AMD Shanghai/Deneb Review Thread

[Ghostbuster]

C2 45nm jumped a lot in Cinebench... thanks to the caches.

The main contributor was the Fast Radix-16 divider which speeds up its FPU operations.

And yes, I'm suspecting what you said. That score was 32 while the Tech Report always benches in 64.

EDIT: Still seems low. Way too low.

http://www.techwarelabs.com/reviews/.../index_5.shtml
http://www.sharkyextreme.com/hardwar...261_3756681__5

Those are 32-bit ones..

We know clock-for-clock that these Denebs are REALLY close to the Xeons in single core rendering... so wtf.
Tech Report 2.7Ghz Opteron Cinebench x64: Single core 3156

If we take this score, scale clock linearly to 3Ghz, times 3.9 (Phenom I speedup), we get 13676. The 32 bit scores are WAY too low.

Nope, you can't scale them up like that. Shanghai scaling probably a bit different in CB10 and this may also apply to Deneb.

TR shows 18616 for 8 cores. Thus speedup is (18616 / 3156 ) = 5.90

Thus scaling per core becomes (5.90 / 8) = 73.7%

Speedup for 4 cores is (73.7% * 4) = 2.95

Thus per core CB10 score hypothethically is (9746 / 2.95) = 3303

Hypothethically speaking, a 3GHz Phenom II scores 3303 per core.

For some odd reason, Shanghai/Deneb doesn't scale well in CB10.. unlike Phenom I...

[flippin_waffles]

Thanks iocedmyself! Looks awesome, and I bet you're having a blast.

[charged3800z24]

The main contributor was the Fast Radix-16 divider which speeds up its FPU operations.

Those are 32-bit ones..

Nope, you can't scale them up like that. Shanghai scaling probably a bit different in CB10 and this may also apply to Deneb.

TR shows 18616 for 8 cores. Thus speedup is (18616 / 3156 ) = 5.90

Thus scaling per core becomes (5.90 / 8) = 73.7%

Speedup for 4 cores is (73.7% * 4) = 2.95

Thus per core CB10 score hypothethically is (9746 / 2.95) = 3303

Hypothethically speaking, a 2.7GHz Phenom II scores 3303 per core.

For some odd reason, Shanghai/Deneb doesn't scale well in CB10.. unlike Phenom I...

But isn't CB10 memory sensitive?...nt so much on AMD's NB side but RAM timings and ram speed seem to show better on my runs...which would slow Shanghai down compared to Deneb

[doompc]

C2 45nm jumped a lot in Cinebench... thanks to the caches.

The main contributor was the Fast Radix-16 divider which speeds up its FPU operations.

And SSE4 support.

[Ghostbuster]

Multicore rendertime was 1minute 1 second.

Within striking range of QX9770... (about -3.6% per core, -1% multicore)

But isn't CB10 memory sensitive?...nt so much on AMD's NB side but RAM timings and ram speed seem to show better on my runs...which would slow Shanghai down compared to Deneb

I suspect the OCW scores lowly due to memory? Their 3DVantage score was a little bit low..

[Zucker2k]

Within striking range of QX9770... (about -3.6% per core, -1% multicore)

I suspect the OCW scores lowly due to memory? Their 3DVantage score was a little bit low..

Not bad at all.

[charged3800z24]

Multicore rendertime was 1minute 1 second.

for a lil comparison..here is a 3.3ghz run I just did on Win XP 64bit this I think, show much improvement...

[iocedmyself]

yeah i was speachless when my uncle called and asked me the model number of my motherboard so he could make sure to get a speed matched Deneb, and after i discovered the bent pins due to playful cat i sat pretty much unblinking trying not cry...so i'm still just happy it didn't get FUBARed 36 hours after i got it.

The only thing i was warned about was the possibility that i shouldn't post CPU-Z or overdrive screens with visible voltages when referencing clock speeds, reporting voltages is ok, Verifying not sure, so until i get verification just gonna avoid potential problem.

