VOGONS


AMD K6 3DMARK, aiming for stars.

Topic actions

Reply 440 of 539, by Necrodude

User metadata
Rank Member
Rank
Member
Sphere478 wrote on 2022-05-06, 21:05:
Freakin awesome!! Is the increased height an issue? […]
Show full quote

Freakin awesome!! Is the increased height an issue?

I was pondering trying to get cnc copper replicas made of the oem spreader. I have a spare lid I can send off.

What cpus are those from? Did you have to shorten or modify them at all? I see you sanded them to the copper on top and bottom perimeter.

Btw idk if you saw my report about recent cpus I’ve modded but I keep getting failed super pi runs on multiple systems. I can’t figure out why. But the cpus are stable in everything else. What’s particularly interesting is that it is happening on my OEM 550 3+. So I’m convinced the cpus aren’t at fault. I’m still unclear what the issue is though. Maybe software, maybe other hardware. Unsure. But it’s on multiple systems.

Re: SUCCESSFUL K6-2+ to K6-3+ Full Cache Enable Mod I added this to the k6 mod guide btw 😀

The height could be a problem. It depends on the cooler you own. G3virus encountered some problems, the he switched cooler, and that cooler fitted just fine.

About the cpus. We scrapped a socket fm1 amd a6 dual core apu, and a 754 sempron cpu. The heatspreader on those cpu:s are identical.

Our K6-2+ cpus modded to k6-3+ halted at the boot screen with l2 cache error codes. So it was a no go.
Still they run 100% stable at 672mhz! ( 112x6 ).
The sad thing is that my k6-3+ at 616mhz is faster then those cpus. That extra 128kb l2 cache makes all the difference.

Reply 441 of 539, by pa1983

User metadata
Rank Newbie
Rank
Newbie
Sphere478 wrote on 2022-05-06, 21:05:
Freakin awesome!! Is the increased height an issue? […]
Show full quote

Freakin awesome!! Is the increased height an issue?

I was pondering trying to get cnc copper replicas made of the oem spreader. I have a spare lid I can send off.

What cpus are those from? Did you have to shorten or modify them at all? I see you sanded them to the copper on top and bottom perimeter.

Btw idk if you saw my report about recent cpus I’ve modded but I keep getting failed super pi runs on multiple systems. I can’t figure out why. But the cpus are stable in everything else. What’s particularly interesting is that it is happening on my OEM 550 3+. So I’m convinced the cpus aren’t at fault. I’m still unclear what the issue is though. Maybe software, maybe other hardware. Unsure. But it’s on multiple systems.

Re: SUCCESSFUL K6-2+ to K6-3+ Full Cache Enable Mod I added this to the k6 mod guide btw 😀

I am G3Virus, its my Discord name on Necrodudes server.

The IHS is from Semprons and other crappy AMD cpu's like Athlon II X2 215, they are not soldered so deliding in the vice is easy.
They are 3.1mm tall lapped and installed, thats 1.5mm taller then original so twice the height, also tualatins are 2mm tall.
Basically nothing with a clip will fit unless modified in some way.
Once I get my new motherboard VRM heatsinks (yea 672Mhz runs the mosfets a bit toasty) I will make a YT video on my build with one of the K6-2+ 672Mhz CPU's that I got to keep.

I initially binned both at 2.2V because 2.0 was not stable and 2.1V I believed was unstable but that was just some driver issue with 3dmark2000 so I think 2.1V is fine.
I rebinned a K6-2+ 570@672 with 2.13V prime95 stable (24h) later.
Planing a build on that myself tough not as fast due to MVP3 chipset limitations, should do 4K+ in 3Dmark2000.
K6-3+ 616 is faster then K6-2+ 672 in both boards I tested with a about 2-6%.

Would have been awesome if the L2 cache mod had worked but it was stuck on C3 = DRAM error when ever the mod was enabled.
Was going to do a howl YT video about it but well after 18h or so I just sad F it and moved on to the LM mod.

