What Were They Thinking

Zarathustra[H];1037873039 said:
Agreed, unless they are harvesting the best cores for Opterons or something like that.

I wonder what the Bulldozer based Opteron figures look like.

Businesses are very serious about reducing the power usage in their server farms, as are those who build super computers, but as I understand sales have been great to both of these customer groups...
Click to expand...

Interesting

According to this the 16 core Opteron 6282 SE at 2.6Ghz has a TDP of 140W.

The 2.1-2.3Ghz models have a TDP of 115W, and the 1.6Ghz models 16 cores only has a TDP of 85W.

Now - granted - while these are 16 cores, they are clocked A LOT lower, but server chips usually are.

They also use the identical silicon as the FX chips, just binned differently.

My theory is as follows:

1.) We know bulldozers power usage increases faster with clock than other CPU architectures, based on the ridiculous load numbers for OC:ed FX chips, and by the fact that these 16 core models can get as low as 85W TDP.

2.) We also know that binning is commonplace in the CPU industry.

I think they are binning the best chips (the ones that will run stable with less voltage) for Opteron production, and the others for FX production.

Server markets are OK as they only use low clocks and thus are not hurt by bulldozers high power use at high clocks.

FX suffer from BD's high power use at high clocks.

High power use at high clocks, or needing excessively high voltage at higher clocks is usually a sign of process immaturity.

This is strengthening my theories that once GloFo's process improves BD clock speeds will go up rather quickly. This also speaks to why it is as delayed as it is, and also why there were rumors over the summer of it not clocking high enough for AMD's satisfaction.
 
eastmen said:
gotta say something is fishy there.

at idle its 110/126 and at load its suddenly 244 and 586 ? something isn't right at all.

He said they are only using prime so its the cpu that is stressed in this. So the cpu is already using 134w more non overclocked , If we compare the over clock its usings 460 watts more than idle ?

Can the motherboard even supply 460w of power to the cpu ?


looking at kyle they have the fx 8150@ 4.6ghz using 452 watts vs 120 idle. That means the cpu overclocked is only pulling 332w more. An extra 200mhz is supposed to cause this thing to pull another 130w of power ? something has got to be wrong there
Click to expand...

Anandtech's review shows similar results in the power consumption tests, though not as extreme as the [H] review. I think the power consumption during load is what disappoints me most.
 
Adam12176 said:
Now that I say it out loud though, for small SMB's/Mid size businesses I'm wondering if an 8C 125W BD part would be competitive with something like an E7-2850 10C/20T at 2ghz. It's the same TDP, so I guess it all comes down to how well ESX/HyperV/Xen can use BD cores.
Click to expand...

Except Opterons at low clocks have MUCH better TDP's than the Desktop/FX BD's.
Code: 
Model		Cores	Frequency	TDP		Pre-order price
Opteron 6204	4	3.3 GHz		115 Watt	$516.13
Opteron 6212	8	2.8 GHz		115 Watt	$303.17
Opteron 6220	8	3.0 GHz		115 Watt	$588.93
Opteron 6234	12	2.3 GHz		115 Watt	$430.00
Opteron 6238	12	2.5 GHz		115 Watt	$516.13
Opteron 6262 HE	16	1.6 GHz		85 Watt		$588.93
Opteron 6272	16	2.1 GHz		115 Watt	$588.93
Opteron 6274	16	2.2 GHz		115 Watt	$720.17
Opteron 6276	16	2.3 GHz		115 Watt	$881.22
Opteron 6282 SE	16	2.6 GHz		140 Watt	$1135.26

The problem with BD seems to be that clocks don't scale well with power. Typical problem on immature processes and with large dies. BD has both.
 
Zarathustra[H];1037873136 said:
Interesting

According to this the 16 core Opteron 6282 SE at 2.6Ghz has a TDP of 140W.

The 2.1-2.3Ghz models have a TDP of 115W, and the 1.6Ghz models 16 cores only has a TDP of 85W.

Now - granted - while these are 16 cores, they are clocked A LOT lower, but server chips usually are.

They also use the identical silicon as the FX chips, just binned differently.

My theory is as follows:

1.) We know bulldozers power usage increases faster with clock than other CPU architectures, based on the ridiculous load numbers for OC:ed FX chips, and by the fact that these 16 core models can get as low as 85W TDP.

