AMD FX-Series (Bulldozer) & 990FX Chipset (AM3+) Discussion

[Linkin]

Just overclock your CPU's CPU-NB. It required little tweaking and zero messing with the base clock speed (like all black edition chips)

My 955BE is a great chip at 4GHz/2.8GHz CPU-NB. It flies through everything I throw at it.

[zero_velocity]

Originally Posted by Linkin

Just overclock your CPU's CPU-NB. It required little tweaking and zero messing with the base clock speed (like all black edition chips)

My 955BE is a great chip at 4GHz/2.8GHz CPU-NB. It flies through everything I throw at it.

Yeah, a nice overclock benefits greatly from a NB OC

[stmok]

Originally Posted by itsmydamnation

ummm.... integer IPC is up by 15% over bulldozer clock for clock...... check again stmok . the power saving features (core parking) in things like mp3 encoding wont be taking effect, light threads are what suffered there. The effects of core parking by bulldozer have been quite thoroughly tested( there are even apps you can download and test/play with specifically targeting it) .

There are obvious additional structures to the front end in pile-driver vs bulldozer they also changed from soft edge to hard edge transistor flops. No they haven't touched anything architecturally for piledriver that happens in steamroller but they still have been very busy.

the architectural choices of bulldozer make sense when your target was throughput, but there target needed to be latency. a trace/uop cache could help, wider instruction decode and being able to decode from both threads in one cycle etc, more load/store bandwidth for FPU, fixing the L1I.

I talked to the author of that Tom's Hardware article on desktop Trinity. He also thinks it could be that the BIOS of the mobos he was using didn't have Cool'n'Quiet implemented properly. He's going to ask AMD about the origins of those performance improvements. We'll see what comes out of his enquiries.

[itsmydamnation]

Originally Posted by stmok

I talked to the author of that Tom's Hardware article on desktop Trinity. He also thinks it could be that the BIOS of the mobos he was using didn't have Cool'n'Quiet implemented properly. He's going to ask AMD about the origins of those performance improvements. We'll see what comes out of his enquiries.

i cant find an original source but this slide lists a whole lot

http://forums.anandtech.com/showthread.php?t=2242815

everyone initially blamed the 20-22 cycle L2 but it rarely seems to be the bottle neck. So its not surprising that improvements all round make a big difference ( having an actual working int divide must help at least a little ) . The other thing to remember is in the bulldozer to trinity test toms hardware did they did turned off all C and P states on both CPU's.

there is an awesome discussion on realworldtech right now about L1I and L1D and how modern caching relates to performance, People like Arron spink (former DEC and intel eng) go into some interesting detail.

[LostBenji]

It's a bloody shame that we can't use these things like we should be able to.

I used to be a AMD fan-boy but then swayed to the dark-side with a Q6600 after being burnt with a claimed X2 6400 BE that was supposed to have open multi's but DID NOT.

I have up until now had a string of s775 CPU's like the Q6600, Q9550 and a QX9650 (@4.0GHz daily) and they were good. I also have my workstation board (Supermicro X7DAe) with twin Xeon 5345's and I am dead-set tempted to put this 6 year old platform back in as it blows my new setup away.

FX8150 + GA-990FXA-UD5 + 8GB DDR-III 1600 + Oc to 4.5GHz stable

Sadly, it struggles to get over 40Gflops while getting too hot for my liking (A/M HSF and high-power server grade 120mm case fan making a 747 blush).
My old xeon smash it and they are only at 2.33Ghz on a 1333FSB with FB-DIMM..... Say no more.

Certainly steers me away from a socket G34 system with 62xx (Bulldozer) CPU's.

[Linkin]

Not even 40gflops? my current overclock nets 50-55

[NNNG]

Well, iirc, Bulldozer isn't' really designed for floating point workloads. If you try something that uses integers it should be a lot more competitive.

[itsmydamnation]

Originally Posted by NNNG

Well, iirc, Bulldozer isn't' really designed for floating point workloads. If you try something that uses integers it should be a lot more competitive.

yes it is, try something designed for bulldozer FP instruction sets

[Linkin]

Originally Posted by itsmydamnation

yes it is, try something designed for bulldozer FP instruction sets

See, the problem with this is that everyone is using things designed with intel instructions in mind and acting like it is fair, or just not thinking about it at all.

[Dave2972]

Originally Posted by synic

So, I've got a Phenomn 2 955 @ 3.2Ghz in an UD3-990FX AM3+ board. What would you think is a cheap upgrade / improvement on what I currently have? I want to use this system for at least several more years; don't game, but have noticed multimedia (1080p films) tearing when playing them on 27" monitor. I know I need to update my gfx card but am considering the CPU first.

I've noticed the FX-4170 @ 4.2Ghz and am wondering if it is a significant enough step up from my Phenonm 2 955 BE to warrant purchasing. I don't overclock so a default speed increase appeals to me.

I just don't see the need for 8 cores (or 4 modules) for what I use my puter for.

Thoughts?

Get better video drivers because if you have tearing while watching standard MPEG2 1080P video and XVID and DIVX MP4, etc videos then it is not the CPU but the video card drivers. My media box has an AM2 X2 5600 (2.8 GHz dual core, 1MB L2/core) and a 9800GT video card and it plays my videos without a problem, no tearing or any other problems unless I screwed up the transcoding via AVIDEMUX when using the wrong settings. My media box has less than half the CPU but a bit better video card than your machine so I can tell you that it not the CPU that is the problem.

I know somebody with an X2 6000 and a 7300GT video card (using windows XP/Win7-64) and it has zero, nought, nothing, SFA problems playing 1080P video, MPEG2 or better compressed. I know this because I built the thing and maintained and upgraded it over the years.

Since you think that a better CPU could cure the problems, the best thing to do is to overclock your CPU. I have an old C2 stepping 955 CPU and it worked fine at 3.6GHz using its stock voltage of 1.35V by just increasing the standard multiplier from 16 to 18, no other adjustments necessary. I am betting that you have a C3 stepping CPU and it should operate at 3.7-3.8GHz by a multiplier increase from 18.5 to 19 on stock volts. Try it out and see what happens because if it works then you do not have to buy a better CPU.

If, on the other hand, the CPU overclock did not cure the video then you have to look at the video drivers you have and then to the video card.

To repeat, upgrade your video drivers.

I almost forgot, hook your monitor up to a PVR/STB and see if it has tearing whilst watching 1080P video, if it does then your monitor is a piece of crap (or you screwed around to much with its settings {or bloody kids did} - restore to factory default).

[Nikon223]

What software are you using to measure GFLOPS?

[AEKaBeer]

The scores look fake to me but here's a LINK to a supposed ES Vishera and some benchmarks.

[remoteman213]

Originally Posted by AEKaBeer

The scores look fake to me but here's a LINK to a supposed ES Vishera and some benchmarks.

The year on the right of cpu says 2011 . Fake?

[.Radiant]

Originally Posted by remoteman213

The year on the right of cpu says 2011 . Fake?

could be an early ES. Either way it won't be indicitave of much if anything at all with that date.

[bobbth]

Originally Posted by .Radiant

could be an early ES. Either way it won't be indicitave of much if anything at all with that date.

As CPU-z isn't updated to support vishera it very well could be mis-reading the date/name and as this chip is probably in an un-supported motherboard (or a motherbaord still in development) the results probably reflect poorly on the expected performance of vishera. It could be real and it could be fake but regardless it's pretty irrelevant.