Bad memory performance of F revision chips fixed!

[Carfax]

This news (http://www.theinquirer.net/?article=29776) from the Inq may provide a reason for Agenda2005's comment in my other post about Conroe:

If you dont know, someone have AMD64 AM2 CPU with DDR2-667 and the conclusion was that infact it run slower than the current 939 CPU with DDR1-400 in many instances, so Intel must have had some way of measuring up the performance of that CPU even by merely overclocking it to DDR2-800 and comparing the performance with Conroe.

I guess the first F revision spin had a bad memory bug which caused memory performance to suck, so this is probably why that "someone" referred to in Agenda2005's post had a lower performing chip than previous revisions.

Anyway, I hope this is true because if it is, it means that the F revision chips will NOT be slower than the E revision :banana:

Even if it does delay them somewhat, I'd rather AMD get it right before they start selling them.

[vapb400]

Sounds good.

Are Rev F chips going to be on AM2/M2/940 or just socket f/1207?
I'm assuming Rev F is for any DDR2 platform.

[Sentential]

This news (http://www.theinquirer.net/?article=29776) from the Inq may provide a reason for Agenda2005's comment in my other post about Conroe:



I guess the first F revision spin had a bad memory bug which caused memory performance to suck, so this is probably why that "someone" referred to in Agenda2005's post had a lower performing chip than previous revisions.

Anyway, I hope this is true because if it is, it means that the F revision chips will NOT be slower than the E revision :banana:

Even if it does delay them somewhat, I'd rather AMD get it right before they start selling them.

What was said above is infact true. However the issue was that the dividers were not being properly applied (instead of 1:1 it would revert the RAM to PC2-3200 not 5300) Memtest showed 2,000MB/S with the RAM at "3200"

(I figured I needed to correct the INQ as they did not explain what the bug was and thus making it appear to be more serious)

[turtle]

Wow, that is definately news to me. Thanks for the for the info Sentential. All the more reason to wait for Brisbane...Maybe that will be tweaked a little more than just a dummy shrink.

Here I was thinking it would be prepared to do ddr-800+ at cas3 when said chips were ready, and with a good deal of performance increase over cas4 or 5. I guess that isn't the case, or at least wasn't initially.

At least they're working on it and are/did fixing/fix it.

[Sentential]

Wow, that is definately news to me. Thanks for the for the info Sentential.

Here I was thinking it would be prepared to do ddr-800 at cas3 when said chips were ready, and with a good deal of performance increase over cas4 or 5. I guess that isn't the case, or at least wasn't initially.

At least they're working on it and are/did fix/fixing it.
Keep a close eye on RAM compatibility. There are alot of RAM related issues on this platform as it stands right now :stick:

[vapb400]

Thanks for the info Sentential. Memory compatibility is NOT what I want to here.

Conroe is looking more and more attractive IMO.

[OmegaMerc]

Sounds like they need to drop DDR2 for socket F untill they can adopt the tech without problems.

[Carfax]

What was said above is infact true. However the issue was that the dividers were not being properly applied (instead of 1:1 it would revert the RAM to PC2-3200 not 5300) Memtest showed 2,000MB/S with the RAM at "3200"

(I figured I needed to correct the INQ as they did not explain what the bug was and thus making it appear to be more serious)

So whats the performance like now Sentential?

[Cybercat]

What was said above is infact true. However the issue was that the dividers were not being properly applied (instead of 1:1 it would revert the RAM to PC2-3200 not 5300) Memtest showed 2,000MB/S with the RAM at "3200"
1:1? Suggesting that the actual bus speed of the new A64s is 333? If that's the case, how is it that they're able to deliver even clockspeeds of 2200, 2400, 2600, etc?

[agenda2005]

Hope things work well for them. Although K8 is not bandwith starved and higher latency DDR2 with increased bandwith is only going to be a mix bag, but lets wait and see.

[biohead]

Hope things work well for them. Although K8 is not bandwith starved and higher latency DDR2 with increased bandwith is only going to be a mix bag, but lets wait and see.
im with you

[nn_step]

Actually if you want to get into details.. the 4.3b revision of the memory control has a 15.8% performance edge over s939..

[Carfax]

Actually if you want to get into details.. the 4.3b revision of the memory control has a 15.8% performance edge over s939..

Where did you hear this? Name your source, or state whether you have insider information...