2.) We also know that binning is commonplace in the CPU industry.

I think they are binning the best chips (the ones that will run stable with less voltage) for Opteron production, and the others for FX production.

Server markets are OK as they only use low clocks and thus are not hurt by bulldozers high power use at high clocks.

FX suffer from BD's high power use at high clocks.

High power use at high clocks, or needing excessively high voltage at higher clocks is usually a sign of process immaturity.

This is strengthening my theories that once GloFo's process improves BD clock speeds will go up rather quickly. This also speaks to why it is as delayed as it is, and also why there were rumors over the summer of it not clocking high enough for AMD's satisfaction.
Click to expand...

I fully agree with all of this.
 
Athlon64 (HAMMER!) was the glory days and top point of AMD's Triumph. Athlon XP was just a precursor to that.
 
I'd call that 6282SE as a wash to slightly slower to the current E7 referenced above. Given BD's need to clock higher to match current offerings anyway.

Believe me I am a firm supporter of competition in the marketplace. I want to find a market for this thing to excel in. It is clearly not the consumer space however.

As much as I want to add "in this iteration"to that sentence; I'm reminded of the rampant delays on the first one. Believing that AMD will adopt a more rapid approach to future iterations would require a large dose of faith - and I have no idea how anyone could have that on this day.
 
Adam12176 said:
As much as I want to add "in this iteration"to that sentence; I'm reminded of the rampant delays on the first one. Believing that AMD will adopt a more rapid approach to future iterations would require a large dose of faith - and I have no idea how anyone could have that on this day.
Click to expand...

Agreed. If they execute to plan (new arch every year for the next 3 years) and release piledriver, excavator, etc. this could lead to something good, but its tough to have any confidence in that right now.

That being said, here's something to put this in perspective:

In early 2009 with cherry-picked parts the absolute highest anyone could get Phenom II x4's to clock on extreme cooling was 4.2Ghz.

Fast forward ~2 years and the production process matured so much that the same parts - as of last December were hitting 4.2ghz on air.

If GloFo's 32nm process matures the same way, we could be looking at 8.5ghz on air out of this architecture by 2013. Sprinkle in a 10% IPC gain once a year with Piledriver, excavator and whatever the other names were, and it really looks like it has potential.

Granted, by then Intel will likely be on the virge of introducing the successor to 22nm, so its tough to say which way this will go. I think it all depends on how quickly process improvements are made. AMD shifting their strategy to only pay for good parts (and hopefully pay more for better parts) should give GloFo some real incentive to get up to better yields fast.
 
can't say this was unexpected.

In all honesty, there was a part of me that was hoping all this silence was to troll the hell out of intel, realistically though, if they had a killer product they would have been slipping that shit all over the place to build more hype. when the official hype train went silent...i had a feeling this was going to be the result.

that said this solidifies my upgrade path, x6 1100t and OC to 4.2 and call it a day...I am going to try to avoid intel as long as possible, however, it may come to a point where i don't have a choice if i want the most out of my home setup.
 
eastmen said:
gotta say something is fishy there.

at idle its 110/126 and at load its suddenly 244 and 586 ? something isn't right at all.

He said they are only using prime so its the cpu that is stressed in this. So the cpu is already using 134w more non overclocked , If we compare the over clock its usings 460 watts more than idle ?

Can the motherboard even supply 460w of power to the cpu ?

looking at kyle they have the fx 8150@ 4.6ghz using 452 watts vs 120 idle. That means the cpu overclocked is only pulling 332w more. An extra 200mhz is supposed to cause this thing to pull another 130w of power ? something has got to be wrong there
Click to expand...


I don't know how much voltage kyle used, but it probably took a lot more power to squeeze out an extra 200mhz to go to 4.8ghz. bittech did state that they raised the voltage. 4.8ghz was their limit of stability.
But even if it were a typing error, 486 watts is not pretty either.

BD has power gating, so idle performance is expected.
 
Zarathustra[H];1037873342 said:
In early 2009 with cherry-picked parts the absolute highest anyone could get Phenom II x4's to clock on extreme cooling was 4.2Ghz.

Fast forward ~2 years and the production process matured so much that the same parts - as of last December were hitting 4.2ghz on air.

