Party like it's 2009

[DrMrLordX]

I am still living with the past on socket AM3. K10.5, why can't I quit you?!? Oh well.

Seriously, I found an Athlon II x2 220 on eBay for $28 shipped, and it was a C3 chip. The IHS said "CACDC" which meant it was probably a failed Phenom II x4 920. I ordered it, got it working tonight with some prodding (after the dreaded "FF" had me stymied for a bit), and . . . it unlocks. Sort of.

Four cores won't get through the Windows boot without crash/autorestart. Three cores (1, 3, and 4) WILL boot to Win7, but as soon as I try to load all three cores, CPU usage drops to 0 and the system gets choppy. Nothing crashes, it just acts weird. Weird as in useless.

I haven't tried running it with two cores disabled to gain the benefit of L3 . . . might try that later.

edit: Unlocked with Core 1 and 4 only, the system seems stable. Behavior is normal (Prime95 is loading both cores). My core temperature readings dropped to 0c, and I now have a Regor with 6mb L3 cache according to CPU-z. Unless I find any odd behavior under this configuration, I can probably conclude that the L3 is not the problem on the chip.

Any recommendations on what voltages I should push to get this thing stable as a Heka? Or a Deneb for that matter? I tried pushing 1.45v vcore to start, and I ran up CPU-NB to 1.22v and NB to 1.2v. It didn't seem to help the chip in Heka mode one bit.

If I can't stabilize it, no big whoop. It's still a big step up from the Sempron 140 it just replaced. Poor thing wasn't even stable at stock speeds below 1.38v vcore. Sad.

 

[Headfoot]

Have a Phenom II x2 rig that will boot with 4 cores unlocked, but will do that same "acting weird" thing you mention as soon as I load it at all. If I only unlock cores 0,1,3 then it will work, but it will eventually crash and become unstable through the week. I've tested the overclocks at stock voltage, at +.1v up from stock and maxed it out at nearly 1.5v vcore to see if it would make any difference. Had CPU-NB and NB at 1.2v, LLC at max. After too much time wasted I've settled with running it as a dual core with the best OC I can get on those two.

Some of the extra cores are likely truly screwed up and it seems like you got one of them. Free L3 is nice though.

 

[DrMrLordX]

Yeah, I do like the L3. As a dual core, this thing seems like a champ, at least by 2009 standards anyway.

It just finished a 12 hour run of Linpack (max memory) at 3.7 ghz, 1.36v vcore. Granted, this is just an isolated clock speed (NB is actually down to 1850 mhz at the moment), but still . . . 3.7 ghz @ stock vcore? I can dig it.

I still think there's something I'm missing about how to stabilize those cores. Maybe PLL voltage or HT voltage? Probably not worth the risk of burning up the chip to find out, but I'm still curious.

 

[Headfoot]

I tried a number of the voltages, PLL and both NB, nothing seemed to stabilize my unlock cores. Spent a solid week testing it. Your chip may be different, so I can't say for sure. But it does sound very familiar to how my experience went. 3.7@ stock is really solid, I think I got 3.6 @ one notch up from stock and 3.7 at +.1v, pretty nice chip you got except for the extra cores!

 

[DrMrLordX]

Thanks. Yeah, I can't really complain. The 14x multi is getting to be a headache, especially since the memory controller seems a lot less stable on 1: 3.33 than 1 : 4 but oh well. If I have to run DDR3-1537 again, then so be it. Using 1 : 4, I can get DDR3-1850 stable, or so it seems.

edit: woops, earlier I said it was cores 1 and 4. It's cores 1 and 3. 2 won't boot windows, and 4 causes the flaky "throttling"-style behavior.

 

[DrMrLordX]

Still no luck stabilizing cores 2 and 4. But . . .

It looks like I was totally wrong about the effect of L3 cache on K10.5's ability to handle programs with a large memory footprint. I had a theory that L3 would actually slow you down if enough system memory was in use on a regular basis. For example, I had expected that SuperPi 32m runs would be faster on K10.5 chips without L3 than those with L3, and at least as far as Deneb vs. Sargas/Propus was concerned, I had thought myself to be correct.

Turns out I may have been horribly wrong. Since this chip has L3 that I can enable or disable at will, I decided to run SuperPi 32M both with and without L3 enabled at the following settings:

CPU: 3.95 ghz
NB: 2736 mhz
RAM: DDR3-1621 6-7-6-18 1T

With L3: 17m 8s
Without L3: 17m 20s

I re-ran both configurations multiple times, and those are the best ones I got. That was totally unexpected. Predictably, the L3-enabled configuration was considerably faster in 1M, but that should come as no surprise.

In any case, I have to step up and admit that my guesswork was just plain off. If there was any one benchmark that could have or should have performed better on an L3-less chip than on one with L3, it was SuperPi 32m (or maybe linpack when using all memory). I'll have to check out linpack throughput and see what happens there.

