Breaking down BD

[Digital Viper-X-]

Good article over at anandtech

http://www.anandtech.com/show/5057/the-bulldozer-aftermath-delving-even-deeper

They go through the short comings and strengths of the architecture and is fairly detailed.

 

[MacLeod]

Good read. A bit depressing though. Looks to confirm that Bulldozer was always meant as a server proc with desktop performance not a big consideration.

 

[cageymaru]

AMD gone out of business yet?

 

[DejaWiz]

Interesting read. Seems to me that AMD clearly gambled on trying to hype up and market a purely server-oriented CPU as an enthsiast consumer flagship king-of-the-hill which has come back to bite them in the ass once FX did arrive.

I mean, really, did AMD actually expect to get away with not being able to deliver on their pre-release marketing hype and performance claims given the sheer amount of users and review sites out there on the web that clearly remember the Pentium 4 fiasco? Same old shit, different brand name.

 

[ebduncan]

i knew most of these problems already from my own testing. What it really boils down to is the FX processors for both desktop and server failed to hit their mark with operating frequency really, then combine that with some cache issues and branch prediction penalties and you get what is called the FX.

Pile driver actually addresses all 3 of these problems, so it will be interesting to see how they fare. I doubt it will be enough though. All I know if Amd doesn't scrape their plans with the CPUs then by the time excavator comes out should be very competitive. 10-15% gains per year pile driver 10-15%, steam roller10-15%, excavator 10-15% On the high side of things Excavator will be 45% faster than FX, on the low side of things 25-30% faster.

Intel managed 5-6% from sandy bridge to Ivy bridge, with most of the advancements coming from the new GPU. Who knows was Haswell will bring but if its another 5-6% then well Amd has a chance.

 

[MacLeod]

Interesting read. Seems to me that AMD clearly gambled on trying to hype up and market a purely server-oriented CPU as an enthsiast consumer flagship king-of-the-hill which has come back to bite them in the ass once FX did arrive.

I mean, really, did AMD actually expect to get away with not being able to deliver on their pre-release marketing hype and performance claims given the sheer amount of users and review sites out there on the web that clearly remember the Pentium 4 fiasco? Same old shit, different brand name.

Ive heard that before that AMD hyped the crap out of BD before it hit. I remember the opposite. i remember being frustrated and pissed that we werent hearing anything at all out of AMD. I remember talking about how their silence was probably a bad sign that things werent working out like theyd hoped.

Now I know that marketing department put out their usual stuff promoting it and everything but I dont really remember all this hype and grandeur that people talk about. Its entirely possible I just missed it but most of what I remember from 2009 til 2011 was nothing but silence from AMD.

Now this isnt counting that JF-AMD guy that was doing his thing, whatever that was. Something like "I know Bulldozer.....I hate when people ask me about Bulldozer...IPC is for retards......I knew I shouldnt have tried to talk to a bunch of knuckle dragging enthusiasts."

I think the biggest thing that screwed up Bulldozer was the delayed release. Bulldozer was supposed to release in late 2009 or early 2010 IIRC which wouldve put it squarely up against the i7-900's and 800's for which it competes pretty good with. But since it came out when it did, it was essentially a generation late and no match for Sandy.

Now the thing that I think could save them is that while AMD is being beat up for concentrating on APU's, Intel kinda is doing the same thing. Ivy wasnt much of an upgrade at all in pure speed over Sandy but its GPU and laptop potential was. That could be a blessing for AMD in as much as it gives them a generation to catch up a little. If they can come up with a big step forward with Vishera and then again with Steamroller, we could be getting back to a competitive CPU market.

 

[AntiSocialMunky]

They did a lot of market hype at the thing and then they laid off a lot of the marketers after the release lol.

 

[GenTarkin]

Well, its nice to know some answers as to why its very picky bout desktop workloads. I for one, am excited to see what piledriver brings. I know the very tweaks done to PD cores were exactly in line w/ this article.
But, I for one give props to AMD engineers to be able to increase the branch predictor accuracy so much from stars core, even though the lengthening of pipeline seems to have outrun the branch predictor opimizations, still the fact they got that much more accuracy out of branch prediction is awesome. Just hope PD branch predictor optimization closes that gap a bit more.
I personally love the BD uarch, even tho its the first iteration and fell short of expectations. I hope AMD invests in tweaking it as much as possible.
I mean, just as the article explains, when barcelona launched it was a major wreck, yet years later we have that same underlying uarch running like 2x as fast, using half the power and a decent IPC increase. This is exactly what BD needs =)

 

[stealthy123]

They did a lot of market hype at the thing and then they laid off a lot of the marketers after the release lol.