If GloFo's 32nm process matures the same way, we could be looking at 8.5ghz on air out of this architecture by 2013. Sprinkle in a 10% IPC gain once a year with Piledriver, excavator and whatever the other names were, and it really looks like it has potential.
.
Click to expand...

While I want to see amd to succeed for at least the sake of competition in the marketplace and better tech developments, I'm having a hard time following your logic.

P II's got 4.2 at launch not on air, 2 years of developments leads to 4.2 on air.

BD gets 4.6-4.8 at launch not on air, 2 years of developments leads to 8.5 on air?

Unless you're referring to the random BD OC released a while back with really no reproduction outside of amd and without very many specifics told. I don't recall the original PII chips being in quite the same boat as BD now.
 
I just threw together this chart based on power usage from various tests on the internet, as well as AMD's published figures.

It illustrates exactly the problem AMD is dealing with:

6238187253_dda84df8a9_o_d.jpg


The top data point was a review at 4.818Ghz

The two clustered together at 4.6 are the [H] tests. I have done some rough calculations to try to compensate for the fact that these are whole system tests.

The rest are AMD's TDP numbers for the full Opteron 6200 and FX ranges divided by core count.

This chart shows very clearly the exponential relationship with power as clock increases. As the process matures the curve should become flatter and allow much higher clocks.

The one outlier that messes this up seems to be the 4core Opetron. maybe they are betting that corporate purchasing isn't going to look into TDP of these as they are likely workstation chips, and corporate IT usually only looks carefully at TDP for server farms. This knowledge may have been used to bin for the 4 core Opteron.
 
Here's my thought. Have you ever been in this situation? I have to write a paper for school. I've written it. It sucks, so I make changes, but it's even worse. I make more changes, but it still sucks. I delete most of my work and start over. I write it again, still sucks. My head hurts. Crap, it's 5am and this paper is due in a couple of hours. Energy drinks aren't helping anymore. Screw this crap, I'll just turn it in and take whatever grade I get.
 
JapaneseGorilla said:
Here's my thought. Have you ever been in this situation? I have to write a paper for school. I've written it. It sucks, so I make changes, but it's even worse. I make more changes, but it still sucks. I delete most of my work and start over. I write it again, still sucks. My head hurts. Crap, it's 5am and this paper is due in a couple of hours. Energy drinks aren't helping anymore. Screw this crap, I'll just turn it in and take whatever grade I get.
Click to expand...

Except after reading these forums, AMD had 7 years to make remake make remake make remake make their BD or chips before it.
SEVEN years, I don't know whats more disappointing this or Gran Turismo 5. Well at least AMD can improve a couple or things or have something up its sleeve. GT5 is still sucking ass, they had an update that puts in-car views for 800 cars. What it looks like? A blacked out silhouette of what the cars interior is SUPPOSE to look like. Fuck Polyphony.

Anyone know how AMD's stock is doing? Wondering if has or know if it will go down.
 
Bulldozer is the CPU equivalent of The Simpsons Movie

Years of will they wont they, loads of re-writes and time spent on it and both turned out average.
 
dccmadams said:
bulldozer release = fermi release
Click to expand...

Hell, no. Fermi was at least the fastest single chip out the block.

It cannot even compare to the monsterous fail that was the HD2900XT release. Why? Because all failings aside, the HD2900XT was FASTER than it's predecessor in every way possible.
 
i dont think its a good idea to go past 6 cores for desktop.

just give me 4 or 6 powerfull cores and im good.
 
jeremyshaw said:
Hell, no. Fermi was at least the fastest single chip out the block.

It cannot even compare to the monsterous fail that was the HD2900XT release. Why? Because all failings aside, the HD2900XT was FASTER than it's predecessor in every way possible.
Click to expand...

he really said that, what an idiot.
 
PRIME1 said:
Fermi was faster than the competition and is currently outselling the competition. It's more like the HD2900XT
Click to expand...

Thats only because of a core respin and some extreme default cooling. Initial GTX 480/470 may have technically been faster a few times but where huge commercial flops.
 
Bullshit Dozer is a big loser for us. Intel now has little competition & We'll be paying for it.
 
Seriously. if they die-shrunk the PII arch, added a bit more L2 cache, duct-taped two more cores on: we would be looking at better benches and a whole lot less disappointment.

