Might be worth trying "Load Performance Defaults" and setting the SDRAM CL to 2 in case it's not picking it up from the PC133 @ 100
And examine the current RAM timings with https://www.cpuid.com/news/66-cpu-z-vintage-edition.html

Check if the BIOS has an option for "in order queue depth" and set it to its highest value

Of course I have forced the memory CAS2 in the BIOS setup.

Unfortunately no option "in-order queue depth" for this motherboard. Some software reports the value of 4 is used (should be 8?).

I believe 4 is the correct value for a BX

Here is the analyse of my problem from ChatGPT, I don't know if it's true but could be an explanation :

---

🧩 Summary: Low “Float iSSE” Memory Bandwidth on Pentium III Coppermine + Intel 440BX

I was testing an old PC with a Pentium III 800 MHz (Coppermine), GA-6BXC motherboard (Intel 440BX chipset), and 256 MB SDRAM CAS2.
In SiSoft Sandra 2002 and 2004, the memory benchmark shows normal “Integer” bandwidth, but the “Float iSSE” score is about 40 % lower than expected.

Older Sandra 2001 (which doesn’t use SSE in its memory test) reports normal and balanced Integer/Float results.

After checking everything — BIOS cache settings, RAM timings, chipset drivers — the cause turned out to be the 440BX chipset itself, not the CPU or RAM.

---

🔍 What’s Really Happening

Starting with Sandra 2002, the memory benchmark uses SSE (iSSE) routines that move data in 128-bit blocks (MOVAPS/MOVNTPS).

The 440BX memory controller, designed before SSE existed, only optimizes 64-bit burst transfers and doesn’t handle 128-bit SSE memory access patterns efficiently.

As a result, SSE memory throughput (the “Float iSSE” test) is about 30–40 % slower than expected.

The Integer test is fine, because it still uses regular 64-bit memory operations which the 440BX handles perfectly.

---

🧠 Why Real-World Performance Isn’t Affected

Games and real applications don’t stream large blocks of floats directly from main memory like Sandra’s synthetic test.

Most data fits in the L1/L2 cache, and game engines use SSE arithmetic on cached data, not on big raw memory transfers.

So this bandwidth drop is only visible in synthetic benchmarks, not in actual 3D or physics performance.

---

⚙️ Katmai vs. Coppermine

With a Pentium III Katmai, Sandra doesn’t use full 128-bit SSE memory moves (the Katmai only supports a limited SSE subset).
→ Memory results are normal.

With a Coppermine, full SSE 128-bit instructions are used.
→ The 440BX can’t burst them efficiently, so the “Float iSSE” score drops.

---

✅ Conclusion

It’s not a defect — just a chipset limitation.

The 440BX predates full SSE support and can’t optimize 128-bit memory transfers.

Everything works as it should; it only affects Sandra’s “Float iSSE” synthetic score.

Newer chipsets like the i815E or VIA Apollo Pro 133A don’t show this issue.

This is probably a problem of a process using the CPU while the benchmark is running. Sisoft sandra memory test can be inconsistent in Windows 98. It helps to re-run the same test multiple times, even reboot the PC. You should be able to get good results at least once.

I have noticed that the performance can be lower when another thread is running in the background, same if you run the benchmark multiple times.
With Sandra 2002 or 2004, the Float memory benchmark using SSE is limited to about 425Mb/s, no matter what I do.
With Sandra 2001 the benchmark is not using SSE, so the results are normal when comparing with other systems.

Maybe someone with a Coppermine CPU on a 440BX motherboard could confirm this behavior?

I have Coppermine on 440BX and do not see that behavior. Try reinstalling OS and install as few drivers as possible. You can also try running it in safe mode.

Even in safe mode the benchmark results are the same.

Maybe this is a bug of the GA-6BXC motherboard ?

It could be BIOS or microcode issue. Maybe you can try updating BIOS. It's a 3 ISA slot board, those tend to be older. You could also try looking at wpcredit with 440bx chipset file. Perhaps there is a setting that can be abled to fix it.

Before updating BIOS try Everest 5.5 or 5.3. 5.5 is from 2010 but still works on Windows 98. It is better than Sisoft Sandra which became bloatware.

It is very difficult to get consistent results from Sisoft Sandra memory benchmark. Everest 5.5 memory test gives me 786MB/s (read), 787MB/s (write), 114ns latency. It would take like 30 attempts in Sisoft Sandra to get results that make sense.

The BIOS was released in 2002 and is ready for Tualatin and C3 CPUs, so there shouldn't be a microcode issue with Coppermine.

Try Everest 5.5, see my results above.

My Everest results are close to yours, we are close to the maximum bandwidth of FSB @100MHz (800MB/s)

So the problem is in Sisoft Sandra, not your board.