Variations in benchmark performance on LC475/Quadra 605 with 68040 CPUs and VRAM

[MrFahrenheit]

Hey everyone!

I’ve encountered very strange behavior with regards to these pizza box Macs.

It started out when I was building my own 60ns VRAM SIMMs and I noticed a performance boost. I was seeing a ~15% performance increase when using 60ns VRAM in a 33mhz overclocked LC475 when compared to 70/80ns VRAM.

I had reached out to Silicon Insider to ask whether he had any 60ns VRAM chips and whether he saw any performance gain from using them and he did some significant testing on a 33mhz LC475. He got the same results no matter if 60 or 80ns VRAM were used, and they were similar results to my “higher” results from using 60ns. Meaning for some reason my OC 475 using 70 and 80ns VRAM was returning SLOWER results.

I’ve thought long and hard how this could be possible.

I have built more than 10 of the 60ns VRAM SIMMs and have tried several to verify. I have tried these in 3 different overclocked 475 machines, with nearly identical results.

All of my 475 that I’m testing on have been recapped with solid tantalum caps. Some have higher voltage caps, and some have the same voltage as the stock ones.

I purchased another 475 so I could test against the stock electrolytic caps and compare. I got the same results as vs when I recapped that same board.

Then I discovered a strange result: when J18 is installed (to identify it as a Quadra 605) I got ~15% performance degradation on the 33mhz overclock. I did not try that before overclocking, sadly.

Nearly all of the variances in speed occur ONLY in 16bit color tests. The 8bit color tests are always within margins of each other.

I just don’t understand what’s going on here.

To add to this mess of strange results, I have more!

Last night I was testing batches of 68040 CPUs I have sourced on eBay. Yes, I’m crazy, but I have obtained 10x full 68040 CPUs marked 40mhz from several sellers. Looking at the silkscreen they appear to be authentic and not relabeled.

I have a variety of masks and revisions, including XC and MC parts, two of which are 57um 2006 manufacture (0K63H, which is just prior to the infamous L88M everyone seeks).

The worst video benchmark score I saw was with the MC68040RC40A with mask 0K63H.

I tested using two different logic boards but consistently keeping things the same. Meaning I tried the same logic board, RAM and VRAM with the CPUs, then another different logic board with another set of RAM and VRAM on that board, with all of the CPUs, and then switching RAM and VRAM between the two boards. Always the same results.

So, it seems that I’m seeing variations in video benchmarks when using different CPUs.

Is there anyone here who can assist me in validating my tests? Someone who has a 475 board that is overclocked to 33mhz, that has a variety of 68040 and 68LC040 chips and a few different speeds of VRAM to independently verify my crazy results? I see little variation when using the stock 25mhz board speed.

If someone is willing to test on at least 2 boards and with at least 2-3 different combinations of CPU and needs a pair of 60ns VRAM I would be willing to ship some out to help nail down what is going on here.

I’d also love to hear the theories and thoughts behind the J18/Quadra 605 slowdown I saw at 33mhz overclock. If you have a 33mhz overclocked 475 and can run a 16bit video benchmark (and 8bit for good luck as well) with and without the jumper I’d love to hear the results.

I also used a variety of PSUs just to ensure for some reason the power supply wasn’t playing a role in the variations. I’m trying to be as scientific as possible.

Your help and support on this is greatly appreciated and I feel it would benefit the community to validate what I’m seeing.

 

[cy384]

when J18 is installed (to identify it as a Quadra 605) I got ~15% performance degradation on the 33mhz overclock.

for this part at least, I may have a hint: in the ROM there is code that decides how long to wait for a VRAM access (some number of cycles), I belive it only picks based on the machine ID

for the same VRAM speed, a faster CPU needs to wait more cycles

if you overclock beyond the standard speed for a machine ID, you will cut into the safety margin of the VRAM access time (but it will still work unless you go really far), since the same number of cycles takes up less actual time

by waiting longer for VRAM access, of course, you lose some performance

 

[dr.diesel]

or this part at least, I may have a hint: in the ROM there is code that decides how long to wait for a VRAM access (some number of cycles), I belive it only picks based on the machine ID

Very interesting! Was the 475/Q605 not released at the same time? Meaning from the factory they would have most likely had the same VRAM chip speeds? I can't think of any other reason the firmware would treat each model different? I could see maybe if in this case the 475 was release later that it possibly could have included faster VRAM.

I have an OCed Q605 that I could test, 33mhz. Can you outline what testing suite/method so we can test apples to apples?

