Here is a good high level slide deck on z/OS 2.5 new features: https://github.com/IBM/IBM-Z-zOS/blob/main/zOS-Education/zOS...

It starts out with the marketing buzzwords but then drills down into the technical nitty-gritty. In terms of what the "AI" stuff actually means, one answer is contained on slide 121 – support for running Tensorflow and ONNX inside a Linux Docker container using zCX (which lets you run z/Linux Docker containers under z/OS). The zCX SIMD support referenced on slide 102 is probably highly relevant to this.

If running containers on Mainframes is of interest, attending this talk at Defcon might be illuminating https://defcon.org/html/defcon-29/dc-29-speakers.html#coldwa... :)

How does that compare dollar-for-dollar to running Tensorflow on a dedicated Linux cluster with GPUs?

Mainframes also have a lot of PCIe channels you can attach GPUs to. IO capacity and hardware offloading has always been the most notable difference between a mainframe and an x86 box.

I wouldn't be surprised if, with some clever programming, one could pipe data from disk directly to GPUs and back to disk without involving the CPUs that run user programs except for the thinnest slice of time.

Do the specific Nvidia cards you'd ideally want for ML have s390x drivers?

That's something you'll need to ask Nvidia.

But, considering what you paid for the mainframe, you probably have the budget to commission them to write drivers just for you (or to license their source and compile it yourself).

(or to serve them a National Security Letter that says "give us your source" and compile it yourself)

Can’t stop laughing sadly.

Btw given they do linux can they twist so that channel is partition into for that linux container. In the new world just driver through.

IBM has built supercomputers with POWER machines stuffed full of GPUs, in close partnership with Nvidia, for years. I think they might still be the only ones offering systems designed for sustained NVLink bus saturation. So yeah, you could do it - and it would be awesome... but it would be a silly misuse of a mainframe where a cluster would make so much more sense.

Probably much more expensive.

Probably depends on whether you've already got the mainframe or not, and whether you've got the dedicated Linux cluster or or not.

"one answer is contained on slide 121" => right there, I decided not to open the slide deck.

You have to go through 100 slides before getting to technical details?

You have to go through 100 slides before getting to technical details for this narrow specific particular topic that is being discussed here

A new OS is a significant endevour and there is a lot of release notes on any and all topics - technical, usability, timelines, compatibility, hardware, architecture, support, upgrade methodology, etc.

If this was the old IBM, back when they had an internal science journal just to keep track of their own accomplishments, I'd say they came up with something like z/TPF[0] - but for AI. But I would be extremely surprised if the new IBM actually started pushing the envelope of computer science again. Not until they reverse course on the whole "Lets get rid of all our hardware expertise and just focus on software" thing.

[0] https://en.wikipedia.org/wiki/Transaction_Processing_Facilit...

IBM is still making deeply custom chips for the z-series mainframes, running them at a higher clock rate than anybody else, and applying water cooling at the system level. Also they have a bus that intermediates between the CPU and RAM, someday the rest of the industry will catch up

https://www.computeexpresslink.org/

Oh boy, that memory bus... I'll get to that in a second. First, yes, IBM creates IP blocks that they then license. Maybe I'm weird, but I view that as software - and I also heartily dislike IP block shops. What makes matters worse is how IBM has an awful track record when it comes to these licenses - they somehow always pick a partner that ensures the tech seeing little to no use. For example, you mentioned their memory bus... guess how they did that: they licensed (or sold?) Centaur to a single party, Microchip, who lists[0] a single memory controller with no price and no apparent interest in actually selling the thing. Also, Synopsys - one of those fabless companies that I love so much, is somehow involved[1]... resulting in memory controller binary blobs. It is pretty awesome really, I don't think IBM could have more surely murdered their tech if they'd tried.

Now I'm pulling for them, I actually own a POWER9 machine. I think it is great that they want to get memory controllers out of the CPU - but they've done it in such a way that leads me to believe that it is an intentional sabotage, which would only happen if they genuinely had no interest in an active hardware role.

[0] https://www.microchip.com/en-us/product/PM8596

[1] https://github.com/open-power/ocmb-explorer-fw/blob/6570112a...

A while ago I came across a really cool technology I wanted to use. Turns out it was patented, and the patent was bought by IBM.

IBM had merged it into Watson, kept the acronym, but changed the words. Then they stopped making it available in any meaningful way.

I was so pissed. I’m sure my company would have paid out the nose for it.