I actually dont know where the limit is on those K6-2+ because the boards are the limit, my External cache is a factor, I can get in to windows at 582Mhz (124Mhz fsb) if the MB Cache is disabled but could not run 3dmark2000 anyways and I have not tried higher because of board issues.
Would not surprise me if the CPU can do 700 ish Mhz.

My experience with k6-2 back in the day and K6-2+ now is that running them colder means higher clocks, (might seem obvious but K6-2 type cpus seems to respond very well to "cooling") and often lower voltage requirements and it also enables higher voltages to be used if needed but I found higher voltage on Necrodudes K6-3+ 400@616 to be more of a problem.

There where 3x K6-2+ 570Mhz in the batch and one was unmodified and went to someone else and he reported 2.2V was unstable at 600Mhz but 2.0V worked fine so yea stock IHS is kind of garbage.

First time I hit 600Mhz on SS7 was with a K6-2 500 running 2.9V delided that could do 24h prime95, that was back in 2000-2001.
From what i recall from 22 years ago I usually gained 50Mhz from deliding or could lower Vcore with 200mV at the same clock after deliding a normal K6-2.

Im not sure I would go trough with this mod again if I knew the L2 cache mod would be a failure, was a lot of work 🤣
But the LM and Sempron IHS do seem to work very well offering high clocks at low Vcore with perfect stability.

Website, YouTube, Discord, Gallery

Reply 442 of 539, by Necrodude

User metadata
Rank Member
Rank
Member

Adding some more pictures from the LIquid metal copper IHS mod

Attachments

  • IMG_2187.jpg
    Filename
    IMG_2187.jpg
    File size
    1016.97 KiB
    Views
    2220 views
    File license
    Public domain
  • IMG_2185.jpg
    Filename
    IMG_2185.jpg
    File size
    875.74 KiB
    Views
    2220 views
    File license
    Public domain
  • IMG_2189.jpg
    Filename
    IMG_2189.jpg
    File size
    1.23 MiB
    Views
    2220 views
    File license
    Public domain

Reply 443 of 539, by bloodem

User metadata
Rank Oldbie
Rank
Oldbie
pa1983 wrote on 2022-05-08, 00:39:

There where 3x K6-2+ 570Mhz in the batch and one was unmodified and went to someone else and he reported 2.2V was unstable at 600Mhz but 2.0V worked fine so yea stock IHS is kind of garbage.

I've modded five so far and only one of them was a dud (so I reverted it back to K6-2+).

In my experience, the stock IHS is more than enough for these chips (as long as you don't plan on breaking any records). Before the mod, all of them were working just fine at 633 MHz @ 2.1V, and this continued to be true after the mod as well. So I wouldn't blame the IHS, you probably just got a bad batch of CPUs.

1 x PLCC-68 / 2 x PGA132 / 5 x Skt 3 / 9 x Skt 7 / 12 x SS7 / 1 x Skt 8 / 14 x Slot 1 / 5 x Slot A
5 x Skt 370 / 8 x Skt A / 2 x Skt 478 / 2 x Skt 754 / 3 x Skt 939 / 7 x LGA775 / 1 x LGA1155
Current PC: Ryzen 7 5800X3D
Backup PC: Core i7 7700k

Reply 445 of 539, by Necrodude

User metadata
Rank Member
Rank
Member
bloodem wrote on 2022-05-08, 10:09:
pa1983 wrote on 2022-05-08, 00:39:

There where 3x K6-2+ 570Mhz in the batch and one was unmodified and went to someone else and he reported 2.2V was unstable at 600Mhz but 2.0V worked fine so yea stock IHS is kind of garbage.

I've modded five so far and only one of them was a dud (so I reverted it back to K6-2+).

In my experience, the stock IHS is more than enough for these chips (as long as you don't plan on breaking any records). Before the mod, all of them were working just fine at 633 MHz @ 2.1V, and this continued to be true after the mod as well. So I wouldn't blame the IHS, you probably just got a bad batch of CPUs.

