AI smarts: IBM pushes out 'faster than X86' POWER9 servers IBM has launched its first POWER9 server, the dual-socket AC922, saying it is designed for compute-intensive AI work, speeding frameworks like Chainer, TensorFlow and Caffe. It has two POWER9 chips and from two to six NVIDIA Tesla V100 GPUs, and shared CPU-GPU memory. The CPUs can have 16, 18, 20 or 22 cores and up to 44 cores …
I'm an ex-IBMer and while I never worked directly in that part of the company, I've always been quite a fan of Power and of AIX (although I know the latter has plenty of detractors). Power offers an alternative to x86 and has always offered an architecture which, with the right code, can be exploited to produce some serious benefits.
The problem has always been the fact that a lot of code and applications are simply not available for Power, and there has been little innovation in the application space on Power architecture. It's much the same reason why Arm hasn't taken over data centre space even though it monopolises consumer products. You can't just recompile code for a different architecture and expect it to fully utilise it.
Power has always been held back by IBM, mostly by their insistence in keeping it so close. Proprietary hardware and protocols like Microchannel, Token Ring and SSA strangled it when it should have been making inroads, and allowed the x86 business to dominate. So Power always remained a niche, just like its competitors from HP and Sun, or its own mainframe business. It's impressive that it remains a force to be reckoned with, albeit a relatively minor one.
Unlike just about every decision IBM have made in the last decade, the decision to open up the architecture and to partner with the likes of Nvidia was a good one, albeit probably far too late. Ideally they should give it a lifeboat and sell it, which may allow it to flourish. I hope it survives personally. A competitive market with x86, Power, Arm, and hopefully others is good for the industry.
I fear you may be right though. IBM will most likely just sell into existing IBM shops; fewer and fewer as each talented engineer retires or is "resource actioned".
Yeah right, and Google are too stupid to realize that.
1) It’s not only IBM, read about Open Power,
2) And it’s not the cost either; Power is still competitive for some workloads and will likely remain competitive simply due to its architecture being most suitable for some workloads.
'Also a) it's IBM and b) the cost.'
IBM? ok, I'll give you that..
Cost? Keep a eye on the auction sites, I've seen Power7 systems show up and go for silly amounts (e.g. the last one I saw was a 750 Express for less than £200..admittedly, that was without an OS installed.)
I have an open-source software project that some users may be interested in running on POWER, but I've no practical way of testing it. It's a set of C libraries which can benefit from some serious CPU oomph, but I can't claim that it works on POWER.
* It runs on Linux, BSD, and WIndows.
* It compiles using GCC, Clang, and MS VC.
* It runs on x83 32 and 64 bit.
* It uses SIMD on x86-64 for a significant performance boost.
* I have an automated test pipeline set up for all of the above, with tens of thousands of tests.
* I'm working on setting up low cost test systems for ARM and expect to be able to claim that as a tested and supported platform.
But what are my options for POWER, other than splashing out on some eye wateringly expensive hardware despite having no paying customers lined up for it? Not good it would appear.
There's a bit of a chicken and egg problem here. I've written some open source software which can really make use of POWER's supposed advantages, and no doubt other people are in the same boat. However I can't call it a supported platform. Users can either take on the testing and any required porting themselves, or they can just use x86 - where my x86 specific SIMD optimisations may overcome any nominal performance advantages of POWER anyway.
I don't know what IBM can do to address this, but it's one of the disadvantages they are labouring under. SPARC is in the same boat. RISC V is going to have to address this question as well, if they want to break out into the general market rather than just embedded niches.
There are clouds with POWER CPUs, but I have not looked at pricing. You can see some listed at https://www.ibm.com/power/solutions/cloud . There is also https://ptopenlab.com/cloudlabconsole/ , but I am not sure what that is. And finally https://openpowerfoundation.org/?resource_lib=ibm-power-development-cloud which is perhaps most suitable.
Lenovo produce (non-IBM) POWER servers running Linux. You don't get to run AIX or IBM i on these but then you didn't really want to do that did you? Only POWER8 for now but expect POWER9 to hit that channel sooner or later.
And they're a LOT cheaper than the IBM-branded servers - competitive with Intel at the hardware level.
The problem has always been the fact that a lot of code and applications are simply not available for Power, and there has been little innovation in the application space on Power architecture.