I still don't remember this market hype so many people mention.

There were a bunch of forum fanboys, but I never saw this marketing push people keep insisting happened.

Reason the marketing department got sacked was they sucked. They didn't push the products enough. This hole(intentional typo) idea that AMDs marketing department pushed a new product falsly is just dumb. Every marketing department lies about their product, for fuck's sake, that's their job.

 

[JohnnyH24]

That was a good article. From a historical perspective...

The K10 (Phenom) was originally designed as a server processor. Like Bulldozer, it was greeted with meh reviews and couldn't keep up with the performance the humble Intel Q6600 offered. And gaming performance was poorer than the older Athlon X2 6000+. It took two years and two iterations of the processor before Deneb became a competitive alternative to Intel.

Hopefully, it doesn't take AMD two years to make Bulldozer right for the desktop.

I guess I should've learned something - I bought a Phenom 9600 BE followed by a Phenom 9850 BE, before settling for a Phenom 2 955 BE. Batting a 1.000, I got the FX-8120.

 

[RamonGTP]

i knew most of these problems already from my own testing. What it really boils down to is the FX processors for both desktop and server failed to hit their mark with operating frequency really, then combine that with some cache issues and branch prediction penalties and you get what is called the FX.

Pile driver actually addresses all 3 of these problems, so it will be interesting to see how they fare. I doubt it will be enough though. All I know if Amd doesn't scrape their plans with the CPUs then by the time excavator comes out should be very competitive. 10-15% gains per year pile driver 10-15%, steam roller10-15%, excavator 10-15% On the high side of things Excavator will be 45% faster than FX, on the low side of things 25-30% faster.

Intel managed 5-6% from sandy bridge to Ivy bridge, with most of the advancements coming from the new GPU. Who knows was Haswell will bring but if its another 5-6% then well Amd has a chance.

Ivy Bridge was a "tick" it was never meant to be a huge gain over IB. Everything about IB performed to expectations with the sole exception of temps. I don't see this grand silver lining you do. Intel his hit every mark they've wanted to hit since the introduction of Core. AMD has fell short every single time since launch of the original Phenom. There's a very clear trend and there's nothing to suggest its slowing down much less reversing like you've suggested here.

 

[ebduncan]

Ivy Bridge was a "tick" it was never meant to be a huge gain over IB. Everything about IB performed to expectations with the sole exception of temps. I don't see this grand silver lining you do. Intel his hit every mark they've wanted to hit since the introduction of Core. AMD has fell short every single time since launch of the original Phenom. There's a very clear trend and there's nothing to suggest its slowing down much less reversing like you've suggested here.

? I wasn't suggesting anything. I was just saying if AMD can get 10-15% more performance per year with the Bulldozer Arch, then They will become alot more competitive with Intel, granted that Intel does not make great strides with Haswell, ie only 5-6% more over ivy bridge.

IE FX- Piledriver 10-15%- Steam Roller 10-15% -excavator 10-15% Per Amd goal slides.

was purely synthetically if FX with Excavator cores has 45% more performance vs FX and Intel only brings 5-6% more with haswell, Amd will be competitive again.....

that's all i was really saying.

 

[RamonGTP]

Yeah, and that has about the same probability of happening as the sun going nova in our lifetime, turning into a black hole, sucking is in, spitting us out, Earth getting captured by another suns gravitational pull with an orbit in the habitable zone and life continuing on earth for another 30 billion years.

 

[ebduncan]

Yeah, and that has about the same probability of happening as the sun going nova in our lifetime, turning into a black hole, sucking is in, spitting us out, Earth getting captured by another suns gravitational pull with an orbit in the habitable zone and life continuing on earth for another 30 billion years.

LOL the odds are way better than that. They are bad odds, but not that bad lol.

 

[RamonGTP]

I was going to stop at "the sun going nova in our lifetime" but didnt think AMDs chances are quite that good.

