At that point, I was happy with the restoration process, except for one thing. The DX2-66Mhz/VLB combo was only able to run DOOM.EXE at 22fps. With FDOOM.EXE I got it to run at 29 fps. Software wizardry was not going to max out DOOM by itself. It was time to boost the hardware.

My 2168-594 came with 128 KiB L2 cache (a very generous amount by '93 standards). There were four empty SRAM slots which, according to the amazing documentation allowed to push the L2 to 256 KiB.

I had no idea how much, if any, performance would improve. But it was such a simple process (or so I thought) that I wanted to attempt this improvement first.

Upgrading required buying not four but five chips. Whether the machine runs with 128 Kib or 256 KiB of L2 cache there must be a TAG chip to store the cache line index. With 128 KiB of L2 SRAM, you need 8 KiB of TAG. With 256 KiB of L2 SRAM, you need 32KiB of TAG.

Accordingly the existing 8 KiB MT5C6408-20 TAG was replaced with a 32 KiB CY7C199-20 TAG.

Manipulating chips was an opportunity to learn how to read the labels. What does "9325 T USA MT5C6408-20" mean?

93   = Year (1993)    
25   = Week of the year (June)
MT   = Manufacturer (Micron Technology)
5C64 = Model 
-20  = 20ns response time

The four purchased SRAM chips where similar to the one already installed. Namely 9133 A USA MT5C2568-25.

In the photo above, notice the DX2-66Mhz in the lower right. Extra pins are visible on the side to allow a Pentium Overdrive (in theory).

When I turned the computer on, I had zero expectation it would work on the first try. And that is exactly what happened. At this point in the restore process, I had gotten used (resigned?) to it. To be fair that was quite consistent with my recollection of the overall PC experience back in 1993.

I posted on vogons^[1], pointing out the documentation, the required SRAM chip specs, the jumpers J16, J17 settings, and asking if anybody had run into this issue.

At first people suggested I got sanded fake chips. Others suggested testing the chips with a reader. I ordered more SRAM and a chip reader.

Eventually, user majesty joined and explained that the problem was far more simpler than others had suggested.

Your TAG chip isn´t inserted correctly, several pins are hanging outside the socket!

- majesty

Ouch. That was embarrassing. And also representative of how error prone the PC field was back in the days.

After making sure the TAG chip was properly inserted, with all the pins in, the machine booted to display 256 KiB in the BIOS.

I took the opportunity to run some benchmarks and measure the impact of cache size over DOOM framerate. Because accessing the L2 chips required removing the riser card, I ran all of the benchmarks without sound and screen size = 10 (fullscreen with status bar).

The impact of cache is tremendous, even a tiny amount of L1 increases framerate by 10x%. The return on investment however decreased with the amount of L2. The first extra 128 KiB brought 10% performance increase. Adding another 128 KiB to a total of 256 KiB improved performances by roughly 5%

Since I had bought quite a stock of chips, I found myself with many 20ns rated ones. These were supposed to be used only for the TAG chip. But I wondered if I would see an improvement if I moved all the L2 to 20ns instead of the recommended 25ns. It made absolutely no difference.

Upgrading the CPU

^

[1]

IBM PS/1 2168 SRAM L2 cache failed upgrade