There is actually quite a lot of software that runs on power.
RedHat offer a full RHEL stack for PPC, which includes all the usual open source software from RedHat - JBOSS, Apache web servers, compilers (gcc etc), email clients, browsers, and so on.
Of course IBM offer a lot of its enterprise software on power, DB2, various (although not all) Websphere products (Application Server, Java, DataStage and so on).
Oracle supports its RDBMS system on PPC.
There is an optimized 'R' statistical package for PPC.
And many others.
>There is actually quite a lot of software that runs on power.
Oh, I know, and I don't deny that. The thing is, there is also a lot more software which doesn't. And I doubt Red Hat spend a huge amount of time optimising the software they provide for Power. It's all open-source and most of it isn't written exclusively by Red Hat. Any changes they made would result in a branch and keeping that up to date with future patches and releases would be time-consuming and ultimately not worth it.
In contrast, IBM have good reason to optimise their own software as they make a lot of money on it. As do Oracle.
We have, sitting both in a rack and in our boneyard, a quartet of Power 550s using the Power6 processors.
Beautiful machines, faster than a greased porpoise in a sea of snot (at least once they were started), and more horsepower than a good portion of our normal infrastructure at the time. (~2011) AIX ran quite nicely on them. Shame about the application they were bought for, though.
(Seriously, it was written entirely in BASIC, and had such entertaining features as a single user accessible user database- i.e., you you had to edit the user list, no one else could log in to the app.)
I compared it to having an F1 racecar, but with the transmission locked out of every gear except first and second, and liberally covered in bondo, baling wire, and 'hello kitty(tm)' stickers.
The only reason we are keeping them around is for data retention purposes. I've love to see what the hardware could actually do if it had something proper to run on it.
Have been involved with a client who has run POWER for many years extremely successfully. They still have several POWER 7 795 machines on the floor (out of maintenance but running happily). In over six years there has been only 1 hardware issue (caused by an overzealous engineer who pulled a wrong cable).
They run POWER VM and AIX and that is where some issues arise---not the products themselves but rather the lack of suitable skills to maintain and optimise both the virtualisation firmware and the OS.
Unfortunately none of the support guys ever played on a mainframe with MDF, PR/SM or MLPF and virtualisation on x86 is trivial (asking for trouble there).
They could have gone Linux but converting/testing/migrating over 300 logical partitions running AIX was a step too far.
Have got a couple of big POWER 8 boxen on the floor now.
The capex for thes products always look horrific cf other platforms but if you do proper costing anaylses over 6 years and rising, with guarenteed performance and availability POWER certainly competes against other solutions financailly.
As much as I love Power you have hit on one reason it fails to make any significant inroads. Proper costing just doesn't happen.
Purchasing is by line of business and project managers. No one has the money or want to 'but the bus' either. Once you have to start talking about TCO you are getting into the same place when trying to sell a mainframe.
Add this to the fact that x86 vs pSeries is almost like a religious war (when talking about utilisation), lack of skills and being made non-strategic by a lot of clients you dont have a good story.
OpenPower is actually a good initiative but suffers greatly from ISV support. IBM do not help themselves at all either in not providing enough routes or incentives for certification. They also are mostly clueless when going to market.
Crap turned into an IBM bashing session.
Gents,
The POWER architecture is quite simply awesome! It is very different from x86 and some concepts take a bit to get your head around. I have worked with IBM i (AS/400 when I first started) for almost 20 years. The machine never misses a beat!
As for applications and code to run on them, you need to look Enterprise. Banks run them. Large Retailers. Insurance companies. There's a niche for this platform everywhere. The compute workload on a large scale multi-user database is something Microsoft dream of.
Seriously, an awesome machine with high Capex to start but, as another poster said, over a period of 6+ years will more than pay for itself.
Oh, did I mention its never been hacked?
Have a wee read of: https://www.scss.tcd.ie/SCSSTreasuresCatalog/hardware/TCD-SCSS-T.20121208.068/IBM-AS400-technical-introduction.pdf to get an insight into this truly magnificent machine.
You're mixing up AS/400 with the hardware it runs on. I agree with what you say about AS/400 (everyone still calls it that, right?) but in reality they could have made it run on x86 or whatever if they really wanted to - that's one of the main points of the OS - a virtualisation layer between code and hardware.