 

[daglesj]

I like some of the ideas and concepts re. Bulldozer. It's interesting to see a different approach.

The thing that stops me buying as a AMD enthusiast is that it's design is not quite there and compromised in certain places. It's a MK1 architecture and I'd rather see a refined/revised MK2 version first.

I have no issue with buying a CPU that goes slower than another one if it costs a little less and still does the work I need. As long as it's working to the best of it's ability.

However, I wouldn't if I knew the reason for it going slower was due to it not being able to fire on all 6 cylinders due to a design flaw/compromise.

If the next gen is more refined and less compromised then they may find their way into my systems.

I don't sit running benmarks all day or transcoding video I may add.

 

[pelo]

That was a good article. From a historical perspective...

The K10 (Phenom) was originally designed as a server processor. Like Bulldozer, it was greeted with meh reviews and couldn't keep up with the performance the humble Intel Q6600 offered. And gaming performance was poorer than the older Athlon X2 6000+. It took two years and two iterations of the processor before Deneb became a competitive alternative to Intel.

Hopefully, it doesn't take AMD two years to make Bulldozer right for the desktop.

I guess I should've learned something - I bought a Phenom 9600 BE followed by a Phenom 9850 BE, before settling for a Phenom 2 955 BE. Batting a 1.000, I got the FX-8120.

That's how all chips have been designed in recent years: laptops and server as their primary focus. That's what people forget, particularly us enthusiasts who believe that we're the center of AMD/Intel's universe but in reality we're an afterthought. Intel makes laptop chips for the desktop and AMD has server chips for the desktop. Intel also has the 2011 workstation chips which are essentially server chips.

"How do we appeal to enthusiasts?"

"Let's release a couple of chips that have unlocked multipliers but give them the high end graphics so they have to pay extra for things they won't use."

"Brilliant!"

Not that AMD would do any better. The only reason we have unlocked AMD chips is because they have no choice but to unlock them

 

[daglesj]

That's how all chips have been designed in recent years: laptops and server as their primary focus. That's what people forget, particularly us enthusiasts who believe that we're the center of AMD/Intel's universe but in reality we're an afterthought. Intel makes laptop chips for the desktop and AMD has server chips for the desktop. Intel also has the 2011 workstation chips which are essentially server chips.

"How do we appeal to enthusiasts?"

"Let's release a couple of chips that have unlocked multipliers but give them the high end graphics so they have to pay extra for things they won't use."

"Brilliant!"

Not that AMD would do any better. The only reason we have unlocked AMD chips is because they have no choice but to unlock them

Shhhh a few here will think you've molested their mum by saying that AMD/Intel don't give a crap about them.

After all outside these forums Joe Average doesn't have a clue much about AMD/Intel technology. All I get asked by customers is "will it be faster than my PC I bought in 2002 and is it cheap?"

 

[MacLeod]

That's how all chips have been designed in recent years: laptops and server as their primary focus. That's what people forget, particularly us enthusiasts who believe that we're the center of AMD/Intel's universe but in reality we're an afterthought. Intel makes laptop chips for the desktop and AMD has server chips for the desktop. Intel also has the 2011 workstation chips which are essentially server chips.

"How do we appeal to enthusiasts?"

"Let's release a couple of chips that have unlocked multipliers but give them the high end graphics so they have to pay extra for things they won't use."

"Brilliant!"

Not that AMD would do any better. The only reason we have unlocked AMD chips is because they have no choice but to unlock them

Totally agree. Why does Intel charge a $250 minimum to overclock an IB? Because they can. This is why I'm pulling for AMD so much. A competitive processor market benefits us [H] types the most.

 

[DejaWiz]

Totally agree. Why does Intel charge a $250 minimum to overclock an IB? Because they can. This is why I'm pulling for AMD so much. A competitive processor market benefits us [H] types the most.

Then why was AMD charging MORE for their inferior FX products at time of release? Because they can, too. From a historical viewpoint, $250 for a quad core unlocked processor is peanuts compared to the $600-800 (or more) cost of mid to upper binned dual cores from the AMD X2 S939 era.

