A Tour Inside the IBM Z17

I/O Architecture and Rack Layout

• Diagram fascinates people: in a 4‑rack z17, most space is I/O drawers, not CPUs; some is intentionally left empty for floor‑loading and power‑density compatibility with previous generations.
• I/O drawers are large (8U) PCIe Gen5 infrastructure: up to 12 drawers, 192 PCIe slots, multiple channel subsystems (up to 6×256 channels), heavy use of PCIe fan‑out and switches.
• Each I/O device is on a “channel” (effectively a separate controller computer), with lots of redundancy and hot‑swap support; this design is key to throughput and reliability.

Mainframes vs POWER and Open Hardware

• Clarification that IBM z and POWER are distinct architectures, though they share ideas.
• Discussion of Raptor’s POWER9 workstations (Talos II, Blackbird) as expensive but open(ish) alternatives to x86, motivated by ISA diversity and firmware transparency.
• Contrast between openness levels: POWER9 has open ISA and on‑chip firmware; OpenSPARC Niagara 2 goes much further with full RTL; neither is fully “free” silicon in practice.

Pricing, Licensing, and Procurement

• No public list prices; everything is negotiated and often NDA‑bound. Estimates range from ~$100k for older/entry systems to “over a million” for large configs; modern z‑class often leased.
• Pricing is dominated by software and capacity licensing (MIPS / rolling averages, “sub‑capacity” models). z/OS and COBOL capacity is most expensive; Linux on z and Java somewhat cheaper.
• Note that an IBM Linux‑only mainframe (e.g., Rockhopper) has mid–six‑figure starting prices but won’t run z/OS.

Who Uses Mainframes and For What

• Widely used by large, long‑lived institutions: big banks, payment networks, insurers, governments, social security systems, healthcare, and other Fortune‑500‑scale orgs.
• Core workloads: high‑volume OLTP plus batch—payments, ledgers, entitlement calculations, fraud detection—often with Java and COBOL mixed.
• Many systems still effectively compatible with IBM 360‑era software; some 1980s assembly still runs unchanged.

Reliability, Performance, and New Features

• Emphasis on extreme reliability (claims of 8 nines), hot‑swappable components, spare processor units, RAID‑like memory, and precise I/O semantics (no “lying” about writes).
• Architecture prioritizes huge caches and very fat cores for single‑thread/low‑latency performance over core count density; good for workloads that can’t be easily sharded.
• Crypto is heavily accelerated in hardware (CPACF), including “post‑quantum” algorithms; AI units are aimed at ultra‑low‑latency inference during transactions, not training.

Legacy, Migration, and Growth

• Mainframe usage is framed as “legacy but growing”: overall compute on z increases as more frontends and analytics are bolted onto old cores.
• Migration off mainframes is described as risky, expensive, and often performance‑regressive; an example social‑security rewrite reportedly failed by being orders of magnitude slower than the original.
• Some argue all workloads could migrate but that keeping mainframes is cheaper and less risky; others highlight that modern replacements often undervalue performance and correctness versus developer cost.

Cloud Comparisons and Alternatives

• One view: cloud is “mainframes gone full circle”—centralized, consumption‑based, specialized hardware; difference is that in the cloud you must build reliability in software across unreliable nodes.
• For many orgs, a fault‑tolerant distributed x86 system is preferred due to vendor plurality and less IBM lock‑in, despite the engineering effort.
• Skepticism about cost‑effectiveness: commodity 1U servers can offer more cores, RAM density, and network bandwidth, though defenders note that raw counts ignore latency and reliability needs.

Security and Other Mainframe‑like Systems

• Side discussion of Unisys ClearPath MCP and BAE’s XTS/STOP as security‑focused “mainframe‑style” systems; debate over whether their security claims meaningfully exceed well‑hardened Linux.
• Some see these systems’ security story as partly marketing and note that MCP now runs as an emulator atop Linux, changing the threat model.