The cost of AS400 is something above and beyond Power. After all Power running Linux costs less than Power running AIX costs less than Power running i. All on the same hardware. Or at least it did last time I looked - feel free to correct me if this is no longer the case.
A lot of enterprises do run AS/400 as you say - I have seen plenty of instances. There are plenty of them that see the value of them, as with the mainframe, but equally plenty who have either got themselves off, or are desperate to.
Re; the relative costs - you're paying for the database in every case and for Linux you're (not) paying for the OS, only the support.
Now it's true that IBM i (AS/400 to you dodos) does need some microcode support for internal disks but honestly, in this day and age who uses internal disks? OTOH, once you have the IBM-suported hardware in the box you do get a really high level of stability, for the same reason you do with MacOS and iOS (and IOS for that matter) which is that the OS and hardware are tested, developed and certified together.
Late to the comment trail, I know, but AS/400 was the hardware platform, and used to have it's own processor types, although they adopted (and some say saved) IBMs PowerPC processor platform, with Rochester picking up 64 bit systems with the Amazon (RS64) processor when Austin dropped the ball with the failed PowerPC 620, which barely saw the light of day outside of IBM.
IBM i was previously called OS/400.
One reason that IBM i persists is because it is a very business friendly system. Before things like Apache and the other open software packages were grafted on top of the POSIX compatibility layer, many organizations did not employ specialist system admin/operations staff. It was sufficiently simple and menu driven that the general running could be given to ordinary admin staff with little training, and all of the hardware type stuff was handled by IBM CEs.
But it is a propriety system, and you have almost complete vendor lock-in, which is why most consultancies will suggest ditching them. But that does not mean that they could still be the best solution for some companies.
IBM i has been hacked, both in labs and in the real world. Or I should say that Java, Apache, SSL and the like on IBM i have all been hacked.
Of course it can all be fixed in a proper implementation - but the default box won't be secure as-is in this new world and the skills to make it so are hard to find. Pro-tip, get your users in groups that don't have any members with high privilege and remove *PUBLIC authority, then add SSL support to interfaces (5250, DRDA and FTP looking at you here) and push it up to TLS1.2. Get current on the OS to avoid the SMB1 issues (IBM i doesn't have the issues, but supporting it means you expose your network) and keep it patched and audited.
So, not that different to any other system really except that doing the above won't give you grey hair (though you might already have it if you're an IBM i professional).
Good points made. Lost count of the time spent convincing people of the risks of the default FTP server config, and enterprise vendors who insisted all their app users had to be members of QPGMR.
I agree it's "not that different" in that it all needs careful configuration, but doesn't the architecture offer fewer attack vectors? Hence the "never been hacked" claims.
Am I right in thinking that the current "fileless" exploits (pushing a replacement *OBJ into main storage without detection) haven't been done on a correctly configured IBM i?
IBM i has been hacked, both in labs and in the real world. Or I should say that Java, Apache, SSL and the like on IBM i have all been hacked.
There have been vulnerabilities in the OS itself as well, going back as far as OS/400. There aren't a lot of OS/400 CVEs, but there have been some.
Most are information-disclosure issues (WRKUSRPRF, LDAP server enabled by default, ...), but there were some more serious ones. Most notable is probably CVE-2006-6836, for bugs in ASN.1 parsing[1]. Details on that one are scarce, but it got a CVSSv2 of 10.
And the '400 and its successors / rebrandings have been a target for hackers over the years, with significant success. While it has some advantages (the capability-esque architecture, address validation in the single-level store, etc), it's not magic.
I have a certain perverse fondness for OS/400 (aka System i aka just plain i because yes, IBM, in an era of web searching that's a terrific name for a product), with its semi-capability design and its industrial-buildings-over-the-ruins-of-Future-Systems landscape. I did most of my '400 work back in the Good Old Days on a little baby B-class with a twinax-connected 5250, using the PDM development system (like ISPF with much less functionality) and SEU editor (like XEDIT with much less functionality). Builds of our product took like half an hour. Good times. But I never assumed OS/400 was bulletproof, and indeed it was not, and is not.
[1] As always. ASN.1 is a gift that keeps on giving. It's both horrible as a specification and a generous source of security holes in its implementations.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