 

[Dan_D]

Totally agree. Why does Intel charge a $250 minimum to overclock an IB? Because they can. This is why I'm pulling for AMD so much. A competitive processor market benefits us [H] types the most.

Nonsense. Intel charges a little bit to overclock their K series processors vs. standard ones. There are Core i7 2500's, etc. which are not K parts and do not overclock well. These are about $20 cheaper or so.

 

[pelo]

Then why was AMD charging MORE for their inferior FX products at time of release? Because they can, too. From a historical viewpoint, $250 for a quad core unlocked processor is peanuts compared to the $600-800 (or more) cost of mid to upper binned dual cores from the AMD X2 S939 era.

They're bigger chips and they had fewer per wafer. They had to cost more and yet AMD was still making less per-chip than Intel was.

That doesn't excuse that they were still far too expensive. Fanboys and the few people that they made sense for actually bought them at that price but yes, they were still too expensive. They're finally sitting at price points which they should have launched at if you consider their relative performance and chip availability yet it's taken AMD a 2011 platform+CPUs and Ivy Bridge to get there and they're still not a good buy.

They don't care about the desktop enthusiast. But neither does Intel really so there's no reason to get upset at just one of them We enthusiasts spend a lot of money on hardware but a majority of it is not on the CPU. The days where a $1000 CPU was a massive upgrade from a $150 one are long over for us and there's no need to buy the $1000 chips anymore.

Someone mentioned the misprediction penalty being improved in Piledriver but I'm not sure that's true. AMD hasn't changed the pipeline length or the uOP placement. The L2 might be a tiny bit faster but that horrible misprediction penalty will likely stay until they completely overhaul the core architecture. Essentially, AMD improved the branch prediction (very slightly) but the branch misprediction penalty is still the bigger issue.

 

[Solomutt]

I think you lost me there. That post seemed to be heading towards a point, then veered off into some other direction. Please note that P4 EE were way up there, too, and still weak.

Edit: this was to the guy referencing S939 procs being expensive.

 

[n370zed]

Very informative read that article.

 

[MacLeod]

Nonsense. Intel charges a little bit to overclock their K series processors vs. standard ones. There are Core i7 2500's, etc. which are not K parts and do not overclock well. These are about $20 cheaper or so.

Ummm, right. That's what I said. $240 ( Newegg price) is the minimum price you have to pay if you want to overclock a Ivy Bridge. $220 is the minimum you'll have to pay if you want to OC a Sandy Bridge. You can't buy a $120 i3-2100 and overclock it. You used to be able to overclock any Intel proc. Hell you used to be able to OC their $50 Pentium dual cores to over 4 GHz!

 

[Dan_D]

Ummm, right. That's what I said. $240 ( Newegg price) is the minimum price you have to pay if you want to overclock am Ivy Bridge. You can't buy a $120 i3-2100 and overclock it. $220 is the minimum to overclock a Sandy Bridge. You used to be able to overclock any Intel proc. Hell you used to OC their $50 Pentium dual cores to over 4 GHz!

No, the difference in price isn't $240. It's the cost difference between non-overclockable parts and overclockables. For like processors, that difference is minimal.

Core i5 2500K = $169.99 at MC or $204.99 at Newegg.
Core i5 2500 = $209.99 at Newegg

In this case you pay more for the non-overclockable part unless you have a Microcenter around. If you want to use your Core i3 2100 example, you are talking about a CPU with two less CPU cores, which is apples and oranges, but let's use it anyway.

Core i3 2100 = $119.99 at Newegg

Even thoough these aren't apples to apples processors, at best the Core i5 2500K is only about $90 more than the Core i3 2100 is. Plain and simple.

 

[defaultluser]

Ummm, right. That's what I said. $240 ( Newegg price) is the minimum price you have to pay if you want to overclock a Ivy Bridge. $220 is the minimum you'll have to pay if you want to OC a Sandy Bridge. You can't buy a $120 i3-2100 and overclock it. You used to be able to overclock any Intel proc. Hell you used to be able to OC their $50 Pentium dual cores to over 4 GHz!

Right, and they stopped that because it wasn't making them any money. People would buy the absolute cheapest chips on the planet with little profit margin, overclock and sit on them as long as they could. Rise, repeat.