Hey AMD: I'm planning on upgrading my whole machine around July of next year. If you guys don't clean up this bull-chip by then, I'm going Intel.
 
AntiSocialMunky said:
its 20% faster...
Click to expand...

Not across the board. I assume you're talking about a particular bench.
Well, it isn't AES Twofish Serpent... It's something like 12% faster than 2600k in that particular test. The deltas for all the scores range from 20-90 points. The largest delta is 96 points in the Twofish bench and 96/647 = 0.1483 * 100 = 14.83% Where does 20% come from?:confused:
 
Faethon said:
The worst in all this, is if the flagship can't cut it, what's the point in buying one of the lower clocked BD or the 6core BD, when you can still grab a Phenom II at less...
Click to expand...

There is a possibility that since programs do not support multiple cores very well that the performance may not be scaling down so bad. IE just like in the core days when people would buy cheaper dual core CPUs and get better performance. The lowest core count they have is 4. That is not bad. If that has similar clocks and is cheaper then just maybe they have some value at the lower end. We will see when low end bench marks start showing up for all the products.


At any rate since comparable CPUs are smaller in size it should be cheaper for AMD to produce these and they will clear the channels of the older CPUs. Overall I do not think it is worse than previous chips and that will float AMD for a while.

The only hope for AMD is that they are making a long term bet that will pay off as programs make better use of multiple cores. Also perhaps they will improve fabrication of the chips over time. Phenom was worse than this I think. AMD just has problems fabbing a new chip. They need to start leveraging their Graphics business to sell this stuff. They need to have alot of AMD bundles which include a mobo, cpu and video card at a good enough discount that people overlook the CPU short comings.
 
me4get said:
Cache requires lots of transistors.
Click to expand...

To be fair, 8MB AMD cache is more cache tham 8MB Intel cache. Intel makes their's Inclusive, AMD, exclusive. Inclusive copies L1+L2 data to L3, while that takes up space, it also prevents the extra clocks needed for cache snooping... (or just going straight to the IMC and requesting the data from there).
 
jeremyshaw said:
To be fair, 8MB AMD cache is more cache tham 8MB Intel cache. Intel makes their's Inclusive, AMD, exclusive. Inclusive copies L1+L2 data to L3, while that takes up space, it also prevents the extra clocks needed for cache snooping... (or just going straight to the IMC and requesting the data from there).
Click to expand...
It actually has nothing to do with whether the cache is inclusive or exclusive. AMD's cache density is nowhere near Intel's. I'm not sure, but I think AMD's cache density is/was 10-15% lower than Intel's.
 
I've been reading around a lot on the failures of Bulldozer, and the best explanation I have found comes from a Real World Technologies forum post here:

http://www.realworldtech.com/forums/index.cfm?action=detail&id=123110&threadid=123055&roomid=2

Here are the relevant discussion points quoted for easy reading:

>They *had* to have had something in mind, so what could it have been?

There are 2 or 3 former AMD MPU designers who've discussed the BD design and what has been happening to design teams and development methodology at AMD over the past year on usenet and at a few web forums (and suffered great personal attack at the latter for their trouble).

Basically AMD heavily adapted ASIC design style from the GPU side to MPU development and a lot of the "old school" engineers from DEC and NexGen left. MPUs have gotten a lot more complicated over the past decade and AMD didn't have the resources to keep doing it old school unlike Intel and IBM.

From K10 to BD the uarch got narrower and more deeply pipelined which cost IPC. The problem was the clock rate bonus from doing so was lost in the switch over to
a ADE heavy/ASIC flow methodology. I think a lot of folks are putting too much blame today on GF's 32 nm process technology and ignoring the changes in team composition and design methodology at AMD.

Why did AMD bring BD to market? It is clear from the former AMD MPU designers that have talked that BD's shortcomings have been known for a long time but it was the best new design they had kicking around and once it was underway the system takes over and that was what was going to be delivered come hell or high water. Don't underestimate institutional inertia.

