This is very impressive!

cyclone3d wrote on 2022-02-21, 07:09:

I'm a native "English" speaker and string/thread are interchangeable though thread is usually used more in relation to sewing.

And thread is used for forums where string usually isn't 😜
I always assumed it had to do with being a long single thread of replies.
Not a native english speaker here, so perhaps I missed something.

Wondering if the CPU works if that 0R resistor is removed but not reinstalled in the other position, and if so what amount of cache it shows? 😀

I wonder if they binned 250nm K6-III chips the same way. For example K6-2P chips with disabled cache.

I just want to add this for anyone else seeking to decap their K6-2+

To soften up the silicone adhesive dunk the CPU in acetone upside down for a while, it completely softens it up and you can slice the lid off like it were attached with butter.
Almost no force required.

Warlord wrote on 2022-02-21, 05:49:

I can't see from a business standpoint how selling a otherwise perfectly good chip for less money, gets you more money. Ok I can understand it with intels logic in convincing you to pay more for a non cripppled version. But were talking about AMD here and in what was already a low cost chip that was for low end laptops. k62+ must be binned for having faulty 256 cache that can't run at full speed. It might work downclocked.

End result might be a chip that could run at 600mhz with 128k cache, but can now only run at 400mhz with 256k.

Well AMD has done this before and after 😀

I have a DX2-80 that can be used as a dx4, I have a 486dx4 with "8kb of write-back L2 cache" that actually has 16kb and can be used as a 5x86, some people spoke about the thunderbird which is true, but there's also the AMD x3 that were x4 with a disabled core, except people figured out that most had their 4th core still working, so you could get a x4 for the price of an x3 😀

Sometimes perseverance is indeed worth it, good job finding this, with the humongous source of k6-2+ and this brand new mod, this may be revolutionary for socket 7 machines

This now begs for benchmarks of the modded vs unmodded K6-2+/570s.

The actual chip allocations and configurations are generally determined by market demand, not ideal configurations.
As such if there is a large demand for K6-2+ 500's but not much for K6-3+ 500's using the same die and at the same time a large population of K6-3+ dies are meeting full spec, but there are not many low yield chips, then AMD and indeed many manufacturers will downbin some of their chips to meet demand of the lower SKU's.

Intel and AMD have both been doing this with every single Ryzen launch and it's why we had those odd Ryzen 1st gen chips that showed up with 8 cores, but should only have 4/6 functional.

It's just Bizniz yo 😜

If the seller ever decides to unload them really cheap maybe some of you guys could make money converting them to K6-3’s.

Yes, the presence of a jumper makes it possible to sell K6-2+ as a product option if yield only really affected the cache, you just disable one half of them. There are probably two jumpers for this purpose each controlling one half. But while people are looking as yield as the reason for these, there is another angle here. It could be that the buyer of these chips wanted a higher frequency. But to deliver at 570MHz, the K6 design caused heat and required higher voltage to run stable. But since the customer is in the mobile space, higher voltage and heat is not good. So in order to meet the spec of the mobile customer, all 570MHz would be marked as K6-2+ to disable 1/2 of the cache transistors to keep it in spec, even if there are no yield issues actually.

So I think it's possible that this run was all marked as K6-2+ even if it was possible that it could run fine, to limit the voltage/heat, since there is virtually no K6-3+ marked at this speed. It might be that all these chips were forced with the same 1/2 cache jumper disabled, except in the rare case, they might have had a yield issue where then they could swap to the other 1/2 good cache. I think this is the case based on the two reports on the mod so far, one for higher voltage, and one about needing better cooling. We are seeing the limit of the design for mobile at 180nm, at least as it was intended.

Well, you can already find the same new 570 chips beenn sold at ebay for 99€. Waiting to see price of the first K6-III+ 570 😁

We could name it the "Xi Jinping looks Like Winnie the Pooh K6-3+ Mod" to keep the Chinese remarkers off it 🤣

0xCats wrote on 2022-02-21, 13:31:

I just want to add this for anyone else seeking to decap their K6-2+

To soften up the silicone adhesive dunk the CPU in acetone upside down for a while, it completely softens it up and you can slice the lid off like it were attached with butter.
Almost no force required.

Excellent tip. Will probably use it to sacrifice one of my 570s to the modding gods.

Warlord wrote on 2022-02-21, 06:19:

I hope I'm wrong and I'm mostly going by the rumor that k6 yields were not good. Where the general consensus has been something like only 1 in 10 k62s would overclock more than 50mhz. As in a 450 would usually get 500mhz but not 550. Which means that a K6450 is not simply a gimped 550, but actually a binned cpu due to bad yields.

K6III+ is also a much more rare cpu than a K6II+ and higher clocked versions of K6III+ are also more rare than lower clocked ones. Which leads me to believe there really was a yield issue.

Well there were phases of that... The initial "chompers" K6-2 production had trouble hitting 400Mhz, most were bogging down at about 350, some you could squeeze up to 366/375, I think they got a handful of 400Mhz parts out of it... then they revised it to the CXT core and they broke 400 easy and found the next walls at 500+ but were able to tune the plant after a few months to release 550 chips... but then the + series was a die shrink, which process node is known to be good to a ghz plus, so the yields should have been fine for 500Mhz plus, but somewhere over 600 they start hitting the architectural wall... though also the power most boards can supply is capped around 30W so that's a factor too.

Edit: Oh I meant to say somewhere that time had caught up with the K6 line, as regards what speeds the market thought were worth buying so they may have preferred to sell "Highest clock possible" 2+ instead of 3+ capped at 450 for power and heat reasons.

BitWrangler wrote on 2022-02-21, 15:07:

We could name it the "Xi Jinping looks Like Winnie the Pooh K6-3+ Mod" to keep the Chinese remarkers off it 🤣

You may have just permanently excluded all of our Chinese members.😦

Ooops, forgot we had some guys there, sorry fellas. I'll leave it to mod decision, if they can tell they're all using VPN to get here anyway since there has not been much filtering of criticism of China in other posts, it might not be a biggie.

I'm curious to see if those cpus are really stable when the cache is enabled, if so this is a game changing discovery!

Nemo1985 wrote on 2022-02-21, 16:27:

I'm curious to see if those cpus are really stable when the cache is enabled, if so this is a game changing discovery!

Prime 95 Blend or large fft test (good for cache) for ~8 hours and 3DMark 2000 in loop

For cache stability test, memtest 86 is a good candidate, oxCats told us that if the cache is bad, with the 256k range set in memtest, errors should appear very quickly. See sphere 478's post about memtest to see what you need to do

NostalgicAslinger wrote on 2022-02-21, 16:29:

Nemo1985 wrote on 2022-02-21, 16:27:

I'm curious to see if those cpus are really stable when the cache is enabled, if so this is a game changing discovery!

Prime 95 Blend or large fft test (good for cache) for ~8 hours and 3DMark 2000 in loop

For testing in Windows plain Prime 95 with small FFT is fine for finding unstable L2 cache. No need for large FFT, large FFT functionally adds nothing to a cache test except more time.

Though I generally prefer just the ease of slamming the L2 in rapid succession with memtest using small test address range that kinda breaks memtests anti-caching logic.
From personal testing I know CPU's with bad caches fail this really quickly.