AMD Bulldozer review thread.

3) Everything that 3dnow could do, SSE did better and faster. Already K10 could do SSE twice faster than 3dnow, and if bulldozer would have supported 3dnow, the 3dnow performance would have been half of SSE performance.

Nope. 1-pass 3DNow reciprocals (both of them) haven't been beaten yet in precision. And subr (reversed two-operant subtract) has just been beaten with AVX, took quite some time - those lazy buggers. :p
 
:rolleyes:

Core 2 Quad and Phenom 2 are pretty neck and neck in terms of per-clock performance and BD is 10-15% slower per-clock then Phenom 2.... So get a clue tbh as it's not BS..

And my post was relevent to the point in hand..... You don't like it, Don't read it...simple.. :rolleyes:

It was not an invitation, and I will not utter it again. Either take the implied warning, or don't that's pretty simple, but rest assured that I wasn't joking about the low tolerance bit. Phenom II is hardly matching Yorkfield per clock, and it's sort of matched with Kentsfield except for when it isn't. BD tends to fall even under Phenom I per clock, quite frequently (not necessarily surprising, but we digress). I will not even endeavour to investigate the somewhat dense claim that it will, and I loosely quote, "equal but be much faster". It's quite similar to saying that someone died once but got better afterwards.

And I'm quite sure you are mistaken if you think that this is a rubbish bin in which nonsense can be thrown, and each should wade through it on his own accord. Sources of noise are frequently and efficiently culled. You are very close to being one, so either strive to move away from that particular basin of attraction, or don't, but B3D will not be drenched in the typical noise that I see oozing in other parts.

On the topic of "it would look good at X clocks", I'm noticing that that particular comparison tends to be made versus competitors at base clocks. Sandy Bridge tends to clock similarly to BD, with significantly better power profiles. Intel never pursued aggressive clocking with Yorkfield, as there was no competitive pressure to do so, but lather spins were very good clockers.
 
At a 4.5ghz base clock, BD would probably look good compared to Phenom II and C2Q, but overclockers are finding it to be stupidly hot. Like it makes Prescott look good. 4.5 ghz is the clock speed I think the thing needed to ship at but obviously it wasn't gonna happen with its power usage.

It would be rather pointless to upgrade from a 4ghz Yorkfield to it. Throwing money away.

I'm tempted just to buy a BD and see what I can get out it and do a comparison to my 4.8Ghz Phenom 2 x6....

More specifically to the CPU-NB scaling as a 3Ghz NB on a Phenom 2 can add upto 16% more performance at any given clock speed.

For example :

3Ghz Phenom 2 x6 + 3Ghz NB clock = 4Ghz Phenom 2 x6 + Stock 2Ghz NB

That's why when ever someone says that when overclock Phenom 2's are on par with socket 775 quads I laugh, People don't overclock the NB and I know that 99% of the Phenom 2 reviews don't either.

A 4ghz Phenom 2 x6 with a 3Ghz NB clock is a lot closer to i7 then reviews make it out to and is a far whack faster then a 775 Quad.

BD could be just as bandwidth starved as Thuban is, If not more so.
 
The only dedicated test of clocking the NB in Bulldozer I've seen, shows next to very marginal BW/latency scaling with the clock-rate. But Bulldozer already show some significant improvements in the throughput for both the L3 and the main memory, even at comparable clocks with Thuban. The problem here really is with the upper levels of caching.
 
More specifically to the CPU-NB scaling as a 3Ghz NB on a Phenom 2 can add upto 16% more performance at any given clock speed.

For example :

3Ghz Phenom 2 x6 + 3Ghz NB clock = 4Ghz Phenom 2 x6 + Stock 2Ghz NB

That's why when ever someone says that when overclock Phenom 2's are on par with socket 775 quads I laugh, People don't overclock the NB and I know that 99% of the Phenom 2 reviews don't either.

A 4ghz Phenom 2 x6 with a 3Ghz NB clock is a lot closer to i7 then reviews make it out to and is a far whack faster then a 775 Quad.