 

[MrFahrenheit]

for this part at least, I may have a hint: in the ROM there is code that decides how long to wait for a VRAM access (some number of cycles), I belive it only picks based on the machine ID

for the same VRAM speed, a faster CPU needs to wait more cycles

if you overclock beyond the standard speed for a machine ID, you will cut into the safety margin of the VRAM access time (but it will still work unless you go really far), since the same number of cycles takes up less actual time

by waiting longer for VRAM access, of course, you lose some performance

That would make sense, except that I'm losing performance on some CPU chips vs others. I get variations in my benchmark results using the same logic board, RAM, VRAM, and 33Mhz overclock, but different 68040 (full FPU) CPUs.

 

[MrFahrenheit]

Very interesting! Was the 475/Q605 not released at the same time? Meaning from the factory they would have most likely had the same VRAM chip speeds? I can't think of any other reason the firmware would treat each model different? I could see maybe if in this case the 475 was release later that it possibly could have included faster VRAM.

I have an OCed Q605 that I could test, 33mhz. Can you outline what testing suite/method so we can test apples to apples?

This would be great if you could validate some of my tests.

I have a bunch of results from Norton Utilities 3.x (don't have the version handy, sorry) System Info. The right one is where System Info compares its results against the Quadra 700. Newer versions are against PowerPC machines. I believe it's Norton 3.2.x

Set monitors to 640x480, turn Appletalk off, set color depth to 8 bit and run the CPU, FPU, and video benchmarks (no need to test SCSI at this point). Then repeat the tests with setting the color depth to 16 bit. Note the speed of your RAM and VRAM please.

Then, if you could, repeat the above tests by inserting the jumper at J18 for 'Quadra 605'. You'll need to shut the machine down first, because it will fail to boot if you jump it on-the-fly (not recommended!).

Once you can do these, please post pictures of your results (the files themselves can be hard to transfer and then download and analyze, so a photo of your screen with the results is sufficient).

I'll try to find another few apps we can use to test them in a standardized way, to try and weed out any discrepancies. I personally have many LC475 boards and 68040 CPUs that I can test against. I'd like results from ones I haven't touched.

Thanks for helping me out on this. I really appreciate it.

 

[MrFahrenheit]

Forgot to mention, I do tests on 7.5.5 but I could use any OS version you use as I have all of them accessible to me.

 

[dr.diesel]

First round of tests:

Globals:
Q605@33Mhz, System 8.1, 36M RAM (60ns), 1M VRAM (80ns), Norton Utilities 3.1, AppleTalk Off.

With J18 installed (Machine ID = 95):
8bit:
CPU = 130
Video = 124
FPU = 131

16bit:
CPU = 130
Video = 69.8
FPU = 131

Without J18 installed (Machine ID = 90)
8bit:
CPU = 130
Video = 125
FPU = 131

16bit:
CPU = 130
Video = 69.9
FPU = 131

 

[dr.diesel]

I'm now questioning the 1M of VRAM, was 99% sure I upgraded this machine? The highest resolution option I have is 832x624@16bit, I seem to recall that should be higher with 1M? I might be remembering incorrectly... The sims are not labeled, will have to Google the chip numbers I guess.

 

[Flux]

For Quadra 605 (Performa 475 and LC475) specifications, try http://quadra605.mcdonnelltech.com.
It sounds like the 1 MB VRAM is being used on your system:

• Standard 512 KB VRAM supports 8-bit color for 256 colors at resolutions up to 832×624 and 16 colors at resolutions up to 1152×870
• Upgrading to 1 MB VRAM provides 16-bit color for thousands of colors at resolutions up to 832×624 and 256 colors at resolutions up to 1152×870

I have a Performa 475, which I've upgraded with a full-fat 33 MHz '040.
I can't easily run benchmarks just now, but I can take a look when I next get it set up.

 

[MrFahrenheit]

Here are my video only results. The remaining results (CPU, FPU) are very similar or identical to yours depending on the run I do.

 

[dr.diesel]

Do you have any machines with an 040 and hardware cpu cache? I wonder how much improvement that makes? I believe MicroMac made a CPU socket cache card for the Q605s and equiv but I've been searching for years without luck.

 

[MrFahrenheit]

Do you have any machines with an 040 and hardware cpu cache? I wonder how much improvement that makes? I believe MicroMac made a CPU socket cache card for the Q605s and equiv but I've been searching for years without luck.

No. I’m just trying to sort out the inconsistencies I’m finding on stock hardware with overclocks.

 

[dr.diesel]

No. I’m just trying to sort out the inconsistencies I’m finding on stock hardware with overclocks.

Any other tests/help I can provide?

 

[MrFahrenheit]

Any other tests/help I can provide?

No. I think that’s good. You verified that what I’m seeing is pretty much only me seeing it. I have no idea why I have variations.

Thanks for doing the tests and providing the results. I really do appreciate it.

