二十年Architecture之长进

Fritz is a German chess program originally developed for Chessbase by Frans Morsch based on his Quest program, ported to DOS, and then Windows by Mathias Feist. With version 13, Morsch retired, and his engine was first replaced by Gyula Horvath's Pandix, and then with Fritz 15, Vasik Rajlich's Rybka.

Fritz 9 Schachbenchmarks (jens-hartmann.at) 节选：

CPU | nxMHz | Relative speed
Celeron Northwood | 2000 | 0.91
Pentium III | 1000 | 1.00
Athlon XP 2600+ | 2083 | 2.06
I5-4690k | 4x3900 | 24.40
i9 9900k | 16x4700 | 70.04
Ryzen 7 3800X | 16x4400 | 69.00

i9-9900K / Q4'18 / 14 nm /8C16T / 3.60~5.00 GHz / 16 MB Smart Cache
Pentium III 1G / Q1'00 / 180 nm / 1C1T / 1.00 GHz / 256 KB L2 Cache / TDP 29W
Celeron Northwood / 2.60 GHz / 130 nm /128 KB L2 Cache / TDP 62.6 W!

Intel Smart Cache refers to the architecture that allows all cores to dynamically share access to the last level cache.  Sarcasm?

In performance, Coppermine arguably marked a bigger step than Katmai by introducing an on-chip L2 cache, which Intel names Advanced Transfer Cache (ATC). The ATC operates at the core clock rate and has a capacity of 256 KB, twice that of the on-chip cache formerly on Mendocino Celerons. It is eight-way set-associative and is accessed via a Double Quad Word Wide 256-bit bus, four times as wide as Katmai's. Further, latency was dropped to a quarter compared to Katmai. Another marketing term by Intel was Advanced System Buffering, which encompassed improvements to better take advantage of a 133 MHz system bus. These include 6 fill buffers (vs. 4 on Katmai), 8 bus queue entries (vs. 4 on Katmai) and 4 write-back buffers (vs. 1 on Katmai).

Under competitive pressure from the AMD Athlon, Intel reworked the internals, finally removing some well-known pipeline stalls. As a result, applications affected by the stalls ran faster on Coppermine by up to 30%. The Coppermine contained 29 million transistors and was fabricated in a 0.18 μm process.

Coppermine (0.18 μm), L1-Cache: 16 + 16 KB (data + instructions), L2-Cache: 256 KB, full speed, MMX, SSE

Intel Pentium III (Coppermine) @ 933MHz, BogoMIPS=1862.05 [link]

所以我放个厥词：20年前，Pentium III做到了一个时钟周期两条指令，然后大家都就是靠主频和核数了。别说什么4发射，cat /proc/cpuinfo不是可以看BogoMIPS么？不要讲14T算力，讲识别率。不要说买了多贵的硬件和用了什么方法，说做出了什么成果。:-)

补充：How BogoMips is calculated in linux kernel?