Yeah its sound 100% reasonable that most k6-2+ works just fine at 630mhz.

Did any of your k6-2+ cpus stay prime95 stable for 24h at 672mhz with the original IHS?

We are blaming the k6-3+ modding failure on a bad batch of cpus. They didnt work modded to k6-3+ at any frequency.

The whole plan with the copper IHS was to get increased overclocking headroom on k6-3+ modded cpus.
Since the mod failed we settled on more overclocking headroom without the k6-3+ mod.

Reply 446 of 539, by pa1983

User metadata
Rank Newbie
Rank
Newbie
bloodem wrote on 2022-05-08, 10:09:
pa1983 wrote on 2022-05-08, 00:39:

There where 3x K6-2+ 570Mhz in the batch and one was unmodified and went to someone else and he reported 2.2V was unstable at 600Mhz but 2.0V worked fine so yea stock IHS is kind of garbage.

I've modded five so far and only one of them was a dud (so I reverted it back to K6-2+).

In my experience, the stock IHS is more than enough for these chips (as long as you don't plan on breaking any records). Before the mod, all of them were working just fine at 633 MHz @ 2.1V, and this continued to be true after the mod as well. So I wouldn't blame the IHS, you probably just got a bad batch of CPUs.

I talked to the owner of the 3:e unmodded 570 from the same batch and it refused 600@ 2.2V until Vcore was lowered.
I seen first hand how much the IHS sucks on K6* processors over the years, the glue used is garbage, the IHS is usually not flat in the middle.

Its just garbage and leaves a lot of potential on the table.

But thats my opinion based on my experience.

Like I also sad, not recommending the LM mod for everyone but it sure keeps the thing cool and stable and 633 VS 672Mhz is a big difference.

Website, YouTube, Discord, Gallery

Reply 447 of 539, by bloodem

User metadata
Rank Oldbie
Rank
Oldbie
pa1983 wrote on 2022-05-08, 13:30:

I talked to the owner of the 3:e unmodded 570 from the same batch and it refused 600@ 2.2V until Vcore was lowered.

That is without a doubt a behavior specific to that particular chip (maybe because of the additional heat that comes with 2.2V, maybe due to other more low level quirks that it might have).
However, what I do know is that this is not the typical behavior. I've tested many of these chips at 2.1, 2.2 and even 2.3V, and most will work just fine as long as you have a good enough cooler.

So, even though the heatspreader is not the best in the world (obviously), it does its job in most cases. In fact, my fastest SS7 build actually has a K6-3+ 400ATZ which is overclocked at 633 MHz (115MHz x 5.5) and it required 2.2V for 100% stability. It has worked like this without issues for a few years now (with a more beefy socket 370 cooler).

pa1983 wrote on 2022-05-08, 13:30:

I seen first hand how much the IHS sucks on K6* processors over the years, the glue used is garbage, the IHS is usually not flat in the middle.

Yes, it does suck and that original silicone glue is definitely not the best. But, again, it's typically fine for normal use and even overclocking (with a good chip, it definitely works up to 633 MHz at 2.1 & 2.2V).

pa1983 wrote on 2022-05-08, 13:30:

Like I also sad, not recommending the LM mod for everyone but it sure keeps the thing cool and stable and 633 VS 672Mhz is a big difference.

Obviously! 670+ MHz is definitely the type of use case (extreme overclocking) where every improvement counts. However, I was talking about the 600 MHz 2.2V overclock issue.

Necrodude wrote on 2022-05-08, 12:47:

Did any of your k6-2+ cpus stay prime95 stable for 24h at 672mhz with the original IHS?

Did not even try going higher than 633 MHz with these chips.
The overclock I was aiming for was a "great overclock but still one that should be achievable by a good percentage of these CPUs" (for a series of youtube benchmarks that I uploaded).

