Mainframes Find New Life in AI Era (msn.com) 56
Mainframe computers, stalwarts of high-speed data processing, are finding new relevance in the age of AI. Banks, insurers, and airlines continue to rely on these industrial-strength machines for mission-critical operations, with some now exploring AI applications directly on the hardware, WSJ reported in a feature story. IBM, commanding over 96% of the mainframe market, reported 6% growth in its mainframe business last quarter. The company's latest zSystem can process up to 30,000 transactions per second and hold 40 terabytes of data. WSJ adds: Globally, the mainframe market was valued at $3.05 billion in 2023, but new mainframe sales are expected to decline through 2028, IDC said. Of existing mainframes, however, 54% of enterprise leaders in a 2023 Forrester survey said they would increase their usage over the next two years.
Mainframes do have limitations. They are constrained by the computing power within their boxes, unlike the cloud, which can scale up by drawing on computing power distributed across many locations and servers. They are also unwieldy -- with years of old code tacked on -- and don't integrate well with new applications. That makes them costly to manage and difficult to use as a platform for developing new applications.
Mainframes do have limitations. They are constrained by the computing power within their boxes, unlike the cloud, which can scale up by drawing on computing power distributed across many locations and servers. They are also unwieldy -- with years of old code tacked on -- and don't integrate well with new applications. That makes them costly to manage and difficult to use as a platform for developing new applications.
Not so much (Score:3)
Re:Not so much (Score:5, Insightful)
Either they're going to run INCREDIBLY - OUTSTANDINGLY power inefficient code on ancient metal
Ancient? The last "non AI" IBM mainframe, the z15 was released in 2019. The newer z16 does indeed have hardware accelerators for AI, specifically a bunch of special purpose hardware for matrix multiplies, convolution, certain other mathematical ops which are popular. "Stapled on" I mean kinda, but IBM have been integrating high speed, low latency accelerators deeply into their mainframes for decades, formerly things like databases, Java, XML, cryptography and compression. I think rapidly shuffling vast amounts of data vrom various sources to various application specific accelerators is kind of their jam.
Re: (Score:1)
Is there a "CobolGPT"?
Re: (Score:2)
I think that's going to take a while. There's OO COBOL already, so we first need Functional COBOL. Then COBOL running on "Web3" blockchains. And only then GPT COBOL. (-:
Re: (Score:3)
Ancient metal? Who is running ancient metal? The current generation of mainframes is 2 years old, and includes dedicated AI inference hardware on each chip.
Re: (Score:2, Funny)
Ancient metal? Who is running ancient metal? The current generation of mainframes is 2 years old, and includes dedicated AI inference hardware on each chip.
Didn't you even read TFS? It says they use "tacked on" ancient code that cannot be maintained. The AI is written in COBOL back in 1972, and it's going to be very expensive to find programmers.
MOVE HYPE TO AI-BUBBLE.
COLLECT UNDERPANTS.
INFERENCE PROFITS.
Re: (Score:3)
What are you talking about? The "AI" was not written in Cobol, it did not exist back then. Apps were written in Cobol. These days, apps are written in more modern languages. So how come they cannot just compile AI apps for the mainframes? It isn't like they are doing anything deeply mysterious, just process a lot of numbers.
Re: (Score:2)
What are you talking about? The "AI" was not written in Cobol, it did not exist back then.
Umm, I was there. Does anyone on Slashdot understand humor, if it's not marked with a /s or whatever? When did reading comprehension get so bad? Get off my lawn!
Re: (Score:2)
Ancient metal? Who is running ancient metal?
Iron Maiden?
Re: (Score:3)
Extant mainframes weren't designed for the specific compute needs of "AI". Either they're going to run INCREDIBLY - OUTSTANDINGLY power inefficient code on ancient metal, or they're going to staple new hardware onto their mainframe and call it an "AI mainframe".
They're still making new mainframes... you know, the big expandable frames with the highly integrated general purpose and special purpose hardware all in one unit? They absolutely can run specialized hardware, it's what they do.
Re: (Score:2)
Re: (Score:2)
See: extendable architectures.
Re: Not so much (Score:2)
Re: (Score:3)
Modern mainframes are manufactured with tech that is the same or even exeeds that of modern servers.
Re: (Score:1)
The mainframe (zSystems) that IBM sells are just Linux clusters with some amazingly high-availability features such as entire board, CPU and RAM hot swapping. These days the workloads are all run in a container-like system anyway. Modern mainframes are basically a cluster but you have unlimited resources on both hardware and software to make sure it is completely redundant and 'never crashes' because of hardware issues.
You can do a lot of these things on Intel systems that you take off the shelf as well, eg
Re: (Score:2)
Mainframes are more than "just linux clusters with HA". People who want to learn more should read a bit more about mainframes and in particular about the IBM Z and the Telum CPU, for instance from https://ieeexplore.ieee.org/do... [ieee.org], that is used in those systems. And yes, a mainframe can run "Linux on zSystems", but most companies buy them to run z/OS (https://en.wikipedia.org/wiki/Z/OS), sometimes keeping some special applications on an Linux on Z LPAR. It is indeed possible to do a lot with HA clusters, bu
Re: (Score:1)
oh... let's also mention backward compatibility... There is very good software written out there that does not need to be changed and that perhaps will never need to be changed as long as there is hardware to run it, and z/OS is backwards compatible (including binary compatible) with decades ago systems (for some applications, binary compatibility can go back to 1964, except in the case of rare CICS macros that can be modified or software emulated).
