular". A T7-1, although it is actually 8 x quad core chiplets, you call it a single chip consisting of 32 cores while calling the 2 socket S824 a 4 chip system which IBM, the vendor of the server does not. That said, your 32 core chip, by YOUR words is 5.5X better performance. Thus, 6 cores vs 32 cores or 5.33X.
Next, you chose the 24 core S824 which is fine but it also runs at 3.52 GHz vs the 4.13 GHz T7. Why didn't you pick the 16 core S824 running at 4.15 GHz? That way you could call it a 4 chip, 2 socket with 4 cores per chip system and your 32 core, single socket, single chip (that is really 8 x quad chiplets) system is 2X the cores allowing for better and equal extrapolation? Even setting the ASMI mode to favor performance is the 2nd best option. There is a better option but since you don't know the platform you either chose not to or didn't know what to select.
Then, you say "On a per core basis the SPARC T7-1 server demonstrated nearly 3% better performance per core compared to an IBM Power System S824 server.". I'll just refer to the above paragraph where the T7 clock frequency was running 17% higher.
Then you say "At the system level, the SPARC T7-1 server demonstrated nearly 1.4 times better performance compared to the IBM Power System S824 server.". 24 cores times 1.4 = 33.6......hmmm. You tested a 1 socket vs a 2 socket server running 17% higher clock.
As stated at by your many, MANY distractors at http://www.anandtech.com/show/10230/ibm-nvidia-and-wistron-develop-new-openpower-hpc-server-with-power8-cpus-nvlink they rightly point out that Oracle ran the Oracle benchmark on the POWER8 server. There is no disclosure on how the server was configured; How many DIMMs were used? Filling all 16 DIMM slots will make a difference than just 8 or 4 since all 3 configuration options can achieve 512 GB Ram. Although a few tunable were disclosed as if to demonstrate that an effort was made, I question the ones used. I won't disclose what I would have done differently as I like that Oracle looks like petty in their effort to show Oracle on POWER8 performance. If they were interested they would authorize IBM to run their own.
Even though you won't go away, won't you just go away?!
OK, So you say that Oracle mis-configured the Power S824 for the HammerDB benchmark, where Oracle shows the 24-core 3.52GHz system did 3,609,832 transactions per minute. And by the way, the HammerDB test simulates the TPC-C benchmark workload quite closely so can be compared to the TPC results on TPC site.
And I just found a IBM slide showcasing Power S824 @ "3585K" TPM, as in 3,585,000 transactions per minute, lower than what Oracle published. Can you explain how Oracle misconfigured the system to get better results than what IBM shows??
https://urldefense.proofpoint.com/v2/url?u=http-3A__www-2D03.ibm.com_systems_data_flash_fr_resources_modernisation-5Fi-5F2015_S17-5F-2D-5FLinux-5Fsur-5FIBM-5FPower-2C-5Flalternative-5Fgagnante-5Fa-5Fx86.pdf&d=CwMFaQ&c=qRq7a-87GiVVW7v8KD1gdQ&r=I2xtURXsO1RWtTHq5QigTWSQEI0oZ0-d4SrzmJeBSTc&m=0MvNexPnvNRNGyiwFK8sYrVDdksU4VrFPMJ32OrGjxE&s=TPR6RebLomi0pzJw7DzZX9FgdCwko4FKfgylMua7q2E&e=
"Can you explain how Oracle misconfigured the system to get better results than what IBM shows"
No! I wouldn't help you improve your garbage benchmark even 1% just so you could half-ass it further only to attempt to exploit some other failure made by Oracle. Thank you for purchasing a POWER8 server.
Your math is wrong!!
A SPARC T7-1 (512GB) = $58,610
Where do you get $12,999? A Power S812LC with same 512GB RAM, configured identically to SPARC T7-1 is $37,845. Pricing is here: https://www.ibm.com/marketplace/cloud/big-data-infrastructure/us/en-us
OK, now lets look at available benchmarks: Theres only 3 published on S812LC:
SPECint2006rate, SPECfp2006rate & SPECjbb2015
In every Price/Performance calculation, SPARC T7-1 comes out ahead by up to 3x over Power S812LC. And if you compare to a Power8 system with more cores, the performance ratio of Power8 gets worse, not better.