I got the feeling it was a capture-and-kill situation.

They've got a hefty patent portfolio, but I can't think of a recent example of them using it to freeze everyone else out à la eInk. Of course, the barrier to entry into their market is so incredibly high that there would be only a small number of companies who'd actually be able to test that. The only instance of them really going after someone (over IP) that I can think of was in the 80s when Hitachi was pretty brazen in the way they were stealing source code. Very low effort, like: Step 1 - acquire from a subcontractor the entire source tree for IBM's current mainframe OS; Step 2 - sed s/IBM/Hitachi/g; Step 3 - package the OS with Hitachi mainframes sold in Japan. They got along famously though, after Hitachi lost the legal battle and started paying them a royalty for the rebadged OS.

IBM does have a reputation for buying competitors in very niche areas, like process planning software, and then not doing anything with them except maybe parting them out several years later. It looks suspicious, but I think that is more to do with the way they stick to hyper-specialized enterprise solutions... the pickings are naturally very slim.

> They've got a hefty patent portfolio, but I can't think of a recent example of them using it to freeze everyone else out à la eInk.

I work in the display industry. I've never heard that e-ink "freezes everyone else out" from colleagues, coworkers or even gossip at conferences. Surely, if this was true we would have heard about it. Could you elaborate on your claim and if you experienced this? Please see my comment history to see why I keep asking about this.

lol, I'm already familiar with you. I'm not really interested in generating billable hours for someone who behaves exactly as I'd expect a reputation management rep to. But I am curious to know if your script addresses this logical contradiction:

Electrophoretic display technology has a barrier to entry that can be vaulted by anyone capable of standing up a moderately sophisticated light industrial facility.

Electrophoretic display technology is so useful that E Ink Holdings Inc rationally leverages the legal protection of the USPO through a large number of patents to maximize profits.

Electrophoretic display technology is so useless that consumer demand is only able to support a small number of manufacturers (who are able to demand very high prices).

Ok, if that's what you want to believe then I can't change your mind. It is pretty clear to me that this is equivalent to me as a display guy claiming Microsoft is blocking progress in operating system development just because they have a lot of patents and they maximize profits. I hope it is clear that I think this is not a valid argument. Also, I'd encourage you to live up to the spirit of HN and try to provide evidence instead of being rude to people who challenge your claims.

> It is pretty clear to me that this is equivalent to me as a display guy claiming Microsoft is blocking progress in operating system development...

I'd need to have a very low opinion of a "display guy" to believe he was genuinely arguing in good faith, and in keeping with the spirit of HN (hallowed be its name), that the market forces an OS is subject to (like user training with regard to network effect) have any overlap with display hardware (which is unaffected by prior user uptake). Because anyone arguing that would have to be ignorant of both subjects... but why would such an ignorant person be so invested in white knighting a hardware company? Oh well, doesn't matter - my opinion was already arbitrarily settled upon for completely irrational reasons.

> a very low opinion of a "display guy

> HN (hallowed be its name

> white knighting

You clearly have very high opinions of yourself and your belief system. It is clear you are unwilling to respond with facts when simple questions challenge your claims and react with hostility. I don't think there's any beneficial discussion possible going forward. Best of luck in your endeavours.

> I think it is great that they want to get memory controllers out of the CPU

The trend is out of the CPU now? I was hoping they would instead start layering DRAM stacks on the CPU die.

The writing has been on the wall for a while when it comes the next major bottleneck. I'm surprised that Intel was able to keep their quad-core game running for as long as they did, but "chiplets" were obviously going to have to happen, which stresses IO. IBM solved that a long time ago - as demonstrated by the massive IO bandwidth on display for several hardware generations. POWER9 came in two flavors, one with on chip memory controllers and one with their traditional off die strategy. So definitely not new - but its always been very boutique and therefor very expensive. Their OMI proposal was supposed to change that - but it won't, given the awful way they've executed it. Anybody remember that time that IBM's PS/2 offering had a major positive impact on the industry and didn't lead to a mess of dead ends? Me neither... this will be like that.

It goes both ways.

HBM on the die is crazy high performance.

With a bus like CXL you could have a card full of memory chips that is shared by multiple CPUs. The z machines have a system like this it is just proprietary for them.

Memory that is less coupled to the CPU won't be as fast as HBM but you can have a lot more of it.

I'd love to do mainframe development. Sounds like it would be fun. I could never afford one though and my industry doesn't use them either. Uptimes of decades is seriously impressive.