The Youtube video links and more information regarding why I did not bother with record breaking overclocks can be found in this thread, where my initial goal was comparing the performance of a VIA C3 Ezra-T/440BX build to that of a very fast SS7 build.
If you're interested, you can also watch the K6-2+ to K6-3+ modding guide I did here: https://www.youtube.com/watch?v=5Aw2U_BZnSk.

1 x PLCC-68 / 2 x PGA132 / 5 x Skt 3 / 9 x Skt 7 / 12 x SS7 / 1 x Skt 8 / 14 x Slot 1 / 5 x Slot A
5 x Skt 370 / 8 x Skt A / 2 x Skt 478 / 2 x Skt 754 / 3 x Skt 939 / 7 x LGA775 / 1 x LGA1155
Current PC: Ryzen 7 5800X3D
Backup PC: Core i7 7700k

Reply 448 of 539, by pa1983

User metadata
Rank Newbie
Rank
Newbie

What Im trying to do is replacing my SIS 530 AT baby board and K6-2 550@600 with a ECS AT Baby board based on MVP3.
Because thats what I have for a case, AT, and Necrodude gave me the board, cpu and the case at some point for recaping and other stuff as a trade.

So Im not aiming for a record, I had a very similar Shuttle HOT-591P board back in the day with a K6-2 500 delided capabel of 600@ 2.9V prime95 24h run and had a GeForce 2 GTS with that, yea makes no sens but that was what I had 🤣.
A GeForce 2 MX makes a lot mos sens for a daily driver SS7 system but I want to do the unbalanced shit I did back then because of reasons haha.

So I recently repaired a GeForce 2 PRO, and I still have my original SB16 from my Pentium 166/K6-2 500 rig and before I went Athlon Thunderbird 1400 I did try to find a K6-2+ in sweden back when but gave up on that.

So this build is not identical to my 2000 setup but it will use some parts from that build and also similar parts and parts I wanted so it would represent the peak of what I was aiming for back then.

Running 672Mhz rock solid, been running the system on my cardboard box for a while now waiting for my new VRM heatsinks from china(they are soldered to the board so got taller versions of that), played some CS 1.5 the other day with Necrodude and his friends and it was a better experience then my Athlon 800 with Voodoo3 3000, idk V3 never really liked the smoke in CS.

So yea aiming for a fast and 100% stable SS7 system that can be used from Friday to sunday on a lan with good performance and stability and be stupidly unbalanced 🤣

So fare I got to 4043 3Dmark2000 but I should break that, I just nuked my install with DX9 (for testing) that screwed the performance a bit so once I reinstall with all the hand picked drivers, DX7 and so on I think 4100 3Dmark 2000 should work.
Quake 3 is doing like 141 fps I think now, with some quality settings about 125fps, max textures no vertex lighting etc.

The aladin 5 board Necrodude uses had a 500 point instant upper hand on my mvp3 board with the same K6-3+ he uses. He has tweaked another 600 points out of that, I got another 200-250 out of my board since then, hoping for another 100 points.

So benchmarking is fun but its mostly to fine tune driver selection, DX, wpcredit, stuff like that.
Once I have a final score I probably post that.
Also I will use my Adaptec 19160 SCSI controller and drive I bought new back in like 2004, getting 72Mb/sec read and write on that, could never afford that back in 2000 so that will be cool, games load FAST!

Website, YouTube, Discord, Gallery

Reply 449 of 539, by Sphere478

User metadata
Rank l33t++
Rank
l33t++

I’d be interested In any info you guys have about how to directly access the + cpu registers in software. I’m wondering if this chip may have gotten a multiplier controller that can give more than the 8 known multipliers that are able to be programmed via the three hardware registers.

It seems through software, the hardware multiplier can be overridden. Maybe there is also a soft mod for this mod also

Sphere's PCB projects.
-
Sphere’s socket 5/7 cpu collection.
-
SUCCESSFUL K6-2+ to K6-3+ Full Cache Enable Mod
-
Tyan S1564S to S1564D single to dual processor conversion (also s1563 and s1562)

Reply 450 of 539, by pa1983

User metadata
Rank Newbie
Rank
Newbie

I finally got my K6-2+ @672Mhz system built and prime95 tested for 24h and its stable!