SPECint2006rate Peak Price/Performance
----------------------------------------------------------
SPARC T7-1 $58,610 / 1,200 Peak =$48.84
Power S812LC $37,845 / 642 Peak =$58.95
SPECint2006rate Base Price/Performance
----------------------------------------------------------
SPARC T7-1 $58,610 / 1,120 base =$52.33
Power S812LC $37,845 / 482 base = $78.52
SPECjbb2015 Maximum jOps Price/Performance
----------------------------------------------------------
SPARC T7-1 $58,610 / 120,603 Maximum jOps =$0.49
Power S812LC $37,845 / 44,883 Maximum jOps = $0.84
SPECjbb2015 Critical SLA jOps Price/Performance
----------------------------------------------------------
SPARC T7-1 $58,610 / 60,280 Critical jOps =$0.97
Power S812LC $37,845 / 13,032 Critical jOps = $2.90
OH and just check out that Power Consumption difference!
a SPARC T7-1 Maximum Power Consumption = 1,318 W
a Power S812LC, per the IBM redpaper, has maximum power consumption of 3,500 W
That’s a 3x power consumption difference! Is that why its called Power??? As in Power Hog??
Yes, this is definitely Oracle marketing. You need to update your marketing FUD Oracle as the S812LC redpaper no longer has uses the 3500 Watt reference. It was no where close...it is 3 digits and less than your T7-1.
Oracle, just accept my "POWER Challenge".
Oracle is predictable and boring. If they would just shut up and put up we could let the world draw their own conclusions.
I stand corrected. IBM has corrected the redpaper and you are correct, the Power S812LC, a 10-core Power8 system, now shows 900 watts maximum. So a SPARC T7-1 @ 1,318 W is higher by 46%, but as Ive shown with all the benchmarks available, the SPARC T7-1 ranges from 28% to over 3x higher performance/WATT. And by the way, a 2 x Xeon E5-2699 v4 with 44-cores is @ ~600-700W, far lower than a SINGLE Power8 10-core chip. So clearly, Power8 is a Power HOG.
Let's see if I have this right....IBM is bad and Oracle is better....oh wait - now that you have facts IBM isn't so bad, actually better than Oracle but IBM still sucks because they aren't as good as Intel. <Bell rings> Hey, is that your 3rd grade bell ringing?........#NoCredibility
Typical response PowerBoy! Actually, I think you’ve never passed 3rd grade math. Clearly, you don't understand the metric "Performance/WATT" otherwise you wouldn't be making such lame remarks.
Bottom line is that a *single* SPARC M7 chip SPARC T7-1 server offers 28% to over 3x higher performance/WATT than the *single* Power8 chip S812LC server.
So you might say, gee, but that’s 32-cores versus a poor, little 10-core Power8 chip. So whats the power consumption of a 32-core Power8 system and how much performance does that deliver?
A PowerS824 with 24-core @ 3.52GHz has power consumption of 2,300 W. That’s 96W per core! A Power E850 with 48 x 3.02GHz cores @ 3,500 W is 73W/core. That’s IBM's slowest/lowest Ghz Power8 chip. So a SPARC M7 is roughly half the Watts/core! And theres several public benchmarks showcasing SPARC T7-1 outperforms Power S824 by up to 40%. You can see them all here:
https://blogs.oracle.com/BestPerf/
So, unfortunately for IBM, the world has significantly passed IBM in multi-core architectures, limited to just 12-cores maximum today, especially with Intel now at 22-cores with E5 v4 and Oracle at 32-cores with SPARC M7. So sure, maybe in two years, IBM will finally show up with 24-cores. But do you seriously think Intel and Oracle will be sitting by waiting?
What's awesome about your half-ass attempts to portray SPARC as anything other than lipstick on a pig while putting down all competitors especially IBM POWER which you seem to have a sick obsession with, is that end up showing how poorly SPARC performs.
You tried to call out the S812LC as requiring 3500 Watts because that is what the IBM Redbook says. I called you out and you admitted you were wrong. As an Oracle troll, you continue to state the POWER8 S824 is a 4 chip system when that IS NOT how IBM classifies it but instead it is a 2 socket server. This isn't unusual for a company such as you work for who is known to purposely misstate a competitors features. It is inconvenient since you want to marginalize IBM POWER as much as possible to highlight your inferior SPARC servers. This goes way back to even when it was Sun (sad to say). One example that comes to mind is with the IBM POWER6 570 server. IBM would refer to it as a 8 processor server which in fact was IBM say 8 cores or 4 x Dual Chip Modules. Sun would then use the fact IBM said "8 processor" even though it was clear it was 8 cores and compare it to a T-series server with 8 processors. In this case, the 8 processor might have 8 or 16 cores because Sun called a socket or chip a processor. Sun would then try to say their T-series was 4 times better performing with their 8 processor vs IBM POWER even though the T-series had anywhere from 8 - 16 times more cores. As I state in my recent blog https://powertheenterprise.wordpress.com/2016/04/17/oracles-at-it-again-stuffing-a-card-up-their-sleeve/, Oracle has ZERO credibility. The only reason I respond to a troll is to not let your lies go unanswered.