A 16% increase in performance for a memory controller rev sounds a bit fishy!

[[XC] Lead Head]

Its not 16. 15.8, theres a difference :p:

[nn_step]

a huge difference if you are really into the benchmarks...:p:

[Sentential]

So whats the performance like now Sentential?
Not exactally sure but as far as I know the divider issue has been fixed. There's no doubt in my mind that AM2 will be signficantally faster than S939, its just a question of by how much.

1:1? Suggesting that the actual bus speed of the new A64s is 333? If that's the case, how is it that they're able to deliver even clockspeeds of 2200, 2400, 2600, etc?
Ive said that it was 333 from the start. The clock speeds listed on various websites are simply incorrect as they are not even numbers.

Honestally tho ive got every confidence that this will all come out ok in the long run, hell its still very very early for any production chip.

Im more impressed that they even decided to go with a 333HTT bus as Intel is only going to use 266 with Conroe. AM2 should have no problem keeping the crown but its gonna be harder on 90nm than it will on 65nm. Certianly wont be seeing as high as 500HTT 1:1 on conroe lol so that will certianly help things along.

Frankly for most A64 owners none of this should be a suprise as most of them are common issues on S939 platforms

[Revv23]

cool, thanks for the info sentinel, inq made this sound like something that would cause huge delay's, and now it doesnt seem like so much to worry about.

[Pinnacle]

Actually if you want to get into details.. the 4.3b revision of the memory control has a 15.8% performance edge over s939..

Where are you getting this from?

[DeltZ]

if these kinda performance gains from new mem controller are true

AMD>conroe?

Even if conroe DOES have than 20% increase in performance based on just CPU itself. Gonna be hard to keep up with something which gets 15.8% off a new MC revision.

is that figure theoretical or practical, got any more specific details?

Then added the fact that when it goes 65 nm next year or whenever.. +10/15% efficiency.

Are there any major improves after conroe by the time scale that when AM2 goes 65nm?
Or are intel just gonna have to get sick clock speeds?

[krille]

Guys, I think nn_step is talking about their DDR3 MC. Because, you still believe M2 will be DDR3, don't you nn_step? And in that case, DDR3 could very well provide 15.8% performance boost (if not more). However, most seem to think DDR3 won't be until early 2007. Is this correct? Please, speak out!

edit: Or were you discussing quad channel DDR2 for socket F (LGA1207) nn_step?

[nn_step]

I will not speak of the Specs but I will say the Latencies in the revised edition are far better than expected and that the clock speed is a multible of 200..
that is practical application improvement..

[dinos22]

Guys, I think nn_step is talking about their DDR3 MC. Because, you still believe M2 will be DDR3, don't you nn_step? And in that case, DDR3 could very well provide 15.8% performance boost (if not more). However, most seem to think DDR3 won't be until early 2007. Is this correct? Please, speak out!

edit: Or were you discussing quad channel DDR2 for socket F (LGA1207) nn_step?
:hehe:

[nn_step]

I am speaking of Revision F Memory controller revision 4.3b.. how is that too confusing for everyone?

[krille]

I am speaking of Revision F Memory controller revision 4.3b.. how is that too confusing for everyone?OMG!! sorry, i thought you meant 4.3c :stick:

[nn_step]

OMG!! sorry, i thought you meant 4.3c :stick:
you know full well that 4.3c wont be ready for another 2 months:stick:

[dinos22]

ok what planet are you guys from :D

[nn_step]

ok what planet are you guys from :D
Planet NDA.:p:

[Pinnacle]

ok what planet are you guys from :D

Not this one! And NN is smokin crack again HEHE

[nn_step]

Not this one! And NN is smokin crack again HEHE
you know full well that you smoked all my crack already..:fact:
I've been forced to drink copious quanities of Caffine...

[dinos22]

Not this one! And NN is smokin crack again HEHE
that figures.....can't believe they picked up our bad habbits first hehehe

[nn_step]

Learn not to mistake Amp and Crack.. Amp is green and legal.. Crack is White and not as powerful.. :p:

[Pinnacle]

you know full well that you smoked all my crack already..:fact:
I've been forced to drink copious quanities of Caffine...