BD could be just as bandwidth starved as Thuban is, If not more so.

Assuming this is all true, I would really like to see it in benchmarks. I have never heard such things and I follow PC hardware pretty closely.

Also, what are the ramifications of OCing the northbridge? Massive power use? There has to be a good reason AMD didn't clock the northbridge that high in the first place if it produces such huge gains.
 
The CPU-NB is perhaps the most important thing when it comes to Phenom 2's.

Increasing it improves the through put of the entire chip and increases memory bandwidth.

There's really no downside to overclocking the north bridge, You do gain a little extra heat as you're adding more voltage to a different part of the chip that's not the cores them selfa but it's 3-4c at most.

It's well known in the overclocking world that when it comes to overclocking AMD chips that the CPU-NB is just as important as the CPU clock itself.

Athlon 2 and Phenom 2 x2/x3/x4's max out between 2.6-2.8Ghz NB clock but you do get the odd chip that can do 3Ghz. Phenom 2 x6's can all pretty much do GHz due to using a newer revision.

As for AMD leaving it at 2Ghz stock, Possibly due to stability reasons, Although it's a common and well shared opinion that AMD should of released Thuban with a NB clock higher then the stock 2Ghz.

Quick article with results can be read here

If you do a google search you'll find a few threads that back up all these results, I did my own testing with slides but they were deleted from my photobucket account long ago :(

Benefits from overclocking the CPU-NB/NB :

Higher memory Bandwidth
Reduced latency
Up to 16% IPC increase depending on application.

Downsides :

Require's voltage and thus produces slightly more heat

Phenom 2's are bandwidth starved at stock speeds, That's why they respond so well to the CPU-NB being raised.

I run my Phenom 2 x6 1075T at 4.8Ghz with a NB clock of 3.2Ghz as my 24/7 overclock and it flies and pretty much slaps my friends i7 960 around the place.
 
Last edited by a moderator:
is there any way they can use Radeon achitecture for X86 cpu
at least on core numbers

could you imagine ( Phenom X800 / X1400 ) ^^
 
Nope. 1-pass 3DNow reciprocals (both of them) haven't been beaten yet in precision. And subr (reversed two-operant subtract) has just been beaten with AVX, took quite some time - those lazy buggers. :p

oh, I forgot these.

But these reciprocals are mostly just for perspective correction of texture maps and normalization of geometry vectors, not used much be "generic fpu code".

theoretical addition and multiplication throughput with SSE code is twice compared to 3dnow in K10.
 
I'm just curious about the numbers.

I have a Sempron installed ATM, Waiting on my new RAM as well... 16Gb of RAM for £65 = Bargain!! And It would be even less once I recoup some money from my old RAM!

Here's my HWBot submission.... Link

I'm 9th fastest in the world but when you consider every overclock above mine is on DICE or LN2 I would be willing to bet a good chunk of money that mine is the fastest 1075T on the planet as unlike the other results 4.8Ghz is my 24/7 overclock ( Well it's actually 5013Mhz now after some tweaking ) and I know of no one else that runs a 1075T at that speed 24/7.

I'm so curious as to how BD scales with the cold that multiple times I've had an FX 6 core sitting in a checkout basket!
 
theoretical addition and multiplication throughput with SSE code is twice compared to 3dnow in K10.
Not if they run on the same execution units.

Horizontal adds were one of the best features in 3DNow, and it took until SSE3 until that was evened out.
 
The problem is that while SSE was adding (catching up) features over time, 3DNow! was entrenched in one place and AMD never bothered to do anything more on it for the last 10 or so years. Actually, AMD reduced the throughput of some arithmetic instructions in K7, as the x87 performance deficit was cleared out and thus the original purpose of 3DNow!. The extended 3DNow! set in K7 was in fact a subset of the SSE (integer op's), prepearing the ground for full SSE compatability in the next core revision. The ISA died with the K6 series and simply walked as a ghost around, until Bulldozer ditched it for good.
 
Back
Top