Bulldozer, Massively unoptimized

[Vesku]

Definitely need a more mature GF 32nm before judging architecture. As pointed out, Llano MHz to power ratio is not much better than the mature 45nm process. Then there is how high AMD has been setting default voltages in order to get enough qualified dies, as shown by how many Llano parts can be run fine with hefty undervolting.

 

[cbn]

I think its a design dud for the desktop. It simply is not well suited to desktop tasks. It is better suited to server tasks, but even then, 2 billion transistors for such mediocre performance is embarrassing.

The die size came out quite large @ 315mm2.

In a polling thread I started last year, only 31% of the respondants thought Bulldozer would come out greater than 276mm2 for the die.

One possible explanation:

This article mentions a 20% larger die size and 20% performance drop when using automated tools. Estimates (in the article) are that quite a bit of the bulldozer die used these automated tools.

The reason why performance of the long-awaited Bulldozer was below expectations is not only because it was late, but because AMD had adopted design techniques that did not allow it tweak performance, according to an ex-AMD engineer.

Cliff A. Maier, an AMD engineer who left the company several years ago, the chip designer decided to abandon practice of hand-crafting various performance-critical parts of its chips and rely completely on automatic tools. While usage of tools that automatically implement certain technologies into silicon speeds up the design process, they cannot ensure maximum performance and efficiency.

Automated Design = 20% Bigger, 20% Slower

 

[podspi]

Not going to comment on the use of automated tools (since I really have no idea if the 20% number is correct or not) but I believe the article is referring to transistor density, not the # of transistors.

I think it is pretty safe to say BD's size has a lot to do with the 2B transistors sitting there doing ... something.

 

[Soulkeeper]

I'm really surprised by how much of the die is dedicated to things that aren't cores or cache.

 

[SickBeast]

YABDT (yet another bulldozer thread)

Bulldozer sucks, we know about it. If it's priced competetively, it's not really significantly worse than the current Phenom II.

Hopefully they can fix it.

 

[skipsneeky2]

Waits for one of those toms hardware charts showing the intel processors kicking amds ass in the new game skyrim...should win the processor war every time !

So who is gonna post it and who is gonna put the nail in the coffin ? LOL

 

[SickBeast]

Piledriver will pound Bulldozer into the ground. Deep as well. There will be religious clergy from all religions present to preside over the ceremony of bulldozer being laid to rest permanently for the benefit of all humanity. ()

To directly answer your question, AMD is going to do it. :thumbsup::thumbsup:

 

[cbn]

Definitely need a more mature GF 32nm before judging architecture.

Yep, New architecture on new process....things are bound to look better in the future for Bulldozer.

I just hope future versions get the IPC up.

 

[AtenRa]

AMDs Interlagos (16 core) and Valencia (8 core) Opterons price list

 

[Rifter]

Why are people still trying to defend BD?

It sucks, get over it.

1. Lower IPC than their own previous gen chips even after there marketing department went on a IPC is better hypetrain for YEARS prior to release.

2. Perf/watt is a JOKE for a 32nm chip

3. Lower peformance at half of the most popular benchs than their own previous gen chip, less performance in most if you compare clock to clock.

4. Price is a complete joke considering the performance, needs to drop 60-80$ per SKU.

5. Relying on complier updates is a joke, they dont have the market share to expect others to do the work for them and they had many YEARS to do it themselves. Too bad they CHOSE not to, now they have to live with that decision.


I know it sucks to be a AMD fanboy when your company shoots themselves in the foot but thats the facts of life. AMD has always tried to leech of others when it comes to developer relations and coding for optimizations for their products but at some point they have to realize they dont have the market share to force others to spend the money and they will have to do it themselves. It seems this issue that has been hurting there GPU division for years is now becoming a issue for their CPU division as well.

Edit for failed spelling

 

[videogames101]

Really? Thank you for adding value to this conversation.