And I didnt even feel it


HEHE

[nn_step]

And I didnt even feel it


HEHE
Dude you don't even feel your own arm.. and by the looks of it just got shot..:rolleyes:

[metro.cl]

nn am2 will be ddr2 when it launches.

sorry but ddr3 is not posible and ram manufacturers are developing good ddr2 sticks

[nn_step]

nn am2 will be ddr2 when it launches.

sorry but ddr3 is not posible and ram manufacturers are developing good ddr2 sticks
Do you even know what Revision F means? :stick:

[DilTech]

Revision F is the AM2 chips nn, much like Revision E is san diego/Venice/etc, Revision D was winchester IIRC, C was newcastle/clawhammer....

So what does your post have to do with metro's post?

[nn_step]

Revision F isn't M2 dude...

[dinos22]

nn you crack smoking alien :p: speak in human English pls :)

[nn_step]

01101110011011110010000001110111011000010111100100 10000001101001011011100010000001101000011001010110 11000110110000100000011011010110000101101110001000 00010010010010000001101100011010010110101101100101 00100000011011010111100100100000011010100110111101 100010
and that is my final answer

[Pinnacle]

nn you crack smoking alien :p: speak in human English pls :)


hes just a nut by nature, you have to exuse him from time to time:D

[dinos22]

01101110011011110010000001110111011000010111100100 10000001101001011011100010000001101000011001010110 11000110110000100000011011010110000101101110001000 00010010010010000001101100011010010110101101100101 00100000011011010111100100100000011010100110111101 100010
and that is my final answer
i love you too mang :) rofl

[nn_step]

i love you too mang :) rofl
01110100011010000110000101110100001000000110100101 11001100100000011011100110100101100011011001010010 00000110001001110101011101000010000001101010011101 01011100110111010000100000011001000110111101101110 00100111011101000010000001110100011100100111100100 10000001110100011011110010000001100110011001010110 01010110110000100000011011010110010100100000011101 010111000000100000011011010110000101101110
All important details will have to wait a few weeks ;)

[metro.cl]

Do you even know what Revision F means? :stick:

from your point of view it seems i dont know, care to explain??

i know am2 will be ddr2 fo sure from 2 really really reliable sources.

[Pinnacle]

from your point of view it seems i dont know, care to explain??

i know am2 will be ddr2 fo sure from 2 really really reliable sources.

Please, share! or you gonna pull an NN_step make a statmeant and not back it up?

[krille]

Seems like we're back in the sandbox guys, but who can help it? :toast:

01101001011100110010000001101001011101000010000001 10111101101011001000000111010001101111001000000010 00100110001001110010011000010110101101100101001000 10001000000110111001100100011000010010000001101001 01101110001000000110001001101001011011100110000101 11001001111001001111110010000001100010011101000111 01110010111000101110001011100010000001101001001000 00011011000110111101110110011001010010000001111001 01101111011101010111001000100000011010100110111101 10001000100000011101000110111101101111001011100010 11100010111000100000011101110110000101101110011011 10011000010010000001110011011010000110000101110010 01100101001111110010000000111010011101110110100101 10111001101011001110100010000000111010011101110110 10010110111001101011001110100000110100001010000011 01000010100111011101100001011100110010000001001001 00100000011000110110111101110010011100100110010101 10001101110100001000000110000101100010011011110111 01010111010000100000011011010111100100100000011001 01011000010111001001101100011010010110010101110010 00100000011100110111010001100001011101000110010101 10110101100101011011100111010000111111001000000111 01000110100001101111011100110110010100100000011011 10011101010110110101100010011001010111001001110011 00100000011000010111001001100101001000000110011001 10111101110010001000000110010001100100001100110011 11110010000001101111011100100010000001110111011000 01011100110010000001001001001000000110010101110110 01100101011011100010000001100011011011000110111101 1100110110010100111111

[mursaat]

I would never think a processor manufacturer, with all that extremely complicated things to research like silicon electrical conductivity, multiple layers, shrinking nanometers could make such a "simple" error like that.

Or maybe it wasn't so simple, but as I see it, its way more simpler than, lets say, 90nm to 65nm right? :D

[zir_blazer]

Do you even know what Revision F means? :stick:
Sure (http://www.anandtech.com/cpuchipsets/showdoc.aspx?i=2476)

[biohead]

Sure (http://www.anandtech.com/cpuchipsets/showdoc.aspx?i=2476)
July 21st :stick: it's not even M2 anymore..

[Syn.]

i think this might help this diccusion a bit.