You can! You need to install the Hercules mainframe emulator and the freely available MVS 3.8 on top of it. There are compilers for COBOL, Fortran, PL/1 and assembly available. The go to guy on YouTube for getting all this working is moshix:

https://www.youtube.com/c/moshix

How close is it to the real deal in emulating something like a modern mainframe?

Very close to the real deal only when emulating a 370 from the 80s. zOS is a different beast though even though it shares a lot of the asm underneath. But there is no way to run zOS for yourself without a license from IBM or access to a real z-series somehow.

Kinda what I figured. Th whole concept of a mainframe is pretty neat in that you can go through the 1000 page manual and have an idea about all that is possible, instead of an infinitely changing software tool ocean. The latter is probably better overall, but it would be neat to check out the mainframe paradigm.

Running z/OS on Hercules is trivially easy, you just need to ignore the license. The z/OS software distribution is usually called ADCD.

Yeah, I'm only interested in using software legally, but it's a shame IBM doesn't make this available. It's not like someone is going to just port all their mainframe apps to run on this emulator right? Or is that a real concern of theirs? I figure the hardware itself is the secret sauce.

I get the impression (as a vaguely interested outsider looking in from a distance and making a LOT of presumptions) that it's "just" institutional drag and "tradition". Create a culture founded upon filling in forms and ticking boxes, and "hi I can haz?" is going to produce $every_windows_error_at_once.wav, a lot of hand-wringing, and an internal processing chain that'll naturally tumbleweed itself in the general direction of the legal department. Network effects / emergent behavior. Soo, you read between the lines ("ah, it's fine, the requirement was fulfilled"), and take it from there. 𝙻𝚘𝚐𝚘𝚗 ===> _

The only reason I haven't pursued this myself is that MVS is ultimately just not objectively compelling enough, because while it's similar to what's out there today, I get the impression I'd only have to unlearn an arbitrary bunch of things once I was on the "real" system.

As an aside, I happened to stumble on a copy of "IBM and The Holocaust" a few years ago. It's a startling insight into IBM's beginnings tabulating and wrangling data on the millions of Jews "processed" throughout Germany in WW2, which turned out to be the first commoditized consumption of what would today be regarded (without a second thought) as big data.

The world has certainly moved on since that point and IBM's DNA is completely different today, but IMO the "data processing" energy still remains as an incredibly strong centerpoint, and I honestly see mainframes as a digitized realization of the punched-card concepts IBM's beginnings were founded upon. (Combined with the institutionalization aspects I hinted at in the first paragraph, this is the main reason I'm very hesitant I'd survive in what sounds like an especially rigorous environment.)

While the UI of reading a printed book hasn't yet been beaten (I'll switch to e-readers once they can pageflip instantly -.-), it's very practical to be able to virtually pick up a book and thumb through any portion. To this end, PDFs of this title seem to readily be floating around, for example (open as URL) data:text;base64,aHR0cDovL3Bvc29oLnJ1L2Jvb2svaHRtL2libS5wZGY

I will note that after reading I kinda looked askance at my ThinkPads a bit and went "wow, what a history. uhhh....". Fair warning :)

I should have said 'legally'.

I used to fix compilers for IBM and we worked on (older versions of) these machines. They were awesome! Coming to the PC business after using mainframes was a sad joke. You should see what it's like working with VM, IBM's virtualization environment. It was amazing and that was 25 years ago.

It is so stressful and alienated that the decades spent on mainframe is good but not great. I worked on the partition Management called lpar. Os and db2 etc. It is sort of more interesting to do dos/vse. I guess instead of thousands of users you have hundreds.

I love my Hp 200 dos calculator whilst waiting for the cartridge to run, … at least it is not open reel tape as the cartridge tape at least always suck in.

It’s interesting that IBM wasn’t sooner to the cloud with superior virtualization technology.

Just like the PC, they didn't realize it was important until it was too late. In my experience, mainframes are a superior way to develop and run every kind of software. The PC is a crippled, stupid stepchild by any reasonable comparison. Just the idea that you need to setup and manage 1,000 of them to come close to a small mainframe is ludicrous. Who needs the overhead and maintenance? Sure, you can swap machines and distribute loads between them, but who wants to? Yuck. It's a lot nicer to have that stuff handled automatically.