 

[MrFahrenheit]

For Quadra 605 (Performa 475 and LC475) specifications, try http://quadra605.mcdonnelltech.com.
It sounds like the 1 MB VRAM is being used on your system:

• Standard 512 KB VRAM supports 8-bit color for 256 colors at resolutions up to 832×624 and 16 colors at resolutions up to 1152×870
• Upgrading to 1 MB VRAM provides 16-bit color for thousands of colors at resolutions up to 832×624 and 256 colors at resolutions up to 1152×870

I have a Performa 475, which I've upgraded with a full-fat 33 MHz '040.
I can't easily run benchmarks just now, but I can take a look when I next get it set up.

If you can validate what Dr.Diesel showed and let me know I would appreciate it.

 

[MrFahrenheit]

I was able to stabilize some results between different 68040 CPUs by applying a heat sink and thermal paste. Measuring the temperatures I was able to cut the CPU temperature on the pink/grey ceramic over half from about 50’c during benchmarks to around 20’c (with ambient room temperature of 17’c).

However, for some reason I still see performance degradation with the jumper installed on a 475/Q605 board in 8 bit video tests. This is validated over 3 different logic boards, returning similar results. When I say performance degradation, I mean by a wide margin. At 33mhz overclock I see Norton System Info (from Norton 3.1) return 134 on a jumper less 475 and when jumper installed I see 120. I validated it against Norton 6.0 and speedometer and both show similar percentages of degradation.

For better heat transfer, I applied a pea size of Noctua paste in the center and a ring of thermal tape around the outside. The one photo shows the spreading of the dab of paste and the area it occupies. The tape transfers heat and also holds down the cpu, but the paste transfers the heat better. By putting only a ring of tape, it is more easily removed as well.

 

[dr.diesel]

This is validated over 3 different logic boards

I realize all of these variants have the same board, but is the firmware the same across all 3 models? I don't see a FW sticker on my Q605 board, not sure how pull this information? But perhaps all of yours are 475 versions and mine is a Q605 thus the difference?

Probably all the same, but might be worth the conformation.

 

[MrFahrenheit]

I realize all of these variants have the same board, but is the firmware the same across all 3 models? I don't see a FW sticker on my Q605 board, not sure how pull this information? But perhaps all of yours are 475 versions and mine is a Q605 thus the difference?

Probably all the same, but might be worth the conformation.

I first noticed the slower speed on a Quadra 605 I bought a month ago. I tried a few things and then thought that maybe recapping it would improve the performance. It didn’t (as you can see from my screenshot).

I’ve compared against another Q605 that I have along with 2 other LC475. On all of them, jumper installed the 8 bit video benchmark is affected negatively, and regular performance with it removed.

I can find no differences between what came in two of my Q605 machines, and the other 8-10 LC475 machines. I have at least 10 of these in working condition.

When I get time I will run multiple different benchmarks on multiple combinations of these 10+ machines and try and find the commonality and differences.

I still do not understand how on every 33mhz overclock that I have, 16bit video performance is affected unless I use 60ns VRAM, and yet multiple people can’t validate that themselves.

What could be happening that is affecting all of my boards simultaneously? And it’s not just that maybe I could have done something wrong when I recapped them, because I have validated on stock non-recapped boards.

Interestingly is that on the same set of hardware but different mask 68040 CPUs I see variations in performance beyond the normal slight variations.

 

[GiGaBiTe]

I'm a few years late to the party, but I came across this thread when browsing information on the J18 jumper.

I've hotrodded one of my Quadra 605s to 40 MHz with a full 68040. Removing J18 made a measurable, but insignificant performance difference, depending on the test. Results were 2-5% better with the jumper off.

I don't have the same benchmark suite you used, just MacBench 3.0.

My machine has 20 MB of RAM, 1 MB of VRAM, both are rated at 70ns. The motherboard RAM is also rated at 70ns. The tests were run at 640x480@8bpp. Changing to 15 bit color cratered performance. At 640x480, the performance hit was 25-32%.

Not sure what mask revision my 68040 is since it has a heatsink attached, and it's not easily removable.

 

[MrFahrenheit]

I'm a few years late to the party, but I came across this thread when browsing information on the J18 jumper.

I've hotrodded one of my Quadra 605s to 40 MHz with a full 68040. Removing J18 made a measurable, but insignificant performance difference, depending on the test. Results were 2-5% better with the jumper off.

I don't have the same benchmark suite you used, just MacBench 3.0.

My machine has 20 MB of RAM, 1 MB of VRAM, both are rated at 70ns. The motherboard RAM is also rated at 70ns. The tests were run at 640x480@8bpp. Changing to 15 bit color cratered performance. At 640x480, the performance hit was 25-32%.

Not sure what mask revision my 68040 is since it has a heatsink attached, and it's not easily removable.

View attachment 74586

Interesting…

Thank you for validating my discovery.

I wonder why the jumper causes that to happen.