Re: (Score:2)
Re: (Score:3)
Re: (Score:2)
Re: (Score:1)
Re:wtf 30,000 transactions per second? (Score:1)
It's FCVFPS: Fake Cat Video Frames Per Second, the standard AI metric.
Re: (Score:3)
Re: (Score:1)
The 30k was probably pulled from the number thrown around around 2015 for the z13. It's of course higher now.
Also no machine was doing 30k transactions per second in 1952 or even 1992. What do you imagine a transaction to be, exactly? How many write transactions a second do you think your workstation can do? Give it a shot with mariadb or postgresql, I'd be curious. And no cheating by disabling file system syncing.
Re: (Score:2)
Also, wtf 30,000 transactions per second? Is this 1952?
Meanwhile, an Intel project I worked on could do more than 4.77 million durable basic transactions per second per thread, on hardware released a few years ago.
Re: Big and Slow trying to remain relevent (Score:2)
Mainframes satisfy needs that a collection of x86 servers can not - there's a reason IBM continues to sell new mainframes, and it isn't just to run COBOL code from 1972...
Back in the day, Sun Microsystems, as anti-mainframe a shop as ever was, had an IBM mainframe to run the organization.
Re: (Score:2)
With a 4.7Ghz processor, that leaves 1000 cycles for you to do a 'transaction'. You won't even have the transaction decrypted and routed to the right handler in that time. A credit card purchase is a transaction. A trivial update to a trivial database is not.
Re: (Score:2)
That's why I said "basic transactions". An atomic, ordered, persistent update to a data structure. You then add actual logic atop of that -- and obviously, all data you use needs to be in a high level of the cache or you suffer a long stall.
That's a basic transaction, a building block you can't sanely skip. Persisting the write is the step that takes longest time, and it's thus something we can benchmark. On the other hand, your "credit card purchase" is business babble that can't be meaningfully compar
Re: (Score:2)
So what was the point of your completely worthless comment? For a really impressive number, why don't you see how many NOPs you can do in a second?
People who are purchasing mainframe-type equipment are interested in how fast THEIR workload will run. Nobody gives a crap how fast a useless transaction can run if everything happens to be in the cache.
Your comment is a perfect example of how to completely mislead people with a so-called "benchmark".
Re: (Score:2)
Mainstream Mainframe? (Score:2)
Mainframes do have limitations. They are constrained by the computing power within their boxes, unlike the cloud, which can scale up by drawing on computing power distributed across many locations and servers. They are also unwieldy -- with years of old code tacked on -- and don't integrate well with new applications. That makes them costly to manage and difficult to use as a platform for developing new applications.
I don't quite get this. I thought you could run Linux on them and if so, why would they not be treatable as any other Linux server?
Re:Mainstream Mainframe? (Score:5, Insightful)
It's bullshit. You can write new applications for mainframes just fine. If you've got a mess of hacked-up legacy software on your mainframe, that's on you as well. You can do the same thing on a regular server (or PC for that matter).
Re: (Score:1)
Re: (Score:2)
Sure, but that isn't what the article is saying. The article is an incoherent mess of buzzwords.
Re:Mainstream Mainframe? (Score:4, Funny)
Re: (Score:2)
We have no Turing complete machines. Look up the definition, they need an infinite amount of tape. Last we checked, memory in our machines is finite.
Re: (Score:2)
Nonsense.
To be Turning complete does not require infinite memory, it's requires unbounded memory. Also, just having unbounded memory is not sufficient. For example, push down automata have unbounded memory but are not Turing complete. This is CS 101.
We absolutely do have Turing complete machines. You're sitting in front of one right now. We know that it's Turing complete because it can simulate a universal Turing machine. Given sufficient time and resources it can compute anything that is computable.
Re: (Score:2)
AI came to mainframes years ago... (Score:2)
Same Old, Same Old (Score:4)
I've been hearing predictions about the imminent death of the mainframe since the 80's. I'll bet that I'll continue to hear that mainframes are sure to go extinct well into my eighties.
Re:Same Old, Same Old (Score:5, Insightful)
Mainframes offer something the cloud and even a regular local DC cannot: Ultimate reliability and availability. They will not go away.
Re: (Score:2)
yet for some reason, we're no longer hearing about the imminent death of the mini.
*sigh*
no, kids, we're not talking about skirts . . .
Standards are forever (Score:1)
Mainframes are either the standard or the results of which concerning all things microcomputing. Vacuum tubes (really!), solid state circuits, the asymmetric memory channels, processes, interprocess communications,..... All developed with mainframes as the target. And most before the modern mainframes were even a standard.
hold 40 terabytes of data (Score:3)
hold 40 terabytes of data
If I stuff all five drive bays on my tower PC with 8TB EVO SSDs, I have a mainframe!
Re:hold 40 terabytes of data (Score:5, Informative)
It's 40 terabytes of RAM, not disk.
Re: (Score:3)
IBM has something called RAIM. https://en.wikichip.org/wiki/r... [wikichip.org]
Why are they called Mainframes ? (Score:2)
$stty -echo
We are not in an "AI era" (Score:2)
We are in an AI hype. The hype will die and the rather mediocre actual advances are not enough to qualify as an "era".
Re: We are not in an "AI era" (Score:2)
Some people still think 3D printing will change something. Let them live in their fantasy world.
Re: (Score:3)
Skirting the big question (Score:2)
Study (Score:1)