Now people are forced to buy the performance they NEED. This has not hurt the value of Intel in the low-end - you can get nearly the same frame rates with an i3 as you get with an i5, and the Pentium dual-core has a respectable gaming and single-threaded performance lead over AMD at every price-point.

If you're addicted to tweaking, you can still get your hands dirty if you spend the dough for a K-series. And yes, there's no K-series dual-core because it would cannibalize quad-core sales to enthusiasts.

 

[MacLeod]

No, the difference in price isn't $240. It's the cost difference between non-overclockable parts and overclockables. For like processors, that difference is minimal.

Core i5 2500K = $169.99 at MC or $204.99 at Newegg.
Core i5 2500 = $209.99 at Newegg

In this case you pay more for the non-overclockable part unless you have a Microcenter around. If you want to use your Core i3 2100 example, you are talking about a CPU with two less CPU cores, which is apples and oranges, but let's use it anyway.

Core i3 2100 = $119.99 at Newegg

Even thoough these aren't apples to apples processors, at best the Core i5 2500K is only about $90 more than the Core i3 2100 is. Plain and simple.

I'm not saying you have to pay $240 MORE for a K series proc over a non-K, I'm saying you have to pay a MINIMUM of $240 to overclock cause the 3570K is currently $240 at the Egg. That's the cheapest IB that you can OC more than a couple MHz.

 

[pelo]

I think Dan's point is that Intel has to compete with itself regardless of whether AMD shows up or not. With their tick-tock model they rely heavily on frequent sales to generate them some income and at the moment it's been working. The issue is that IB isn't as appealing for enthusiasts as SB was and it doesn't look like Haswell will be either as both IB/Haswell are for the most part GPU-centric upgrades.

They've certainly segmented the desktop space, though. You can clearly pick out which chip is an enthusiast chip and which isn't and there's only 2 of them.

 

[DejaWiz]

They're bigger chips and they had fewer per wafer. They had to cost more and yet AMD was still making less per-chip than Intel was.

That doesn't excuse that they were still far too expensive. Fanboys and the few people that they made sense for actually bought them at that price but yes, they were still too expensive. They're finally sitting at price points which they should have launched at if you consider their relative performance and chip availability yet it's taken AMD a 2011 platform+CPUs and Ivy Bridge to get there and they're still not a good buy.

They don't care about the desktop enthusiast. But neither does Intel really so there's no reason to get upset at just one of them We enthusiasts spend a lot of money on hardware but a majority of it is not on the CPU. The days where a $1000 CPU was a massive upgrade from a $150 one are long over for us and there's no need to buy the $1000 chips anymore.

Someone mentioned the misprediction penalty being improved in Piledriver but I'm not sure that's true. AMD hasn't changed the pipeline length or the uOP placement. The L2 might be a tiny bit faster but that horrible misprediction penalty will likely stay until they completely overhaul the core architecture. Essentially, AMD improved the branch prediction (very slightly) but the branch misprediction penalty is still the bigger issue.

I completely agree and what you just posted is basically an elaboration of my point.

The only thing that Intel and AMD puts care into with regard to us enthusiasts is getting more money out of us, because it leaves us with no choice but to do so.

I am completely happy that the days of $1000 overpriced dual cores are long gone. My point was, $250 or less for a quad core that is the fastest consumer level mainstream CPU in the world and can also be easily overclocked is a bargain compared to what the market was like for us enthusiasts 5-7 years ago. These are some pretty awesome days we're living in.

 

[Gman1979]

I've always wondered how much difference the absorbing of the ex DEC ALPHA staff by Intel affected our current cpu situation from both companies. Some of the previous members of that team went to AMD when Compaq bought DEC and were responsible for the design lead AMD had over the Pentium 4 at the time. When HP bought Compaq, they inherited what was left of the DEC engineering teams and made a deal with Intel to ship them off to work with them when HP killed off ALPHA development. Couple years later we get the CORE architecture and an improved Itanium design also.

All those high paid engineers AMD lost when the company was massively in debt? Yeah guess which engineers were first out the door..the ex-DEC guys. Now AMD has a pentium 4 on their hands

 

[reb00tin]

I've always wondered how much difference the absorbing of the ex DEC ALPHA staff by Intel affected our current cpu situation from both companies. Some of the previous members of that team went to AMD when Compaq bought DEC and were responsible for the design lead AMD had over the Pentium 4 at the time. When HP bought Compaq, they inherited what was left of the DEC engineering teams and made a deal with Intel to ship them off to work with them when HP killed off ALPHA development. Couple years later we get the CORE architecture and an improved Itanium design also.