I suppose some people are just happy that they can hate something and don't care about anything beyond that.

Bulldozer is a very promising CPU architecture that can only become more relevant with time - software will get more threaded, not less. I will probably not buy an FX now and will wait for a stepping or generation update, but I'm very much looking forward to what AMD can make of it. You, meanwihle, hate on!

We're not here to speculate on performance 2 years out. Right now, as of 11/14/2011 bulldozer has absolutely nothing to offer most users. If Bulldozer starts performing better on new applications than it might become a better value, but you cannot possibly argue the fact that bulldozer, RIGHT NOW, sucks balls.

 

[podspi]

AMDs Interlagos (16 core) and Valencia (8 core) Opterons price list

About time! Hopefully we'll get some benchmarks now on proper server platforms...


Particularly interesting looking at the performance/watt of the 35W part...

 

[cbn]

I've seen talk of Integer and floating point in this thread.

However, I have yet to find SPECint_2006 results during my internet search.

Does anyone have links to those results?

 

[AtenRa]

I've seen talk of Integer and floating point in this thread.

However, I have yet to find SPECint_2006 results during my internet search.

Does anyone have links to those results?

 

[Idontcare]

I find those numbers to be impressive, but I don't see any of the review sites talking about it. So what's the rub? I don't seen anything that's not to be liked here, I'd expect a little more excitement over a product like this.

 

[AtenRa]

I believe SB-E launch has something to do with it

I will have my review ready (sorry no benchmarks) tomorrow with more information and more slides.

 

[podspi]

IDC, it is launching on the same day SB-E


But I agree, it doesn't look bad. I am not surprised there aren't a ton of reviews, it is a server chip after all. Has Anandtech even had a server CPU review @ launch? They usually review servers that have the chips in them...

Edit:AtenRa beat me to it...

AtenRa, if you can say, how long do you think it'll take before they get reviewable platforms out to people?

 

[AtenRa]

I have no idea because i havent asked for a sample. AMD have said that their partners will have systems ready at launch, that's today.

 

[hooflung]

The intellectual dishonesty is so bad in here it stinks.

That watt/thread chart is so stupid if you read anything into it then you are devaluing your point, not empowering it.

The Intel chips cannot, ever, execute more than 1 thread per core even with HT on. It is physically impossible. It doesn't have the the scheduler support. So the question is in what time can it complete 8 threads and how much wattage did it take in comparison to the BD in the same time. Then look at which got the most amount of work done. In some cases Intel my very well push out 8 threads in the same time it takes the BD to do 8. But it is impossible to have that covered in all cases, which is why the BD wins in some workloads.

Until that test happens I call shenanigans.

And how is that SB-e doing for you. Having to buy into a new motherboard yet again. And buying 4 - 8 sticks per install, and having a 500+ dollar cpu. With the same amount of silicon as a BD with minimal real performance benefits from SB.

 

[videogames101]

The intellectual dishonesty is so bad in here it stinks.

That watt/thread chart is so stupid if you read anything into it then you are devaluing your point, not empowering it.

The Intel chips cannot, ever, execute more than 1 thread per core even with HT on. It is physically impossible. It doesn't have the the scheduler support. So the question is in what time can it complete 8 threads and how much wattage did it take in comparison to the BD in the same time. Then look at which got the most amount of work done. In some cases Intel my very well push out 8 threads in the same time it takes the BD to do 8. But it is impossible to have that covered in all cases, which is why the BD wins in some workloads.

Until that test happens I call shenanigans.

And how is that SB-e doing for you. Having to buy into a new motherboard yet again. And buying 4 - 8 sticks per install, and having a 500+ dollar cpu. With the same amount of silicon as a BD with minimal real performance benefits from SB.

Can't tell if serious?

 

[Chiropteran]

FX-8150 still out of stock. Despite all the bad press they seemed to sell as many as they could make. Should have charged more per CPU for the initial batch, probably.

