Comment by kragen

It's worth reading John Mashey's post on why the VAX was displaced by RISC: https://yarchive.net/comp/vax.html

Basically, making VAXen go faster was too much work, so they fell farther and farther behind, and they were already behind MIPS on performance in 01988. Mashey was at MIPS at the time and tells his recollections of how DEC started buying MIPS chips to build DECstations around.

The 11/750 came out in October 01980; one of the last VAXstation 3100s came out in November 01989: https://gunkies.org/wiki/VAXstation_3100_Model_38/48 (not sure about the M76 specifically). The Dhrystone performance results from https://netlib.org/performance/html/dhrystone.data.col0.html puts the 11/750 at 0.50–0.62 MIPS (very close to your 0.65) and the last VAX listed, a VAX 8650, release date December 01985, at 5.0–6.3 MIPS.

A factor of 10 performance gain over 62 months was pretty amazing; that's 56% performance improvement per year, a doubling of performance every 19 months, slightly faster than Dennard scaling. But you'd expect it to be faster than Dennard scaling because the VAX 8000 was ECL, not CMOS; it ran at 18MHz in 01985!

And that's pretty much when the rotting VAX whale splattered at the bottom of the cliff. If you're right that in 01989 they were still stuck at 7.6, four more years later, they'd only gained an additional 35% in performance over those four years, about 8% per year. If we average over the 9 years in your comment, that's 14× performance increase, 34% per year. In isolation that would be staggering—it's faster than the precipitous drop in solar power prices over the last 20 years—but it's far slower than the performance improvement CMOS got from Dennard scaling.

Other architectures (RISC, and the less demanding CISC of the 80386) were speeding up much faster. The 20MHz 80386 released in 01985 had already very nearly caught up with the VAX 8650 in computrons (though not I/O), despite being mere CMOS instead of ECL. The Netlib page documents that an AlphaServer 8400 5/300 was getting 464 MIPS on Dhrystone 2.1 in, say, 01996 https://top500.org/system/166696/. That's 800× the speed of the 11/750 16 years earlier, a 52% speedup per year.

(This is a little bit personal for me because my first internet access was on a VAX, and I think a friend of mine committed suicide due to the failure of the VAX line.)

Speeding up has slowed down since the collapse of Dennard scaling in 02006. Roy Longbottom reports http://www.roylongbottom.org.uk/Raspberry%20Pi%20Benchmarks.... that his Raspberry Pi 3 gets 2469 VAX MIPS and his 3.9GHz "Core i7" gets 16356 VAX MIPS. The i7 is of an unknown vintage (the brand name covers chips from 02009 to 02017) but I'm going to guess it's closer to the end, about 02016, 20 years after the AlphaServer and only 35× faster, only a 19% speedup per year. But this time it's affecting RISC architectures too, and it's GPUs that have left conventional single-thread-focused CPUs in the dust.