4038 points so no record breaker but I am a bit limited due to the chipset used and I also aimed for a usable stable system with some hardware I used back in
my old K6-2 500@600 system and similar hardware i used.
I hade a GF2 GTS back then so a GF2 PRO had to stand in for that and also the motherboard was a Shuttle HOT-591P but a very similar ECS board had to stand in for that.

System Specs:
MB: ECS P5SD-B+ with modded bios, recaped plus extra filer caps and improved VRM cooling and PS2 port added.
GFX: GeForce 2 PRO 64MB
CPU: K6-2+ 570@672Mhz @ 2.13V
RAM: PC133 128MB stick, 2-2-2-5 timings
HDD: 10K SCSI with Adaptec 19160 controller.
Soundcard: CT2890, Soundblaster 16.
Nic: ADMtek AN983B

YAamJRlh.png
a8Pw4ill.jpg

I made a Youtube video about the build.
aoHmd2rh.png
https://youtu.be/A_OAFrsyC9o

KOHYUS2l.jpg
P1G9mhSl.jpg
xw9WYnDl.jpg
Also the K6-2+ to K6-3+ L2 cache mod did not work on this one, gave DRAM post code error .
We tested it on my friends board to and his K6-3+ also worked on both boards so seems the extra L2 was nuked on this one.
Sempron IHS, conformal coating on the components. Luiqid metal added to die and IHS and then glued back on with silicone.

Website, YouTube, Discord, Gallery

Reply 451 of 539, by Repo Man11

User metadata
Rank Oldbie
Rank
Oldbie

My system is plateaued between 4,500 to 4,600 points, and I guess it will remain there unless I buy another K6-2+ 570 and win the CPU lottery with a chip that will do 672 MHz. Interestingly, my Quadro 2 Pro score is virtually identical to my score with the Ti 200.

One thing I didn't expect; in 3D 2001 my highest score was achieved with an FX5600, which scores about 500 points higher than a GeForce 4 Ti 4600.

Attachments

"I'd rather be rich than stupid" - Jack Handey

Reply 452 of 539, by bloodem

User metadata
Rank Oldbie
Rank
Oldbie
Repo Man11 wrote on 2022-06-17, 19:12:

My system is plateaued between 4,500 to 4,600 points, and I guess it will remain there unless I buy another K6-2+ 570 and win the CPU lottery with a chip that will do 672 MHz.

Is SDRAM Bank Interleaving enabled? If it's not, you could squeeze a bit more performance out of the system (make sure to fill all memory slots).

Repo Man11 wrote on 2022-06-17, 19:12:

Interestingly, my Quadro 2 Pro score is virtually identical to my score with the Ti 200.

Yeah, not unexpected. CPU bottleneck at its finest. 😀
If you remove most of the CPU bottleneck (with a fast Core 2 Duo), a GeForce 2 Ti/Quadro 2 Pro can score over 12 000 points in 3DMark2000, while a GeForce 3 Ti500 scores ~ 16 000 points (the Ti200 should be around 14k points). So on slower platforms, when using identical drivers, they will both score about the same.

Repo Man11 wrote on 2022-06-17, 19:12:

One thing I didn't expect; in 3D 2001 my highest score was achieved with an FX5600, which scores about 500 points higher than a GeForce 4 Ti 4600.

That's interesting, indeed. What drivers were you using?
Either way, it's not exactly a relevant benchmark for such a slow platform... I personally don't run 3DMark2001 on anything slower than at least a late Athlon/Tualatin.

