Barcelona Opteron 2350(B1) arrived

[mutambo]

So, the conclusion is we can fit barcelona's performance less than Quad Intels and greater than dual opterons/Xeons ??

OR

Yet to see the B2 stepping ?

some thing else ?

PS: No offense ...trying for a honest opinion.

[JohannesRS]

seems that B1 and BA are "almost" the same so.

My questioning would be: in which cases does it perform better, and which does it performs worst?

And: how much better or worst?

[mad_skills]

is B2 coming out?

[s7e9h3n]

Now CPU-Z can show NB clock, each core clock/multiplier, memory acccess ganged/unganged...

Great work, my friend! Ok these questions are for either you or Franck ...

One thing I'm a bit confused about - I see the CpuZ shots in this picture identify logical cores as an individual processors i.e., Processor #1 and Processor #2. Now what happens when you insert a second PHYSICAL cpu? Will CpuZ assign each core #5-8? Or will there be some way which it differentiates between physical cpu #1 and physical cpu #2?

Will CpuZ be capable of reading the Vcore on Barcelona? To this point, the only tool that I believe accurately depicts Vcore is AMD's power monitor. It'd be great if CpuZ could do the same.

[cpuz]

Hi S7e9h3n,
I'm confused as well. This should not report this way, but one processor with 4 cores / 4 threads, as on previous screenshots.
Kyosen is modifying MSRs, so this can make cpuz wrongly detect some things, this may explain. He will help me figuring out anyway.

AMD's power monitor reports the Core VID, I'm currently adding that in cpuz as well. But the real VCORE is always better, and in this case cpuz did not detect the sensor chip.

I'm also preparing a small tool to change multiplier on the fly for each core. Kyosen already does it, but this is not very convenient. A slider will make things much easier for him.

[s7e9h3n]

Hi S7e9h3n,
I'm confused as well. This should not report this way, but one processor with 4 cores / 4 threads, as on previous screenshots.
Kyosen is modifying MSRs, so this can make cpuz wrongly detect some things, this may explain. He will help me figuring out anyway.

AMD's power monitor reports the Core VID, I'm currently adding that in cpuz as well. But the real VCORE is always better, and in this case cpuz did not detect the sensor chip.

I'm also preparing a small tool to change multiplier on the fly for each core. Kyosen already does it, but this is not very convenient. A slider will make things much easier for him.

OK...thanks for clearing that up.

BTW, if you need some data/dumps from a system running 2 x Barcelonas, feel free to send me a PM (of course that's dependent on when my system comes back from repair )

[Lightman]

I presume multiplier is upwards locked in new Barcelona's?
Also is it possible to set vcore higher than 1.2V using C'n'Q or this is maximum allowed?

What is maximum vcore for northbridge?

Can you use half multipliers on Barcelona? Phenom 9600 is running 11.5x so it looks like Barcelona should be capable of doing same. If yes will we see any performance drop from using .5x as on K8?? (mainly lower memory bandwidth).


Lots of questions for our Guru's!

[cpuz]

I presume multiplier is upwards locked in new Barcelona's?

Yes

Also is it possible to set vcore higher than 1.2V using C'n'Q or this is maximum allowed?

Max VID is 1.55V.

What is maximum vcore for northbridge?

This depends on dual-plane or single-plane configuration, and I still miss the register values to tell. On Kyosen system, the NB VID is 1.2V on the B1 and 1.15V on the BA.

Can you use half multipliers on Barcelona?

Yes.

If yes will we see any performance drop from using .5x as on K8?? (mainly lower memory bandwidth).

I don't know. On K10, memory clock is computed from bus speed, and it therefore not dependant from CPU speed anymore, so this is tempting to say "no", but only a real test can tell.

[s7e9h3n]

This depends on dual-plane or single-plane configuration, and I still miss the register values to tell. On Kyosen system, the NB VID is 1.2V on the B1 and 1.15V on the BA.

Is it possible to figure out the value by looking at which register Amd's Power Now tool is accessing? Also, I think the B1 vs. B2 NB VID's are different because Kyosen's cpu's are HE's

[Lightman]

Yes


Max VID is 1.55V.


This depends on dual-plane or single-plane configuration, and I still miss the register values to tell. On Kyosen system, the NB VID is 1.2V on the B1 and 1.15V on the BA.


