On Thu, 21 Aug 2025 17:25:14 -0400, Phil Smith III <li...@akphs.com> wrote:
> the mere fact that they did Telum and Spyre doesn't convince me that there's > a real business case there. Understand that AI does not function without a CPU (e.g. Telum, x86 or ...). AI engines can be TPU's, APU's, TPU's, NPU's and more. The one thing they are not is CPU's. They specialize in math and lack the capabilities found in a CPU (e.g. TCP, moving data, formatting data, interfacing with users and all the other tasks we expect from a CPU. Today, every company and many home computers do AI. As such, they need CPU's. Often, they want to buy an integrated product. The market exists but can IBM sell their concept. IBM makes innovative products and companies should be flocking to their products. Google has 2.5M servers with AI. They could replace them with 250,000 z17. IBM can't sell water to a man dying of thirst, otherwise Google would have a boatload of z17 but they don't have a single one. We must stop talking about the virtues of the mainframe. Instead, we need to educate the computer industry how they are getting everything wrong. Vector instructions are just one screwup. Until we debunk the lies from the computer industry, the mainframe doesn't stand a chance despite being the most innovative computer to date. > Telum and Spyre doesn't convince me Technically, this is the most innovative design. Every other implementation pairs the CPU and GPU. In 1960's, IBM went NAS (Network Accessible Storage). Today, there is not 1 device connected to a z17. That Spyre card is accessible to every CPU drawer. Maybe someday, they'll move Spyre outside the z17 to make it available to multiple z17 at the same time. Technically, IBM is the only company to truly understand distributed computing. IBM z and El Capitan supercomputers are the only computers that do not have disks and are designed to use NAS disks. When a person says they know NAS, then ask them if they can get 100,000 transactions per second using a single computer. El Capitan proves the hardware is capable but Unix is not. Marketing wise, the computer industry is not capable of understanding the mainframe. For this to be successful, we must talk about this in their terms instead of the virtues of the mainframe. >Who is using expensive IBM Z MIPS for AI? Do you consider El Capitan cheap MIPS at $600M? At 12TB per second transfer rate between blade servers, it's about moving data to and from the APU's. It's about PCIe 5.0 with up to 256 lanes. each Each blade server has 8 CPU at 1.8Ghz (slow). Each CPU has 24 cores = 192 cores. Each CPU has 128GB dedicated ram (not shared between CPU's). Even if it were shared, it would be 1TB. 32KB L1 cache per core, 1MB L2 cache per core, 32MB L3 cache shared between 8 cores and the APU, per 4 cores and 256MB L4 cache shared between 24 cores and the APU. IBM Z at 5Ghz has 192 PCIe 4.0 at 16 lanes. 40TB storage. 32MB L2 per core. 512KB L1 per core. Everyone should be using IBM Z if they want an entry level supercomputer that can "potentially" perform as well El Capitan after adjusting the price performance. I can't say who is using it. I doubt IBM Z is expensive considering it replaces 10 to 20 servers, personnel, power and other factors. Sadly, Unix is not designed for IBM Z despite it running on Z hardware. Cost will be significantly more. z/OS on the other hand can achieve 100,000 IMS transactions per second on the same hardware. IBM Redhat Linux is closed source and IBM should make zLinux performant. I personally believe IBM has the talent and skills to build the greatest computers. The problem is they don't have the ability to sell great innovation.