The thing clocks high and easy on suprisingly low volts, and is drastically cooler then my 9850 was with same DuoOrb heat sink, I've taken the voltage up to 1.6 and still haven't seen it go above 50c under sustained load despite the fact that i just used the white thermal grease that came with the HS instead of the usual AS5. In 24C ambient it's idling around 28-29c @1.35v clocked to 3.7, load temps are 35-36c with all 4 cores @100%.

As much as i love the motherboard, i suspect the 4+1 phase power is a bit hindering to stability, i was running my 9850 in this board for a week or so before i got the deneb and while i was able to crank it up an extra 350mhz over what i could on the DFI 790FX/SB600 board, Using this asus board even running the 3.3ghz DFI rock solid speed was somewhat flakey. So im going to pick up on 790FX to run this chip in, have always relied on DFI for uber overclocking but would be grateful for any recomedations from anyone that's had first hand experience with more then one MFG.

Also, i must admit that i'm kind of at a loss with the ACC function, i've more or less been randomly setting the % either plus or minus, so any insight that could be shed on its functionality would be appreciated.

So just ran prime, and too lazy to take a screeny and btw @ 3ghz SuperPi runs 1mb in 23secs


AMD Phenom(tm) II X4 940 Processor
CPU speed: 3007.47 MHz, 4 cores
CPU features: RDTSC, CMOV, Prefetch, 3DNow!, MMX, SSE, SSE2
L1 cache size: 64 KB
L2 cache size: 512 KB
L1 cache line size: 64 bytes
L2 cache line size: 64 bytes
L1 TLBS: 48
L2 TLBS: 512
Prime95 32-bit version 25.6, RdtscTiming=1
Best time for 768K FFT length: 15.087 ms.
Best time for 896K FFT length: 18.044 ms.
Best time for 1024K FFT length: 20.564 ms.
Best time for 1280K FFT length: 25.551 ms.
Best time for 1536K FFT length: 31.511 ms.
Best time for 1792K FFT length: 37.425 ms.
Best time for 2048K FFT length: 42.566 ms.
Best time for 2560K FFT length: 55.744 ms.
Best time for 3072K FFT length: 68.293 ms.
Best time for 3584K FFT length: 81.211 ms.
Best time for 4096K FFT length: 92.011 ms.
Best time for 5120K FFT length: 124.520 ms.
Best time for 6144K FFT length: 150.859 ms.
Best time for 7168K FFT length: 182.677 ms.
Best time for 8192K FFT length: 210.213 ms.
Timing FFTs using 2 threads.
Best time for 768K FFT length: 10.022 ms.
Best time for 896K FFT length: 12.980 ms.
Best time for 1024K FFT length: 14.818 ms.
Best time for 1280K FFT length: 18.493 ms.
Best time for 1536K FFT length: 22.310 ms.
Best time for 1792K FFT length: 26.568 ms.
Best time for 2048K FFT length: 30.100 ms.
Best time for 2560K FFT length: 40.088 ms.
Best time for 3072K FFT length: 47.670 ms.
Best time for 3584K FFT length: 56.505 ms.
Best time for 4096K FFT length: 63.814 ms.
Best time for 5120K FFT length: 81.174 ms.
Best time for 6144K FFT length: 100.351 ms.
Best time for 7168K FFT length: 121.644 ms.
Best time for 8192K FFT length: 138.646 ms.
Timing FFTs using 3 threads.
Best time for 768K FFT length: 8.209 ms.
Best time for 896K FFT length: 13.247 ms.
Best time for 1024K FFT length: 14.495 ms.
Best time for 1280K FFT length: 17.185 ms.
Best time for 1536K FFT length: 19.838 ms.
Best time for 1792K FFT length: 23.091 ms.
Best time for 2048K FFT length: 25.199 ms.
Best time for 2560K FFT length: 37.169 ms.
Best time for 3072K FFT length: 42.726 ms.
Best time for 3584K FFT length: 47.814 ms.
Best time for 4096K FFT length: 53.474 ms.
Best time for 5120K FFT length: 62.424 ms.
Best time for 6144K FFT length: 77.315 ms.
Best time for 7168K FFT length: 93.946 ms.
Best time for 8192K FFT length: 106.442 ms.
Timing FFTs using 4 threads.
Best time for 768K FFT length: 7.475 ms.
Best time for 896K FFT length: 12.338 ms.
Best time for 1024K FFT length: 13.448 ms.
Best time for 1280K FFT length: 15.917 ms.
Best time for 1536K FFT length: 18.593 ms.
Best time for 1792K FFT length: 21.290 ms.
Best time for 2048K FFT length: 23.759 ms.
Best time for 2560K FFT length: 34.387 ms.
Best time for 3072K FFT length: 39.402 ms.
Best time for 3584K FFT length: 44.715 ms.
Best time for 4096K FFT length: 49.535 ms.
Best time for 5120K FFT length: 56.587 ms.
Best time for 6144K FFT length: 66.567 ms.
Best time for 7168K FFT length: 80.962 ms.
Best time for 8192K FFT length: 90.527 ms.
Best time for 58 bit trial factors: 3.960 ms.
Best time for 59 bit trial factors: 3.953 ms.
Best time for 60 bit trial factors: 3.937 ms.
Best time for 61 bit trial factors: 3.953 ms.
Best time for 62 bit trial factors: 7.273 ms.
Best time for 63 bit trial factors: 7.265 ms.
Best time for 64 bit trial factors: 7.261 ms.
Best time for 65 bit trial factors: 7.233 ms.
Best time for 66 bit trial factors: 7.212 ms.
Best time for 67 bit trial factors: 7.181 ms.