 

[voltronn]

FX-8150 still out of stock. Despite all the bad press they seemed to sell as many as they could make. Should have charged more per CPU for the initial batch, probably.

In stock in Europe if you wanna order
http://www3.hardwareversand.de/artic...1673&agid=1242
http://www4.anobo.de/articledetail.jsp?aid=51673&agid=1242&pvid=4n7ldwjdx_guzmdors&ref=13&lb

 

[Chiropteran]

In stock in Europe if you wanna order

Don't want one that badly, just making an observation. If I can't buy something on newegg.com or amazon.com I'll find an excuse to not buy it

 

[cbn]

Some Integer results from AMD Zone http://www.amdzone.com/phpbb3/viewtopic.php?f=532&t=138927&p=214043#p214019

BD 6282 SE 2.6GHz@ 2s/32c = 526 SPECint_rate2006
Xeon E7-4870 2.4GHz @ 2s/20c = 553 SPECint_rate2006
Xeon 5690 3.46GHz @ 2s/12c = 393 SPECint_rate2006
Opteron 6180 SE 2.5GHz @ 2s/24c = 430 SPECint_rate2006

SB 2600K 3.4GHz @ 1S/4C = 156 SPECint_rate2006

BD = 16.4 SPECint_rate per core @2.6 GHz
K10 = 17.9 SPECint_rate per core @2.5 GHz
Westmere = 27.7 SPECint_rate per core @2.4 GHz
Westmere = 32.7 SPECint_rate per core @ 3.46 GHz
Sandy Bridge = 39 SPECint_rate per core @ 3.4 GHz

I have to admit I am disappointed.

Hopefully AMD can fix this and beef up those integer cores.

 

[Gikaseixas]

More to read on the Opteron:
http://hexus.net/tech/news/cpu/32595-amd-bulldozers-raison-detre/

AMD announced today the immediate launch and availability of its Opteron 6200 and 4200 series of Bulldozer-based server processors, previously codenamed Interlagos and Valencia, designed for Socket G34 and C32 motherboards respectively.

The new silicon offers exceptional value for money for those that need the horse-power, with AMD claiming up to 84 per cent greater performance and 73 per cent greater memory bandwidth for its Operton 6276 when benched against the Intel Xeon X5670, made possible thanks to the Inerlagos' monstrous, hexadeca-16-core design. Quite rightfully, by cramming such performance in to a single chip, cost effectiveness for large server farms goes through-the-roof as you require fewer racks to achieve comparable performance, with AMD also claiming that the technology requires two third's less floor space and money, in comparison to the Intel Xeon 5600 series.

AMD is also claiming the greatest performance-per-core with its 4200 EE series. The math works out at 4.375W per core, with the lowest comparable Intel server offering at 10W per core, the Xeon L5630. The firm has also announced that the new line of Opterons are the only server processors currently supporting ultra-low 1.25v memory.
Initial pricing appears more than competitive, with, for example, 1k unit costs for the AMD Opteron 6276 at $788, compared to $774 for an Xeon E5640. AMD's frontrunner will be the Opteron 6282 SE 16-core, 4-way Interlagos Bulldozer, clocking in at 2.6GHz per core, with a turbo frequency of 3.3GHz, 16MB level three cache and a 140 Watt TDP, with a 1k unit cost of $1019.

It's about time to see AMD's Bulldozer architecture finally do what it was made for, which as we've always suspected, is highly-threaded server work; with AMD's initial figures seeming to show that it does this job exceptionally well. This new product line does, however, give us an interesting insight into the large scalability of the Bulldozer design in general; proof that it can utilise low-power memory, range from 4 to 16 cores, scale from TDPs of 35W to 140W and from previous over-clocking evidence, scale incredibly well from low high-core-count server frequencies up to record-breaking single module frequencies. Along with our prior post on scheduler performance optimisations, perhaps this all shows that there's plenty more left to come out from the Bulldozer architecture.