Yes.


I don't know. On K10, memory clock is computed from bus speed, and it therefore not dependant from CPU speed anymore, so this is tempting to say "no", but only a real test can tell.

Thank you for your answers Franck!

With 0.5x multipliers I'm also tempted to say it will not make any difference, but ....

Regarding vcore is it then possible to set 1.55V from Windows?? You know that X2 are hard locked between 1.30 and 1.45V depending on model and version....

BTW have you all guys noted AMD's official pricing for Phenoms??
It is what I was waiting for .

[informal]

.

BTW have you all guys noted AMD's official pricing for Phenoms??
It is what I was waiting for .

Pricing is very competitive

[mnemonix]

BTW have you all guys noted AMD's official pricing for Phenoms??
It is what I was waiting for .

source?

[zornundo]

Pricing is very competitive

dood, you know we're lazy! where's the linkage?

[informal]

dood, you know we're lazy! where's the linkage?

Here's your "linkage":
http://www.xtremesystems.org/forums/...93&postcount=1

[linhvndiy]

I just got Tyan S2915-E. Seem oc a little bit higher. Now i can run 225 mhz htt.. But can not reach higher than 230mhz. Seem mainboard need a lot of tweak. Barcelona 2.3ghz is coming, i saw preorder on web now.
Very good news.
And my client confirm One barce 1.9ghz faster than Two opteron 280 in web & streaming servive in their datacenter.
And desktop software do not gain much and using full barcelona power. Becos i saw no gain when using 3dmax9 render(single chanel vs numa enable).

[cpuz]

Is it possible to figure out the value by looking at which register Amd's Power Now tool is accessing? Also, I think the B1 vs. B2 NB VID's are different because Kyosen's cpu's are HE's

I miss the value of the register, not the register.
This register is not "standard" and the generic dumps do not show it. It needs a special access. The next test will answer to that question (max NB VID)

[MAS]

as far as i can understand according to kyosen's SS:

[IMC clock]:[phys. memory clock] == 27:5 when IMC clock == 1800 and ddr2 667 installed
and
[IMC clock]:[phys. memory clock] == 24:5 when IMC clock == 1600 and ddr2 667 installed, etc.

and respectively

[L3 cache clock]:[phys. memory clock] == 27:5 when IMC clock == 1800 and ddr2 667 installed
and
[L3 cache clock]:[phys. memory clock] == 24:5 when IMC clock == 1600 and ddr2 667 installed, etc.


right?

[kyosen]

SuperPI4M time with different NB clock

2350(B1) @1.8G=200x9, NB@1.8G=200x9: 3m47.453s
http://www.oohashi.jp/c-board/file/S..._NB1.8_UnG.png

2350(B1) @1.8G=200x9, NB@1.6G=200x8: 3m47.359s
http://www.oohashi.jp/c-board/file/S..._NB1.6_UnG.png

Almost same time...difference is within fluctuation, I think.
NB@1.6G may be enough for single thread program with dual DDR2-667 access!?
or I suspect my NB multiplier changing method is incomplete
I have no detail info about L3 clock and its behaviour...


3DMark06 score with different HT Link clock

2350(B1)@2.0G, GeForce8800GT@default clock, HT Link@1000MHz=200x5: 9418
http://www.oohashi.jp/c-board/file/3..._HTx5_88GT.png

2350(B1)@2.4G, GeFoce8800GT@default clock, HT Link@962.8MHz=240.7x4: 10926
http://www.oohashi.jp/c-board/file/3..._HTx4_88GT.png

We can observe slight improvement compared to previous results, 9353 and 10796:
http://www.oohashi.jp/c-board/file/3....0G-200x10.png
http://www.oohashi.jp/c-board/file/3....4G-240x10.png


K10 support in MchbarEdit and MemSet
FELIX provided me beta programs, both are working fine

Now we can change NorthBridge & HT Link multiplier with MchbarEdit,
and memory timing with MemSet, on Windows...both 32bit and 64bit
(Good old WPCREDIT has no support for Win x64)
Screenshots are located on my BBS:
http://www.oohashi.jp/c-board/c-boar...ne;no=5239;id=


CineBench10 32bit/64bit

32bit, 2346(BA)@1.8G, NB@1.6G, DDR2-667: 1483, 5731, 3.86 x
http://www.oohashi.jp/c-board/file/C..._NB1.6_333.png