[iocedmyself]

At 3.4ghz calculates in 19s =)

[Caveman787]

I just wanna know if it'll perform as good or better in gaming then core2, yorkfield and i7. Both single card, sli and crossfire.

Thats the main issue for most more mainstream people.

iocedmyself....I don't want to doubt you but those temps seem unreal only 4 or 5 degrees over ambient deneb must have some sort of cool n quiet that lowers core clock, voltages and core activation to manage that. That's all I can figure...unless one of the pins for that isnt functioning right.

Also what did you use to stress it?
Maybe that's why temps are so low.

Otherwise deneb looks to be amazing so far.

[charged3800z24]

So im going to pick up on 790FX to run this chip in, have always relied on DFI for uber overclocking but would be grateful for any recomedations from anyone that's had first hand experience with more then one MFG.

I use the asus M3A79-T Deluxe with 8+2 Phase...I and others here have a lot of fun on this board..790FX/SB750 +ACC it is a good board in my book..solid and stable..lots of options..

[flippin_waffles]

seems to be a pretty good improvement there in SPi from PHi to PHii.

[Macadamia]

iocedmyself, can you try raising NB/L3 speed and Cinebenching again?


Justapost says it doesn't matter for Phenom I, but that might be because PhI already scaled at 3.9x at default. We have 0.2x here to catch up.

[donitsi]

After a small scare in which my hardware loving cat (black of course) knocked the chip as it sat in it's blue factory carrier on the ground knocking a fair deal of pins out of whack... got them straightened out last night.

Keep your :banana::banana::banana::banana::banana: away from your HW

THANKS for sharing this PII info with us

[justapost]

@iocedmyself, thanks for the info's. Early shanghai (p)reviews showed enhancements in sandra's intercore test. Is it possible you tell us the numbers you get for intercore bandwidth and intercore latency? Also interesting would be the L3 latency sandra reports.

My 9950 results (mem unganged, winxp 32bit)
CPU 2.6GHZ NB 2.0GHz Mem 533 5-5-5-15-26: intercore bandwidth 2,99 latency 140 / L3 latency 54
CPU 3.2GHZ NB 2.0GHz Mem 533 5-5-5-15-26: intercore bandwidth 2,94 latency 135 / L3 latency 64
CPU 3.2GHZ NB 2.4GHz Mem 533 5-5-5-15-26: intercore bandwidth 3,43 latency 126 / L3 latency 56

(Sandra calculates the latency with the cpu frequency that's why the latency seems to get bigger with scaling cpu frequency).

BTW: You know if you select 1.5V vcore in the bios some apps will only show a VID of 1.3V and not the +0.2V voltage offset?

I posted a few recommendations and results with ACC here. Hope it helps but it's timeconsuming.