You are now searching through IBM docs looking for anything you can criticize. If you knew half as much as you thought you would be dangerous. Right now you are just ignorant. IBM has a public tool which shows real world energy consumption based on how the server is built. As such, a heavily configured 24 core S824 like I regularly sell to customers to replace aging legacy SPARC crap it shows the server would consume 980 Watts at 50% utilization. So, your 8 chipset T7-1 by your own admission uses roughly 1750 Watts. That puts IBM POWER8 at 490 Watts per socket compared to 1750 Watts for Oracles 1 socket server. Break it down by core and POWER8 is 40 Watts per core vs 55 Watts per core for the SPARC T7-1.
Go back to your Mama's basement troll.
Powerboy, you need to go into Politics! Instead of proving or countering my facts with substantiations, you come up with all sorts of stories and arguments that are completely irrelevant. You sure know how to dance and distract! But...I'm sure most readers here will see right through you though.
And as a fact, its you that’s needing some credibility! I don't think many here care how IBM or anyone markets chips or sockets, or even cores for that matter, but what does matter, is what you can buy and run, how much it costs, how much performance it delivers, and how much power does it consume? Sure, theres many more technical & business requirements but these are the basics.
And so as Ive said before:
For Server Price/Performance:
A SPARC T7-1 server, with a list price of $39,821 (w/128GB RAM/2x600GB Disk), outperforms an IBM Power S824 with a list price of $65,291 (w/192GB RAM/2x146GB Disk), by 40% using an OPENSOURCE Database OLTP testing tool, HammerDB.
That’s a 56% price/performance advantage! The results of just one of many benchmarks can be found right here: http://www.hammerdb.com/benchmarks.html, second link down "https://blogs.oracle.com/BestPerf/entry/20160317_sparc_t7_1_hammerdb"
For IBM's OLTP TPM "estimate", which is actually lower than what Oracle tested, can be found here:
http://www-03.ibm.com/systems/data/flash/fr/resources/modernisation_i_2015/S17_-_Linux_sur_IBM_Power,_lalternative_gagnante_a_x86.pdf slide 12 under "OLTP"
Pricing can be found here for SPARC T7-1: https://shop.oracle.com/pls/ostore/f?p=dstore:5:1741895351289::NO::P5_PROD_HIER_ID,P5_LPI:126821605095251205695842,126822405061891214221509:
IBM PowerS824 Pricing here: http://www-03.ibm.com/systems/power/hardware/s824/browse_aix.html
For Server Performance/Watt:
A SPARC T7-1 server with MAX power consumption of 1750 Watts is 40% faster than the PowerS824 consuming MAX 2,300 W.
That’s a 46% Performance/Watt advantage!
SPARC T7-1 MAX power consumption of 1750 Watts found here:
http://www.oracle.com/us/products/servers-storage/sun-power-calculators/calc/t7-1-power-calculator-2398701.html
IBM PowerS824 MAX power consumption of 2,300 W found here:
http://www.redbooks.ibm.com/redpapers/pdfs/redp5097.pdf
Now if you're a real "PowerMan" and not a politician, could you demonstrate (transparently), system to system, with substantiations, where *any* Power8 system shows superior HW List Price/Performance & superior Performance/Watt using a public, Database or Application benchmark vs SPARC M7? Show us you have something hangin between your legs!
Yawn! After Oracle accepts my Power challenge to run them head to head using a real customer workload I'll happily show you the HW List Price / Performance & Performance / Watt. Nothing says believable like a real world workload.
We can also meet at any major US airport hotel for a platform / technology debate.l El Reg has stated they would happily moderate when I first offered this to Oracle over a year ago. Bring your best 2 speakers and I'll bring my wing-man. You'll get a projector and a whiteboard. We will record and publish unedited to YouTube.
Phone lines are open ..... don't call collect, Larry can afford the dime.
@PowerMan4Evr,
Well, instead of you ducking all the hard numbers and claiming your opponents are Trolls, how about you show us some benchmarks where IBM POWER8 is faster than SPARC M7? Put up, or shut up? Do you have any benchmark links that proves POWER8 is faster than SPARC M7? No? Then how can we believe that POWER8 is faster when there are no benchmarks out there? It makes you look like a FUDer?
-Trust me, POWER8 is faster than SPARC! Do I have benchmarks or any evidence at all, to back this claim up? No, you just have to trust me.
This is not really credible, eh? Scientific method is to back up your claims with some kind of evidence. And as of now, there is no evidence at all. That makes you look like a liar and FUDer, right? If you can show us benchmarks, then fine, then POWE8 is faster - but you can not show us hard facts. Because they don't exist, POWER8 is way slower. Wishful thinking does not convince anyone?
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