This has also forced me to reconsider the effects of not including L3 on all of AMD's APUs, from Llano up to Kaveri. Die space is at a premium, granted, and I'm not entirely sure what good L3 would do for the iGPU side of the chip. AMD also has problems with cache density (at least as compared to Intel, anyway), so coming out and saying, "hey, AMD! Put L3 on your APUs!" may not be the most reasonable thing to do.

That being said, the Pentium G3258 seems to get by okay with a paltry 3MB L3 cache (yeah, I know, different uarch, different cache implementation, etc etc). Including 2-3 MB L3 on the APUs might not be that bad of an idea, if they could find room to fit it in there. Methinks they probably know that, and just can't "find the room" while keeping costs down.

edit: Linpack throughput:

L3: peak 27.5 Gflops
no L3: peak 26.2 Gflops

(same settings as the Pi 32m runs)

L3 helps here quite a bit more here than it did on Pi 32m. I'll leave it to others to draw conclusions about that

 

[SPBHM]

L3 always had a big impact when it comes to gaming, if you compare PII to AII X4 for example,
the same for the APUs, you can compare the 5800k-6800K to the 4300-4350 and it's also apparent in some games that the APU is loosing performance because there is no l3 cache

but for many things L3 can have no impact,

 

[Chicken76]

you can compare the 5800k-6800K to the 4300-4350 and it's also apparent in some games that the APU is loosing performance because there is no l3 cache

If you downclock a 4350 to 4300 level, is there a measurable difference in games? (not mocking here, I really don't know. I've always attributed the difference to the clock difference)

 

[PG]

If you downclock a 4350 to 4300 level, is there a measurable difference in games? (not mocking here, I really don't know. I've always attributed the difference to the clock difference)

To your question yes, clock speed can make a small difference sometimes, but I think you are missing his point.

See this page, and scroll down to L4D performance: http://www.tomshardware.com/reviews/athlon-l3-cache,2416-6.html
They compared a phenom II to an Athlon II, both at just 2.6 Ghz.
"Left 4 Dead is different. The processor with L3 cache delivers almost 20% faster frame rates."

L3 by itself sometimes made a huge difference with the Athlon II / Phenom II architecture.

 

[sm625]

With L3: 17m 8s
Without L3: 17m 20s

I re-ran both configurations multiple times, and those are the best ones I got. That was totally unexpected. Predictably, the L3-enabled configuration was considerably faster in 1M, but that should come as no surprise.

Uh, that's 1.17% faster. Hardly what I'd call "considerably".

 

[Chicken76]

To your question yes, clock speed can make a small difference sometimes, but I think you are missing his point.

See this page, and scroll down to L4D performance: http://www.tomshardware.com/reviews/athlon-l3-cache,2416-6.html
They compared a phenom II to an Athlon II, both at just 2.6 Ghz.
"Left 4 Dead is different. The processor with L3 cache delivers almost 20% faster frame rates."

L3 by itself sometimes made a huge difference with the Athlon II / Phenom II architecture.

Your example compares a CPU without any L3 cache, with another that has L3 cache. While the presence of the L3 cache is a benefit undoubtedly, adding more will always provide diminishing returns.
So my question remains: is there a tangible benefit in the case of the 2-module Vishera in having 4 MB of extra L3 cache in games?

 

[zir_blazer]

I through but never confirmed that the few Athlon II X2 models that had 512 KB Cache L2 per Core instead of the full 1 MB (Which is exclusive to Regor) could use heavily harvested Propus or Denebs. A bit too late, but I'm happy that someone confirms that.

 

[DrMrLordX]

Uh, that's 1.17% faster. Hardly what I'd call "considerably".

Those were 32m results. You're reading my statement that 1m times were considerably faster. My Sempron 140 put up times in the 20s range, but the x2 220 with L3 enabled is in the mid 17s range

 

[DrMrLordX]

I through but never confirmed that the few Athlon II X2 models that had 512 KB Cache L2 per Core instead of the full 1 MB (Which is exclusive to Regor) could use heavily harvested Propus or Denebs. A bit too late, but I'm happy that someone confirms that.

Don't know why I ignored this post, but I thought I'd add something more: not only were there 512k L2 cache Athlon II x2s that were failed quads, but there were some 512k L2 cache native dual-cores as well. Those were the Regor parts that were made into Sempron 140s/145s/etc. when one core was bad.

But, I digress! I have stabilized a third core (Core 4) by raising the PLL voltage by .1v (from 2.5v to 2.6v). At least, I think that's what did it. It's prime testing right now with everything else at stock. Dunno why I didn't try this earlier.

If any of you scrape up a cheap x2 on eBay and luck out on a CADAC or CACDC chip, try some extra PLL voltage. It might help with your unlock (or it might not).

edit: Now it's stable at stock (2.5v) PLL voltage with three cores. I guess torture-testing it at high voltages convinced it to make a third core work? Weird.