Update: Just ran the benches at CPU 3GHz NB 2GHz under Vista 64bit

CPU 3.0GHZ NB 2.0GHz Mem 533 5-5-5-15-26: intercore bandwidth 3,17GB/s latency 142ns / L3 latency 60

[Chad Boga]

After a small scare in which my hardware loving cat (black of course) knocked the chip.

You should put up some pictures of this monster.

[iocedmyself]

I will play with the NB speed and bench the diff in a bit when i wake up as well as post the pics i took before my :banana::banana::banana::banana::banana: came and scared my hardware

Also as far as the temps go i was under the same impression of the temps displaying wrong but after checking the Bios, overdrive, Asus PC probe, thermal sensor from fan control and Infrared Digital thermometer temps and having them be within 1c of one another it runs about 5C above ambient with the normal consistaint voltage and clock speed, C&Q drops the voltage to .8v and clock to 800mhz.

[justapost]

Cool, on the event screenshot from the internal tool, p-state-3 uses 800MHz and 0.8V. Seems denebs come with four usable p-state's instead of two on pI's.

[Chad Boga]

I will play with the NB speed and bench the diff in a bit when i wake up as well as post the pics i took before my :banana::banana::banana::banana::banana: came and scared my hardware

I was wanting pics of your kitty, not hardware.

[Macadamia]

Very nice C&Q. I just hope this time it works nicely (unlike the 9XXXe Phenoms :x )

[Caveman787]

wow.... the idle power consumption should be so low that it should be off the charts.

Lower power consumption and TDP then core i7, cool and quiet puts it at lower clocks add in an undervolt and you've got the ultimate low energy power chip.

[Macadamia]

I don't think idle will win i7 though. They have per-core power gating. D:

Normal 10-25% usage however... should be different.

[iocedmyself]

there is blurry pic of kitty inspecting the the shiny spiny thing under its curious a$$, i'm sad to discover the pics i took of the chip in the blue carrier are lacking detail enough to show what's printed on chip so i'll have to pull the heatsink later and take new shots.

But back to bussiness at hand, i've been trying to get the HT speed to boot at 2.6ghz with stock cpu freq, but the bios only lets you specify 2.2ghz and i'm not having much luck playing with the multi's. Still learning this board personality so, again advice or suggestions would be welcome.

USing overdrive i can take the FSB up to 250x10 to up the HT speed, but i ran CB10 with HT link at 2.2 which uped cpu to 3.3ghz and gave me 3819 single core, 13450 about for multicore which showed scaling to be 3.52. So i'm pleased at least to see that the FSB is a little more pliable than with the phenom 1 chips.

C&Q on the 45nm chips is really much improved, but full features aren't enabled yet, you'll actually be able to specify clock speeds you want to scale down to on the AM3 boards. As well as specify individual clock and power per/core i believe. I'm not sure if i can specify just how low the voltage and TDP has been tested at, but it really does put I7 to shame, for example have 1.0 Vcore @ 2.8 rock solid stable

Just wanted to give update i'm going to reinstall sandra and test interconct bandwidth and cache speeds now. So again any advice suggestions or even requests lemme know

there, he sprawled back on the case again so snapped a better pic =D

[donitsi]

Cute little kitty you got there
edit: looks a bit scary on the sec. pic though

[Donnie27]

That's why NB/L3 runs on separate power plane??
You don't need crazy Vdimm volts in order to OC PhII.You have memory dividers and you use multies to OC the CPU,so the ram stays at the same clock all the time.I thought you knew this already ?
All we know is that PhII ran at 1.9V .That's 1.9V for vcore.Uncore part(NB/L3) ran at much lower volts and clocks using constant NB multi and HTT of 200Mhz.

I know about Phenom II, my comments were about folks being them up on Phenom 1 ANd that they WOULDN'T have to resort to such stuff. If you don't know what I'm talking about, ask Rammsteiner? 1.9v is a lot no matter what processor it's used on IMHO. Short term is one thing, think folks will be able to use 1.9v without LN, I don't think so. Yes, I know what an Un-Core is and even Intel is using that term now as well