Toms Hardware has a PREVIEW of the new AMD chip's. Benchmarks are to be released later today.

http://www.tgdaily.com/2006/02/20/toms_hardware_amd2_preview/

[Vapor]

....do they have a proc with a malfunctioning divider or a new one? Meh, guess we'll find out soon enough.

[biohead]

since when does Tom do AMD :confused:

[gundamit]

Does this mean other AM2 NDAs on the engineering samples are also over today? Would be nice to have a thread with all the links.

[mursaat]

....do they have a proc with a malfunctioning divider or a new one? Meh, guess we'll find out soon enough.since when does Tom do AMD :confused:I pressume they have a bad one because of the hurry to review it :D j/k... or not :rolleyes:

[Vapor]

Does this mean other AM2 NDAs on the engineering samples are also over today? Would be nice to have a thread with all the links.Or they managed to get one from somewhere that ignored the NDA and they themselves never actually signed one....like what Denny does :p:

[Carfax]

Benchmarks are up (http://www.tomshardware.com/2006/02/21/a_look_at_amds_socket_am2_platform/)

I'm pretty sure that they used an early sample which had the memory problems..

If not, AMD will be in a world of hurt when Conroe/Merom hit the deck.

[krille]

Benchmarks are up (http://www.tomshardware.com/2006/02/21/a_look_at_amds_socket_am2_platform/)

I'm pretty sure that they used an early sample which had the memory problems..

If not, AMD will be in a world of hurt when Conroe/Merom hit the deck.According to THG the story is The Socket AM2 generation is certainly not going to outperform comparable Socket 939 (DDR400) processors at DDR2-667 memory speeds-see our benchmarks. This is why AMD decided to wait for DDR2-800 and launch in June.

I think someone owns Lev a cookie!

[Vapor]

Considering it's Tom's and DDR2-667 4-4-4 vs. DDR-400 2-2-2, it's actually not bad. I expected worse from finalized retail chips. :shrug:

[Carfax]

Look what he says:

The processor we used for our tests was an Athlon 64 X2 4800+ for Socket AM2, which made its way to motherboard companies in mid-December 2005. It has the same technical characteristics as its Socket 939 brother: 2 x 1 MB L2 cache, and 2.4 GHz clock speed. However, while we could select DDR2-800 speed in the BIOS of the engineering test motherboard, it obviously did not set the memory faster than 667. As already mentioned, the memory is suspected to suffer from a performance bug, which may be true or simple a matter of the early product stage. The upcoming processor (revision F) will fix any of this; the dual core version is being shipped to motherboard partners starting this week, with single cores to follow in the middle of May.

Like I said, I'm pretty sure that chip is suffering from the memory bug.

[dinos22]

hmmm i thought AMD will have 65nm chips ready to go with AM2..........guess it's better this way when they both some out with big guns late this year......will be able to make up my mind then i guess heh

[biohead]

I smell cookies....

[agenda2005]

Well, This finally open up the can of worms that people don't want to see. In my opinion, without any architectural changes, even a full working DD2-800 will require good MB support that can run cas3 at DDR2-800 in order to show any significant improvement.
If things doesn't change significantly btw now and June 6, then Conroe at 2.66GHz for $531 will make FX-60 look like a bum.
I smell price crash here buddy.

[Sheik]

since when does Tom do AMD :confused:

Actually I think is a brilliant marketing move letting Toms of all sites do the first preview LOL :D

[[XC] Lead Head]

Seems like we're back in the sandbox guys, but who can help it? :toast:

01101001011100110010000001101001011101000010000001 10111101101011001000000111010001101111001000000010 00100110001001110010011000010110101101100101001000 10001000000110111001100100011000010010000001101001 01101110001000000110001001101001011011100110000101 11001001111001001111110010000001100010011101000111 01110010111000101110001011100010000001101001001000 00011011000110111101110110011001010010000001111001 01101111011101010111001000100000011010100110111101 10001000100000011101000110111101101111001011100010 11100010111000100000011101110110000101101110011011 10011000010010000001110011011010000110000101110010 01100101001111110010000000111010011101110110100101 10111001101011001110100010000000111010011101110110 10010110111001101011001110100000110100001010000011 01000010100111011101100001011100110010000001001001 00100000011000110110111101110010011100100110010101 10001101110100001000000110000101100010011011110111 01010111010000100000011011010111100100100000011001 01011000010111001001101100011010010110010101110010 00100000011100110111010001100001011101000110010101 10110101100101011011100111010000111111001000000111 01000110100001101111011100110110010100100000011011 10011101010110110101100010011001010111001001110011 00100000011000010111001001100101001000000110011001 10111101110010001000000110010001100100001100110011 11110010000001101111011100100010000001110111011000 01011100110010000001001001001000000110010101110110 01100101011011100010000001100011011011000110111101 1100110110010100111111


