Comments by "" (@ronjon7942) on "The End of Dennard Scaling" video.

  1. 13
  2. 2
  3. 2
  4. 1
  5. When I was just a pup, I was a hardware tech for IBM on the RS/6000, although it may have been rebranded as Power (after the Power processor family), or maybe it was pSeries…whatever. Anyway, it was around the year 2000, and I was sent to work on some new Power3 SMP servers - think it was the p660 H- and M-Series, which were up to 4-way and 6-way SMP. Back then, a 6-way SMP machine was a big deal. I was really struttin’ my stuff with my colleagues when I got trained on the p680-S85, which was a giant of a standalone server. When it rolled out, it was the fastest standalone SMP server IBM had. Maxed out, it had four 6-way ‘books’ for up to 24 600MHz processors, with 16MB of L2 cache. Dizzying. For memory, it had up to 96GB of ECC SDRAM on either four or eight large cards (about the size of an oven cookie sheet), where the memory was soldered on. As far as I know, every server before and after used SIMMs or DIMMs, but those memory cards were a feature for reliability. So, around 2001, the biggest, baddest, fastest standalone server in the world at that time was a 24-way, 600MHz, 96GB RAM, SMP server. It was contained within two, refrigerator-sized racks - one rack for the processors and memory, the other rack with four IO drawers with a maximum of 56 pci adapters. The rack could also hold SCSI and/or SSA drawers. If it was using SSA adapters, it could conceivably connect to hundreds and hundreds of SSA disk within dozens and dozens of tall racks of disk drawers. Amazing system. If it was clustered using HACMP, it could be one of 32 nodes, sharing those disk with every other node. It was also unique as it could also act as a node in an SP2, Scalable Processor/Scalable PowerParallel cluster. The SP clusters were huge, parallel processing, shared disk, shared memory (using fast HPS switches) supercomputers with up to 512 nodes. The actual cluster nodes were in their own special racks, the latest nodes being 4-way, 450MHz, 16GB RAM SMP servers, so having a number of p680s as SP nodes could have been a big jump up for specific applications and jobs. The p680s would connect to the SP cluster’s HPS switch using special PCI adapters. Some universities had 512-node SP supercomputers; I heard from other techs the NSA had a LOT of them. ASCI White was a famous, 512-node SP2 supercomputer that was ranked number one in the Top500, from Nov2000 to Nov2001. ASCI Blue Pacific was another supercomputer at Lawrence Livermore; Seaborg at Lawrence Berkley. Deep Blue was another SP supercomputer, an older, microchannel-node, SP1 cluster that was good at chess. All of these SP supercomputers ran AIX as their OS. I really miss AIX, it’ll always be the best OS, even today. Wow! Sorry, this REALLY brought back memories. So, it’s around 2000, and I’m being trained in Austin. Austin was where, at that time, the Power hardware was designed and manufactured. AIX may have been as well, but I’m not sure. Anyway, we’re in the lunchroom, and it’s just buzzing with engineering excitement. Then a banner is put up: “Power4 breaks 1GHz!” Everyone goes crazy - almost SpaceX crazy. The next year, the Power4, 32-way, 1.3GHz, 256GB RAM, p690 came out, and changed things into what we see today. It contained the first dual-core (today’s terminology) 4-way MCM (multi-chip module) processor complex. Four MCMs were (carefully) installed to make a 32-way SMP. A big deal. The thing was in a rack well over 6 feet tall. Same size as a mainframe, and internally, it could hold 96 disk drives, and 160 PCI-X adapters, connected to a near infinite amount of storage, as fibre channel had gone mainstream about the same time. This is also when LPARs (logical partitions) (VMs/Virtual Machines, to use todays VMware terminology, went mainstream for the first time outside a mainframe environment. In fact the people who started VMware were LPAR engineers who left IBM. With this first iteration, the individual lpars could use any processor, and any memory, and any individual adapter - nothing was virtualized then. You could have a 32-way SMP server down to 32 single processor servers. Today, memory, cpu, and IO are all virtualized as you’d expect, and the core count has gotten significantly higher; also, as you’d expect. I left IBM around this time, to become an AIX sysadmin. I was lucky enough to admin p660s, a p680 (!!!), and engineer the layout of the p690 (so fun), and also the p595 - a mainframe-sized Power6 machine, w up to 1TB RAM. Before that, I got big into IBMs processor/memory and VIO/PowerVM virtualization, and thoroughly into fibre channel director management. Replacing the beloved p690 with the p595 was pretty cool. Oh yeah, plus we used HACMP/PowerHA, clustering one p690 with another, and then one p595 with another, with cross-site, synchronized, fibre channel storage, in two locations about 8-10 miles apart. Man, that was a fun job. What a world. Time to boot up my 6-core, i5, 32GB PC.
    1