32bit, 2346(BA)@1.8G, NB@1.6G, DDR2-800: 1487, 5730, 3.85 x
http://www.oohashi.jp/c-board/file/C..._NB1.6_400.png

32bit, 2350(B1)@1.8G, NB@1.6G, DDR2-667: 1484, 5743, 3.87 x
http://www.oohashi.jp/c-board/file/C..._NB1.6_333.png

32bit, 2350(B1)@1.8G, NB@1.6G, DDR2-800: 1493, 5736, 3.84 x
http://www.oohashi.jp/c-board/file/C..._NB1.6_400.png

32bit, 2350(B1)@2.4G, DDR2-640: 1973, 7654, 3.88 x
http://www.oohashi.jp/c-board/file/C..._NB1.8_320.png

32bit, 2350(B1)@2.4G, DDR2-800: 1977, 7672, 3.88 x
http://www.oohashi.jp/c-board/file/C..._NB1.8_400.png

64bit, 2346(BA)@1.8G, NB@1.6G, DDR2-667: 1793, 7315, 4.08 x
http://www.oohashi.jp/c-board/file/C..._NB1.6_333.png

64bit, 2350(B1)@1.8G, NB@1.6G, DDR2-667: 1789, 7327, 4.09 x
http://www.oohashi.jp/c-board/file/C..._NB1.6_333.png

64bit, 2350(B1)@2.4G, DDR2-800: 2402, 9642, 4.01 x
http://www.oohashi.jp/c-board/file/C....4_x10_400.png

64bit, 2350(B1)@2.3G=230x10, DDR2-768: 2279, 9191, 4.03 x
http://www.oohashi.jp/c-board/file/C....3_x10_384.png

64bit, 2350(B1)@2.3G=242x9.5,DDR2-808: 2306, 9239, 4.01 x
http://www.oohashi.jp/c-board/file/C...3_x9.5_404.png

Summary from above:
* B1 and BA show almost same performance.
* DDR2-667 or 800 makes almost no difference.
* 32bit -> 64bit score gain is about 20%.
* K10 has no performance degradetion at half(*.5) multiplier.

BTW, Franck provides me CPU-Z latest beta
Now we can see K10 VID on CPU-Z(for example, in the 64bit, 2346(BA)@1.8G case above),
and also can choose a clock showing core#


I intend to test a rendering software, Shade...familiar in Japan,
and then, I'll try voltage modification

[STEvil]

thanks for all the work kyosen, franck, felix, and others

[Eternalightwith]

Summary from above:
* B1 and BA show almost same performance.
* DDR2-667 or 800 makes almost no difference.
* 32bit -> 64bit score gain is about 20%.
* K10 has no performance degradetion at half(*.5) multiplier.

I intend to test a rendering software, Shade...familiar in Japan,
and then, I'll try voltage modification

Why do you think that there is a 20% improvement from 32bit to 64bit?

[MAS]

Why do you think that there is a 20% improvement from 32bit to 64bit?

actually depending on the task

20 per cent is an average mean


kyosen, and where's 64bit 3dmark?

[macci]

Almost same time...difference is within fluctuation, I think.
NB@1.6G may be enough for single thread program with dual DDR2-667 access!?
or I suspect my NB multiplier changing method is incomplete

correct, this (NB FID) can not be changed on the fly (the register can be adjusted but NB clock doesnt actually change). a BIOS item will allow NB multiplier (NB FID) adjustment.

[MAS]

correct, this (NB FID) can not be changed on the fly

are you sure? then how CnQ will work with IMC?

CnQ lowers both CPU and IMC multi in idle, on the fly

[kyosen]

correct, this (NB FID) can not be changed on the fly (the register can be adjusted but NB clock doesnt actually change). a BIOS item will allow NB multiplier (NB FID) adjustment.

Thanks for your comment
OK, I read the document again........yeah, it looks that I need 2 more steps, i.e.
1. write register NbFidEn=1
2. then, reboot.
I'll try it tonight!

And MAS, well, I'll also try 3DMark on WinXP x64

[KTE]

Thanks to all guys, and kyosen.

Kysoen: Are you sure there is almost no difference with different DDR2 clocks in CB10? It should be around 100 CB between 667 and 800.