It is also a bit unfair comparing K10 and BD because the K10 is the end of a long line of processors that have changed little topologically in critical areas since
the original K7. These have been carefully optimized and refined over multiple re-implementations under the old school methodology. A K10 redone in 32 nm using
the BD methodology would likely also look worse than 45 nm Thuban. Future versions of the BD uarch should be better from lessons learned but IMO don't expect any
miracles. Intel has a huge lead in per core performance and that will probably continue to grow larger over time.

AMD has to consolidate its business and decide what x86 segments it realistically can compete in. Llano seems like a much more successful product because of its focus on value, system integration, and power in a segment which values these attributes.
Click to expand...

And an actual repost of an AMD employee from Semi-Accurate forums:

http://semiaccurate.com/forums/showthread.php?t=5623&page=7

"I don't know. It happened before I left, and there was very little cross-engineering going on. What did happen is that management decided there SHOULD BE such cross-engineering ,which meant we had to stop hand-crafting our CPU designs and switch to an SoC design style. This results in giving up a lot of performance, chip area, and efficiency. The reason DEC Alphas were always much faster than anything else is they designed each transistor by hand. Intel and AMD had always done so at least for the critical parts of the chip. That changed before I left - they started to rely on synthesis tools, automatic place and route tools, etc. I had been in charge of our design flow in the years before I left, and I had tested these tools by asking the companies who sold them to design blocks (adders, multipliers, etc.) using their tools. I let them take as long as they wanted. They always came back to me with designs that were 20% bigger, and 20% slower than our hand-crafted designs, and which suffered from electromigration and other problems."
Click to expand...

Sounds like AMD is stuck in low-performance land because they can't afford to optimize their designs anymore. If this is the case, then it won't magically fix itself anytime soon.
 
defaultluser said:
.....

And an actual repost of an AMD employee from Semi-Accurate forums:

http://semiaccurate.com/forums/showthread.php?t=5623&page=7

Sounds like AMD is stuck in low-performance land because they can't afford to optimize their designs anymore. If this is the case, then it won't magically fix itself anytime soon.
Click to expand...

WOW! This actually explains a LOT!!!! Too bad... Cheap automation in, real talent out, some horrible management decisions...
 
mzs_biteme said:
WOW! This actually explains a LOT!!!! Too bad... Cheap automation in, real talent out, some horrible management decisions...
Click to expand...

Not necessarily horrible management decisions. More like, painful management decisions.

They could have chosen to design the old school way and due to lack of resources never get done, or design the new way, and at least get something out, even if it wasn't perfect.

As far as I am concerned, Bulldozer - while flawed - is better than nothing.

Life is hard when your biggest competitor is ten times your size and can throw engineering talent at stuff you simply can not afford.
 
Zarathustra[H];1037878804 said:
Not necessarily horrible management decisions. More like, painful management decisions.

They could have chosen to design the old school way and due to lack of resources never get done, or design the new way, and at least get something out, even if it wasn't perfect.

As far as I am concerned, Bulldozer - while flawed - is better than nothing.

Life is hard when your biggest competitor is ten times your size and can throw engineering talent at stuff you simply can not afford.
Click to expand...

The advantage of the design tools, is the CPU, or at least part of it, will be more synthesizeable, and can be ported between different fabs more easily. So if GloFo sucks and goes under, TSMC/UMC can help out. Same if the reverse happens (of course, there are still teething problems, like the Sumo GPU of the current Llano APUs, but the concept works).

EDIT: that and Intel has much more in terms of seperate design teams, too. Even as they were developing the failed Netburst chips, other teams were looking for another design goal that would eventually result in: Atom and Core uarch.
 
mzs_biteme said:
WOW! This actually explains a LOT!!!! Too bad... Cheap automation in, real talent out, some horrible management decisions...
Click to expand...

It's all part of the current western management technique. Treat your best staff like crap so they leave and then wonder why you are losing market share and no one buys your products.
 
You guys should keep your minds open too. Let me play the devils advocate. Any person whom is being replaced by a machine or sees the threat of his way of life by a machine is almost always going to regard that as the core reason for any short coming.

I am not taking sides it is just that you have ex employees hating on job threatening tactics. I have seen that in every industry but sooner or later the machines move in. It is very possible this is part of the reason bull dozer is not great but it could also be somewhat sensationalized. AMD has been having issues for a very long time. Was this same practice what caused similar issues with Phenom.
 
You must log in or register to reply here.
Back
Top