From mboxrd@z Thu Jan 1 00:00:00 1970 From: kevin.bowling@kev009.com (Kevin Bowling) Date: Fri, 29 Dec 2017 04:04:01 -0700 Subject: [TUHS] Why did PDPs become so popular? In-Reply-To: <20171228232852.GC30269@thunk.org> References: <20171228140551.B6F9418C079@mercury.lcs.mit.edu> <20171228160811.GA13474@mcvoy.com> <20171228232852.GC30269@thunk.org> Message-ID: I can mostly only comment from my readings, but this is IMHO a non-trivial question even in retrospect because it depends on what you or your workload valued. I do like how lmbench categorizes across multiple "interesting" vectors, showing some of the design tradeoffs of CPU architecture and then how the OS plays a role in real world applications. Alpha generally maintained integer/ALU and clockspeed leadership for most of the '90s http://www.cs.columbia.edu/~sedwards/classes/2012/3827-spring/advanced-arch-2011.pdf For intel the Pentium Pro was pivotal in being able to position vs the contemporary RISC CPUs with multi-MB caches. And federating cost effective CPUs together with Laissez-faire mainboard/system packaging/OSes turned out to be more important than pretty much anything else. I know that IBM Austin did some wizard of oz stuff in early POWER.. Bi-CMOS superscalar that put it 4-5 years ahead of most competing microprocessor designs in floating point, though SGI and HP traded respectable volleys. But POWER also had fat memory bandwidth, they tended to make well balanced braniacs that didn't blindly pursue the clockspeed war (http://www.cs.virginia.edu/~mccalpin/papers/balance/). They got into a weird spot in the second half of the 90s with an out-dated bus, internal competition from PowerPC and RS64, but returned to leadership with POWER4 and up to today. POWER's IO leadership made it a good fit for many early internet routers in the early '90s, NSFNET and others. Interestingly modern POWER retains that leadership on memory and I/O which can put it 2 to 4 years ahead of everything else depending on what you value (Google seems to see the light using it for gluing GPUs together..). From IBM Rochester, RS64 was kind of interesting in being the first use of SMT and having incredibly low TDP which were both interesting clairvoyance into the next decade. In retrospect I don't think SPARC ever had a reasonable showing pound for pound vs other microprocessors :) Regards, On Thu, Dec 28, 2017 at 4:28 PM, Theodore Ts'o wrote: > On Thu, Dec 28, 2017 at 08:08:11AM -0800, Larry McVoy wrote: >> > I think you're right. The disinterest in marketing and advertising >> > (Ken Olsen, and therefore DEC, had a "build it and they will come" >> > mentality) was one aspect of the corporate culture. An example of its >> > negative impact: When the Alpha EV5 came out, it was several times >> > faster than anything else around. >> >> Got a reference for that performance claim? Wasn't that mid 1990's? >> If so, I was heavily into benchmarking and performance work during >> that period. If there was a processor that was 2x faster, let alone >> several times faster, I would have noticed. >> >> http://www.bitmover.com/lmbench/lmbench-usenix.pdf > > "Digital's 21164 Reaches 500 MHz: Alpha Regains Performance Lead, > Leaves Pentium Pro in Dust" -- Microprocessor Report July 8, 1996. > > http://noel.feld.cvut.cz/vyu/scs/alpha/164_500.pdf > > Looks to me from the article that Alpha was certain participating in > the clock speed wars, though. > > - Ted