Why Do Mainframes Still Exist? What's Inside One? 40TB, 200+ Cores, AI, and more!

20 thoughts on “Why Do Mainframes Still Exist? What's Inside One? 40TB, 200+ Cores, AI, and more!”

• But can it run Doom?

• 2:36 'mainframe' with a z-what, not a Power(tm), uh, I was thinking this was about the larger Supercomputers. The z16 is cool and all, I just feel a bit let down that you didn't get to see Summit or Sierra. To paraphrase someone working with DNA protean simulations a few decades ago, a mainframe is a machine that is a generation behind what you really need to do the simulations you want to run today, lol.
  8:40 OK, I do like that PCIe expansion slot setup, that's nice for things like lots of AI compute cards. Also good point about the AI scanning of transactions, I was thinking more along the lines of 3D mapping of GPR data of other planets kind of stuff.
  18:22 The same goes for Supercomputer centers, aside from the 3-phase power there is no copper wire, it's been fiber for a few decades now.

• And where are our small Power-based systems, Z systems are beasts. But Powers vastly outnumbers them, while sitting somewhere in between of pure mainframe and an x64-based world.

• looks so sweet.
  what a nice cooling there 👌

• Bro, Your videos amazing when I look at my ZX81 and try to compute how many ZX81s I would need match that amount of cache it hurts my brain.

• This excellent, very informative, and well presented video has earned you a rare subscription with "All Notifications" turned on. I look forward to similarly good videos in future.

• 40TB of RAM, 192 PCIe slots, monster CPUs with 256MB of shared cache and cores that can automatically hot-spare… And IBM still uses count-key disk.

• Those inter-drawer bandwidths are unbelievable. As in, I don't believe it. They are higher then HBM bandwidths. IBM may treat these as "cost is not object" devices, but the laws of physics still apply. I think some numbers got mixed up.

• …silicon tech? come now…is this 1950?

• Any modern server has redundant and hot swap everything, also in almost any HA setup things are virtualized and if any one piece of hardware fails the VM seamlessly rolls to another piece of hardware. Anything can be fitted with fiber interfaces. They are not unique to mainframes. Most of the advantages you were touting are not unique. I often joke that most server class hardware will become obsolete to even the hobbyist before it dies.

• If IBM is truly powerful it wouldn't take the weather channel webpage longer to load than my first morning, post coffee and bran muffin poo. Just saying. The performance is abysmal yet they proudly declare 'an IBM company'. If judging by real world results I'm not impressed.

• When 'eventually consistent' just isn't an option, in other words? Commodity hardware and increasingly sophisticated distributed transactions, parallelism, etc. is capable of addressing some traditional mainframe workloads, but instantaneous distributed commit at seven nines of availability? Some problems still have to be solved absolutely from the iron up, and you have made me realize how happy I am that IBM is still on the job!

• I work in payment processing technical support, can confirm, System z and System i are very much alive, as is COBOL and RPG.

• I was doing real-time 24/7 transaction processing on a network of PCs starting in 1993. These were retail prescription claims adjudication for more than 40,000 US pharmacies. Response time was ~1 second in 1993, reduced to less than 1/4 second by the 2000s. That was for each individual transaction, with the ability to handle multiple transactions in parallel on an array of machines. Similar to, but somewhat more complex, than processing ATM/Credit/Debit card transactions. Yes, it included some fraud and abuse detection and avoidance in that.

  Most portions were fault-tolerant via redundancies, but there were a few single points of failure that did result in 3 downtimes exceeding 2 hours. One of those 3 was planned outage as we moved the data center, the other two were due to equipment or data comm failures (i.e. a contractor elsewhere cut a large cable that happened to include our data circuit), Still, better than 99.96% uptime over 16 years, including those 3 outages.

  Definitely wasn’t processing tens of millions of transactions daily, mainframe is probably the only way viably scale to the volumes of financial transactions large banks process, but I know it’s possible to deliver the reliability and performance needed for many applications on much less costly hardware. Our communication link was originally 9600bps, later 57600 bps, so we had a limit of about 28 transactions per second at the end, and we stress tested at ~5,000 transactions per hour in 1993 when we were using slower PCs and had only a 9600bps circuit. I estimate we could have easily handled 50,000 per hour in the 2000s with the equipment we had, and could have scaled to at least 250k/hr with more equipment and a faster data circuit. Since there are the equivalent of about 10x your peak hour volume daily, that’s about 2.5M per day possible.

  Yes, backup and most maintenance was live, no downtime needed for those. There were a handful of exceptions over than 16 years, but I don’t remember if it was 4 or perhaps 10 times, and those we scheduled for late night/early morning when transaction volumes were tiny, and most involved no more than 30 minutes of downtime.

  The point is much of what Dave discussed is possible on PCs, but not all of it, and not at the same capacities. No hot-swapping CPUs, RAM, NICS, or similar. ECC RAM, but not the type of stuff this mainframe has. Was using RAID 10, redundant Ethernet connections to redundant switches, redundant power/power supplies, etc. Of course, the cost of this equipment was around $50k, nowhere near the cost of even a low end mainframe. Was also using “spinning rust”, not SSDs. SSD’s were still really expensive for the amount of storage we needed.

• Always neat to see what going on in the IBM world versus the world that I exist within – PCs and the IoT device of my employer.

• Is it just me or do some of these interconnects look messy. Figured cables would be neatly ran as everything should be held to a high standard.

• My first job was assembler language programming on an IBM 360 mainframe.

• HP did RAIM too; I once built a used rack for a client and one of the machines was a 7RU chassis with 5 warm-removable cartridges that you could pull out and swap while the box was still running. Neat.

• I love the intro shots 😍

• Don't we call mainframes 'the cloud' now?

Comments are closed.