[Translated Binary]
is it ok to "brake" nda in binary? btw... i love your job too... wanna share? :wink: :wink:

was I correct about my earlier statement? those numbers are for dd3? or was I even close?[/Translated Binary]

[Sentential]

Well, This finally open up the can of worms that people don't want to see. In my opinion, without any architectural changes, even a full working DD2-800 will require good MB support that can run cas3 at DDR2-800 in order to show any significant improvement.
If things doesn't change significantly btw now and June 6, then Conroe at 2.66GHz for $531 will make FX-60 look like a bum.
I smell price crash here buddy.
No offense but you are assuming way too much. As I said *SPECIFICALLY* before this article came out, there is a bug with the dividers that is forcing the ram to obscenely low speeds (memtest showed pc2-3200) which is what is causing these issues. In time it will be fixed and M2 will be significantally faster than S939 and Conroe (on memory bandwith)

[Revv23]

yeah thats not a rev f cpu...

[agenda2005]

No offense but you are assuming way too much. As I said *SPECIFICALLY* before this article came out, there is a bug with the dividers that is forcing the ram to obscenely low speeds (memtest showed pc2-3200) which is what is causing these issues. In time it will be fixed and M2 will be significantally faster than S939 and Conroe (on memory bandwith)

Time will surely tell. Remember, memtest was programmed to read data for known CPU xteristics, so It mght need an update to read DDR2 speeds correctly from AMD new design. I'm not trying to justify the poor performance due to memory problem, but I just don't think that the extra bandwith from a full working DDR2-800 will make any significant increase in performance knowing quite well from DDR1 experience that AMD K8 CPUs are not bandwith starve in anyway. Infact K8 benefit more from low latency than high bandwith.

[nn_step]

Time will surely tell. Remember, memtest was programmed to read data for known CPU xteristics, so It mght need an update to read DDR2 speeds correctly from AMD new design. I'm not trying to justify the poor performance due to memory problem, but I just don't think that the extra bandwith from a full working DDR2-800 will make any significant increase in performance knowing quite well from DDR1 experience that AMD K8 CPUs are not bandwith starve in anyway. Infact K8 benefit more from low latency than high bandwith.
AMD CPU's never were really bandwidth starved with DDR.. Bandwidth isn't the issue it is Bandwidth/latency..

[Carfax]

Sentential may have the right of it I think..

Look at the synthetic memory scores. Those are far too low for DDR2-667.

DDR2-667 in dual channel mode has a max theoretical bandwidth of 10.8 GB/S, and with the efficiency of an ODMC, those synthetic scores should be much higher than what they are.

Instead, they seem to match the bandwidth given for DDR-400, which is only 6.4 GB/S.

Yeah, I must say that Sentential's comment about f*cked up memory dividers seems credible.

[[XC] leviathan18]

do i win my cookie????

[metro.cl]

Please, share! or you gonna pull an NN_step make a statmeant and not back it up?


i cant because of nda, i can only tell you its from 2 ram mannufacturers that i know this.

[DilTech]

Revision F isn't M2 dude...

You're confusing revision F and socket F.

Revision F is the next step in A64 chips, Revision E was the SD/Venice/etc. Revision F is what we'll see on AM2.

[onethreehill]

AMD newest road map confirmed AM2 supports DDR2-800
http://www.hkepc.com/bbs/news.php?tid=564763

[agenda2005]

AMD newest road map confirmed AM2 supports DDR2-800
http://www.hkepc.com/bbs/news.php?tid=564763

As I said before, even DDR2-800 will not give AM2 CPU any significant boost. If it does, AMD would be toutting those things with higher PR rating.This is not a good time for AMD.
Check my previous statement about Conroe at 2.66GHz vs FX-60. Yesterday benchmarks on Anandtech and Hexus obviously proved me right. GL AMD.