old x86 memory performance flaw+benchmarks \ VOGONS

old x86 memory performance flaw+benchmarks

Topic actions

Post a reply

First post, by auron

Posted on 2019-12-22, 18:51

auron Offline

Rank Oldbie

Rank: Oldbie
Posts: 641
Joined: 2015-08-20, 01:56

i have randomly stumbled across this and thought that some might find it interesting:
- 1998 press release
- benchmarks of contemporary processors

there's more technical info and user benchmarks there too if you dig around the site. what's interesting to me is the pentium mmx vs k6 memory benchmarks. the k6 does not benefit from their fixes anywhere near as much as the pentium mmx does, which shows 50%-100% gains in some cases. the k6 also has noticeably lower video memory write/transfer rates by default and does not have MTRRs available to boost that, i wonder if this is one of the main reasons why it was not all that much faster compared to the pentium mmx in games, despite having a significantly more advanced core on paper? were the fixes mentioned on this page eventually implemented in programs?

Reply 1 of 2, by bakemono

Posted on 2019-12-23, 12:08

bakemono Offline

Rank Oldbie

Rank: Oldbie
Posts: 693
Joined: 2018-01-15, 06:56

The main link sounded a bit like snake oil, but after checking out the technical page I decided to experiment on my 83MHz Pentium Overdrive.

Reading the first dword of the next line before reading the remaining dwords of the current line does result in higher throughput compared to REP LODSD. 74,000,000 bytes/sec instead of 49,000,000 bytes/sec.

Not sure how to extend this to faster copies though, and FPU writes don't gain anything on this system (32-bit bus with VLB...)

I guess Intel finally fixed this on the P4 or tualatin, as REP LODSD suddenly became much faster even with the same old SDRAM. 6x86 and K6 already executed each iteration of REP LODSD in only one cycle and spent the rest of the time waiting on the bus anyway.

Not sure about writes on a K6., were they worse than K6-2?

edit:
BTW, my benchmark was reading from the L2

also, I did find that using sparse reads to blast some stuff into the L1 before writing it out using a normal REP MOVSD was slightly faster: 33,000,000 bytes/sec over 30,000,000 bytes/sec (in the L2 again, writeback)