1 x PLCC-68 / 2 x PGA132 / 5 x Skt 3 / 9 x Skt 7 / 12 x SS7 / 1 x Skt 8 / 14 x Slot 1 / 5 x Slot A
5 x Skt 370 / 8 x Skt A / 2 x Skt 478 / 2 x Skt 754 / 3 x Skt 939 / 7 x LGA775 / 1 x LGA1155
Current PC: Ryzen 7 5800X3D
Backup PC: Core i7 7700k

Reply 453 of 539, by Repo Man11

User metadata
Rank Oldbie
Rank
Oldbie

Bank interleaving is enabled - using Sandra to check memory bandwidth, there's no difference between using a single 256 DIMM and a pair of 128 DIMMs.

I tried both 43.45 and 45.23 with the FX5600 and I used 28.32 and 40.72 with the Ti 4600. I overclocked both to the limits, but the gap remained.

My FX 5500 also beats the Ti 4600 on this system. Again using the 45.23, and this result is with the card right at the overclocking limit.

Attachments

"I'd rather be rich than stupid" - Jack Handey

Reply 454 of 539, by Skalabala

User metadata
Rank Member
Rank
Member

In 3dmark 2001, the score is greater the faster the GPU is in nature test resulting in a higher score.
I proved this with a 5950 Ultra. in the 3dmark game tests the FPS are within margin on Geforce 3 vs etc but in nature test the 5950 gets double the FPS.

Reply 455 of 539, by Oerg866

User metadata
Rank Member
Rank
Member

image.png

- Gigabyte GA-5AX Revision 4.1 + Recap (with increased capacities) + direct wire from ATX 3.3V to AGP 3.3V
- ALi M1541 Revision E
- EliteMT 64128 cache chip (used as 512K, but I wanted the 4ns for extra headroom)
- AMD K6-II+/570ACZ modded to K6-III+ -- FSB 110 x 6.0 -- 2.2V core
- Hyundai 128MB @ CL2-2-2-4
- Few register tweaks in M1541
- AGP driver 1.90 set to "Turbo" mode in ALi AGP utility 1.40
- ASUS GeForce 2 GTS @ 234 MHz Core + 387 MHz Mem -- Driver version 7.76
- Windows 98 Second Edition + 98Lite Sleek + DirectX 8.1

Reply 456 of 539, by Sphere478

User metadata
Rank l33t++
Rank
l33t++
Oerg866 wrote on 2022-12-18, 22:30:
https://cdn.discordapp.com/attachments/1018891218710696046/1054162235775275018/image.png […]
Show full quote

image.png

- Gigabyte GA-5AX Revision 4.1 + Recap (with increased capacities) + direct wire from ATX 3.3V to AGP 3.3V
- ALi M1541 Revision E
- EliteMT 64128 cache chip (used as 512K, but I wanted the 4ns for extra headroom)
- AMD K6-II+/570ACZ modded to K6-III+ -- FSB 110 x 6.0 -- 2.2V core
- Hyundai 128MB @ CL2-2-2-4
- Few register tweaks in M1541
- AGP driver 1.90 set to "Turbo" mode in ALi AGP utility 1.40
- ASUS GeForce 2 GTS @ 234 MHz Core + 387 MHz Mem -- Driver version 7.76
- Windows 98 Second Edition + 98Lite Sleek + DirectX 8.1

What if we could make something like this for agp?

Re: Adding 3.3v to pci adapter(Released)

Not to convert the voltage or anything, but just to give it a little more amps?

Sphere's PCB projects.
-
Sphere’s socket 5/7 cpu collection.
-
SUCCESSFUL K6-2+ to K6-3+ Full Cache Enable Mod
-
Tyan S1564S to S1564D single to dual processor conversion (also s1563 and s1562)

Reply 459 of 539, by Oerg866

User metadata
Rank Member
Rank
Member

No, I just soldered a wire outright, YOLO 😜

Also - 1MB Cache on GA-5AX is happening

IMG_20221219_163408.jpg
IMG_20221219_163545.jpg
IMG_20221219_165903.jpg
image.png

It seems to work even without internal tag on Rev E northbridge!!!

I will test a bit more and then write a blog post for TRW detailing this mod (i hate people keeping secrets, seems to be commonplace around here)