Also like the PC, IBM didn't realize that their boxes could be used for this new funky thing called the internet. Talk about blind-sighted. Their first web product was an abomination they inherited from their purchase of Lotus. It didn't work and produced unusable websites, boding poorly for IBM's entry into that market which was already taking off with Active Server Pages and XMLHTTP.

If IBM had loudly (and correctly) proclaimed that mainframes were the place to run web services, everyone and their mother would be running mainframes right now instead of Linux boxes. IBM's eventual realization of this led to their purchase of Red Hat!

> If IBM had loudly (and correctly) proclaimed that mainframes were the place to run web services, everyone and their mother would be running mainframes right now instead of Linux boxes.

No, they wouldn't. I could take one Linux box, get a domain name, and have a web server. I wouldn't be able to afford a mainframe, though. And if your alternative is renting a fraction of a mainframe, the answer is no. No, I will not trade a Linux box that I control for renting a piece of a mainframe that I don't control. Just no.

For enterprise is what I meant. Not for individuals.

Well, not just mainframes generally, but if they created AWS on mainframes and hosted them… that would have made it accessible for startups.

It was so simple internally, they could have ported those tools to a GUI. But GUI's and user interfaces are not something IBM has ever done well (other than command line ones).

The organisation is also in the way. Both in Ibm and for those teams using ibm. Very bureaucratic and like token ring over management kill. Good and have its niche. But murder by the liberal Ethernet.

I would argue neither has AWS yet…

The most logical place for a cloud service to develop inside IBM would have been IBM Global Services. IGS had a sort of proto-cloud set up in 1998-1999 (treated physical systems as cattle that could be remote wiped, virtualization would have been a logical next step) but sold off the organization building that to AT&T as part of the Global Network sale.

I wrote a horrible, hacky, CMS PIPES based gopher/http server for VM some time in 1993 that was used internally for various "non serious" things. A colleague and I designed a native HTTP server for MVS in 1995 and had started coding it when we were more or less reprimanded and told to defer to the then Network Services division which was porting something (I think CERN httpd, they had a bizarre fascination with CERN) to MVS.

It just wasn't in anyone's best interest to even experiment with web services on VM or MVS, let alone try to build a business out of it.

Cloud happened after I left, but the infrastructure to build or respond to the rise of AWS was never in place, sort of by design.

Sidenote: In VM, you could spawn a new machines with "physical" storage or drives (Nothing was physical on the mainframe; everything is faked just for you! Awesome.) by typing a single command at the command prompt! If you needed to fix a bug in some combination of operating system and IBM products, you could spawn that exact customer-replicating environment in about 2 lines of text.

Try that with Kubernetes!

LPAR makes LVM a joke.

It's sad to read about the dev headaches caused by today's poor virtualization environments. I distinctly remember that stuff was no work at all at IBM. It didn't interfere one bit with the actual work of coding on the products. In fact, we really never thought about setups or creating environments for testing or development.

In peer reviews, for example, it was standard to bring output from both the failing case and the working case as printed handouts. Those were easy to generate; just spin up two machines and fix one while you leave the other broken. They can share resources if you like so you don't have recreate everything. That is horribly messy today. Testing is so hard most people don't do it.

Also on simpler BSD's like OpenBSD, bioctl is dumb simple. Setting up a encrypted disk is a no effort task.

On firewalls, PF has a syntax simpler than most text adventure games.

Every time I get into the GNU world, well, I get better HW support sometimes, among a slight bit of performance, but, FFS, configuring a Linux firewall gives me nightmares everyday. Just compare NAT on Linux vs OpenBSD. Or TUN/TAP devices, where in OpenBSD it's just a matter of a SINGLE line at pf.conf and two hostname.if files to define the bridge and the tap device with a subnet. Done. Oh, and sysctl to set IP forwarding, but that was the easiest task on any OS.

In a sense, a parallel sysplex with remote hosts and storage is a very sophisticated private cloud environment that'll run your CICS transactions a lot like you do with AWS Lambda.

IBM has a event where you can get access to a mainframe and complete different tasks https://www.ibm.com/it-infrastructure/z/education/master-the...

You can also get free access to small vps running Linux on a mainfram on the IBM cloud

> You can also get free access to small vps running Linux on a mainfram on the IBM cloud

What's the point? VM with linux is hardly interesting to anyone.

Amusingly, I signed up for this a couple days ago (https://mtm.masterthemainframe.com/), which worked fine at the time, but the following day when I went to follow up I discovered the HTTPS cert had expired :D

That was several hours ago. The site's still down.

