Intel TX chip set and more than 64MB

[SimMike2]

My brother is using an Abit AX5 with 64MB of SDRAM. What would happen if I add another 64MB and get 128MB total? I have heard many stories of performance hits and other stuff because the Intel TX chip set doesn't cache memory when you have above 64MB?

What is the the real world story? Will he see any performance increase?

 

[AndyHui]

If you are running programs that actually use up more than 64MB of RAM, you may see a performance improvement with more than 64MB of RAM.

Normally, performance will be degraded above 64MB of RAM; the chipset caches the first 64MB, but because of Windows' top down memory structure, all of the critical items will be loaded at the top of memory first, outside the cached area.

If all of the programs that you run total less than 64MB of RAM, then you will see a performance slowdown. However, even if you did have 128MB of RAM, and actually utilised more than 64MB of RAM, it would be faster than having only 64MB of RAM. Using uncached RAM will ALWAYS be faster than having to hit the swap file on the hard drive.

 

[SimMike2]

So it sounds like the extra ram above 64MB is exactly like the swap file, only it is in memory instead of on the hard drive? If this is true, it seems like most people will see a performance increase, no matter what they are doing on the computer.

It seems to me that no matter what you are doing on the computer, Windows 98 will use and like having more than 64MB.

 

[Peter]

One thing to add, the "uncached" memory range isn't really uncached, it's just L2-uncached. The level(s) of cache that reside within the CPU will still cache it, only the mainboard's won't.

Regards, Peter

 

[mechBgon]

I had an Epox 430TX motherboard with 64Mb of EDO. I was using some 3D modelling software and needed more memory. I benchmarked a render job with my old memory, then replaced it with 128Mb of SDRAM. To my great disappointment, render times increased 20%. At that point I decided to go to a 440BX/PC100 setup.

The added RAM will help keep Windows from swapping to disk, which is a plus, but can carry a performance hit at the same time.

 

[SimMike2]

So the gist of this is that having more than 64MB of ram in an Intel TX chip set mainboard, basically disables, or, renders useless, the 512K or 256K cache on the mainboard. If this is what happens, this can be a serious hit and performance. My brothers computer has a K6-200, which has a relatively large 64K L1 cache, so it might not be too bad.

 

[AndyHui]

No, it does not render the L2 cache useless. It's just that due to Windows' top down memory structure, the important/frequently accessed stuff won't be cacheable. The L2 cache is still used though.

 

[SimMike2]

So it isn't quite as bad as totally disabling your L2 cache?

 

[Serville]

This has been my previous debate with someone in another forum. He kept saying that >64MB in TX chipset slows system down. My experience with TX97-E with 128MB however does not say so.
I have done very intensive test between 64MB & 128MB in this board using any benchmark I can think of, Quake2, Quake3, UT, Unreal, SySoft Memory/CPU/Multimedia, HDTach, WinBench99, Wintune, WinStone99, and .....god what else.
In every result, I do not see any slowdown in any of them. That is why I keep wondering whether the issue of 64-MB cacheable area is just theoritical or a proven fact.
At least with my system TX97-E, I can not see any degradation in performance due to 128MB.
In fact, the sytem runs generally faster using 128MB, especially during Windows startup, games loading, and even Quake2 & Quake3 score, although it is so subtle , possibly only 1 fps.

I'm interested in this thread. Would like to hear other comments about this 64MB-cacheable limit.

 

[AndyHui]

The 64MB cacheable limit is a proven fact. All Intel 430 PCISet series chipsets, the FX, HX, VX and TX are subject to this limitation. The 430HX can cache up to 512MB of RAM though the use of a cache TAG RAM, which most motherboard manufacturers later included.

However, the problem of the slowdown really only now applies to users of Windows 3.1x or a very clean Windows 95, where the usage of RAM does NOT exceed 64MB. Due to Windows' top down memory structure and the 64MB cacheable limit, the most critical and frequently accessed system data will be located outside the 64MB cached area, resulting in the slowdown.

These days, when most people use Windows 98 and the usage of RAM exceeds 64MB, it will still be advantageous to have more than 64MB of RAM. This is because although the system data is not cached, it is still located in RAM, which is much faster than having to hit the swap file on the hard disk, which would have been the case had you only had 64MB of RAM.

 

[Peter]

To put matters straight, all of those Intel chipsets use TAG RAMs.

With those, 8-bit-wide TAG limits cacheability to 64 MBytes. 11-bit TAG (on HX only) gives you 512 MByte. HX boards usually had an 8-bit TAG RAM onboard, and a socket for a second 8-bit chip that would then provide the extension to 11-bit.

Very CPU centric stuff with very little RAM traffic, as well as "streaming" type applications with really big amounts of data, might not even notice the L2-uncachedness of certain areas of RAM. That's why some people see a difference and some don't.

In all cases, not having enough RAM for whatever you're trying to do there does have a much worse effect than L2 cacheability limits.

Regards, Peter