UPDATE: THE DOMAIN EXPIRED HAHAHA

Archive link: https://web.archive.org/web/20210729160256/https://mtm.maste...

Update-update: Bureaucracy, Inc™ has belatedly noticed and exercised untold power to bring the domain back in a great hurry.

Everything's back to exactly how it was before, even down to using "" (ie, empty string) where my name should be. Excellent.

I must say that site responsiveness is much, much slower than it was when it was serving spam. :D

That is pretty amusing, I also just received a mail from them telling me that my opinion was important to them. Heh

I used it when I was doing some porting of smaller programs to run on z/Architecture for example setting up the de/compression co-processor correctly. It's true that Linux on x86 or arm is not unusual, but this is Linux on Z

"Sounds like it would be fun."

Interesting at times but really really not fun

Interesting to me is a big part of the fun. Learning something new...all that.

Well "to do mainframe development" is more than just the casual learning part.

Touche'

There is also stuff like a sorting co-processor which I find interesting https://blog.share.org/Technology-Article/peeking-under-the-...

Perhaps it is COBOL bindings for the Watson API?

I think it is an AI-processed dialect of COBOL and INTERCAL where you have to prepend PLEASE every few statements to get the compiler-AI to parse it.

In order to get the compiler to optimize your code, you have to say PRETTY PLEASE, though, and THANK YOU VERY MUCH. ;-)

    PRETTY PLEASE
        ADD SALES_TAX TO PRICE GIVING NET_PRICE
    THANK YOU VERY MUCH

Yeah, that's irritating. I do recognize that mainframes had many of the things, forever, that are buzzwords now.

But trying to tie an extremely compute intensive requirement (AI) to an environment with probably the highest $$/MIPS is just...dumb.

If you’re doing transaction processing on the mainframe I could imagine wanting to score something (eg fraud) mid transaction, which might call for some additional platform capabilities. But it’s also plausible someone in marketing demanded more AI.

I can see that, but mainframes are perfectly capable of calling out to external services, and that's not new. I guess I really don't know what they are saying in regards to AI and mainframes.

I do not know if IBM will do this with the AI work.

but in many other similar circumstances they give the option of buying Application Assist Processors. That takes the work away from the main processor.

It is possible that IBM will offer some custom zAAP optimized for running the AI workloads.

IBM also open up running AI in their cloud.

> but in many other similar circumstances they give the option of buying Application Assist Processors. That takes the work away from the main processor.

zAAPs, zIIPs and IFLs are exactly the same hardware as main processors. The difference between them is purely at the licensing level. The firmware sets a flag saying "don't run classic z/OS workloads". z/OS sees that flag set and then will only schedule "new" workloads (such as Java and XML) to that processor. Actually there is an undocumented API you can use to schedule any (enclave SRB mode) workload to the zAAP – IBM keeps it under NDA, and ISVs have to agree to obey IBM's rules about what workloads are allowed to run as part of that NDA. (One ISV, Neon, started selling a tool to run an arbitrary workload on the zAAP, called zPrime, IBM immediately sued them, and Neon settled the lawsuit in 2011 and agreed to drop the tool). Other operating systems, such as z/Linux, ignore that flag completely and to them a zAAP is just the same as a normal main processor (CP).

> It is possible that IBM will offer some custom zAAP optimized for running the AI workloads.

Yes but that wouldn't be a zAAP. They do have hardware accelerators for various functions (crypto, compression, sorting) but those aren't zAAPs (or zIIPs or IFLs). Crypto acceleration is CPACF (extra CPU instructions on main CPUs) and Crypto Express (external accelerator card). Compression hardware acceleration is zEDC. Sorting hardware acceleration is the Integrated Accelerator for Z Sort.

AI inference on existing model instances need not be very compute-intensive at all. The "training" (i.e. model fitting) phase is what requires a lot of compute resources.

They are talking about customers spinning up Linux instances on their existing (mostly not Linux) mainframes, and special software to allow SIMD calls from Tensorflow to work, etc. I think that making a remote call out to a commodity box is probably easier in most cases.

Maybe there's going to be new AI specific hardware on new Mainframe processors or AI accelerators in the future that will require z/OS 2.5 to run but they are not ready to announce the hardware? I think buzzwords is more likely though.

I think they have been making a push on this front with POWER so maybe this is leading up to a similar hardware push for z systems. Still guessing buzzwords