All those high paid engineers AMD lost when the company was massively in debt? Yeah guess which engineers were first out the door..the ex-DEC guys. Now AMD has a pentium 4 on their hands

Microsoft made a version of Windows 2000 for the DEC Alpha, try it, performance is still competitive, not bad for 20+ year-old technology.

 

[pelo]

Some of the engineers that worked on Bulldozer were from SeaMicro, the same company AMD acquired a couple of months ago.

Intel and AMD swap engineers as well; or used to, rather. Now when someone leaves they go to Qualcomm or Samsung

 

[ebduncan]

how did this topic go from the faults of Amd Bulldozer Arch, to comparison of price between unlocked intel cpus and non unlocked ones?

i mean that's some serious thread hijacking right there.

Honestly Both Amd and Intel charge a small premium for unlocked parts. (a8 3870k, 2500k, 2600k, 2700k etc...)

Bulldozer being a performance desktop cpu aimed at enthusiast it made sense to leave them unlocked. Granted Amd Fx systems are geared towards parallelism, or things that will make use of 8 cores with good scaling. In these work loads it is competitively priced. Encoding, file compression/ encryption, Virtual machines, and various server work loads. While it doesn't Best Intel even in these fields (intel has other processors which are more expensive that will fit these bills) Making the FX processors rather niche.

Even if Piledriver manages a 10-15% leap in IPC, for general purpose computing Intel will be a better choice. Granted in the applications which benefit from parallelism, Amd would likely have good size lead over Intel for the $. Sad part not many of these applications actually exist.

Now if Amd is on Target with Steam Roller and is able to introduce HSA to the desktop world, then Combining A FX 8 core integer Cpu with discrete glass GPU for Fpu instructions, you might have something remarkable. Then again will probably be in the same boat as they are in now, being held back by software that takes advantage of it.

Someone mentioned the misprediction penalty being improved in Piledriver but I'm not sure that's true. AMD hasn't changed the pipeline length or the uOP placement. The L2 might be a tiny bit faster but that horrible misprediction penalty will likely stay until they completely overhaul the core architecture. Essentially, AMD improved the branch prediction (very slightly) but the branch misprediction penalty is still the bigger issue.

Yes i said Amd was working on tweaking the branch prediction. To help minimize how often prediction penalties occur. I don't think they can overhaul the actual penalties, but minimize when they happen at least. Amd's Branch prediction is pretty good atm though, just the penalties that occur when a branch is mispredicted are quite bad. FX doesn't have uOP cache. Sandy bridge/Ivy bridge have these not AMD FX.

 

[pelo]

I was surprised to see their branch prediction was so good. But like you said, you can't improve much there and what you do improve isn't going to make too much of a difference. The mispredict penalties will only come with a restructuring of their pipeline and cache access. The fact that it's slow and the clock speeds are low doesn't matter too much on the server end (though it definitely makes a difference), it's the performance hits on the desktop where the L2/L3 heavy apps rely on fast access and proper predictions with low penalties where it hurts (gaming, for instance). The added cache helps but it doesn't overcome the issues.

I thought they do but it's placed post decoder and not pre?

Bulldozer is a deeply pipelined CPU, just like Sandy Bridge, but the latter has a µop cache that can cut the fetching and decoding cycles out of the branch misprediction penalty.

Err, you're right.

What's worse is that the write-coalesce cache that AMD added to serve between the L1 and L2 is hampered by the L2 cache's problems. I'm not sure we'll see a write-through approach unless AMD can iron out the issues with the misprediction penalties and L2 cycles. Although they might not even need a WCC if they increase their 4-way to 8-way access? I don't know

 

[DejaWiz]

I don't see it as hijacking in this particular case...just how the conversation flowed and steered. The price is relevant to the FX architecture and performance... Since we all vote with our wallets.

 

[Dan_D]

I was surprised to see their branch prediction was so good. But like you said, you can't improve much there and what you do improve isn't going to make too much of a difference. The mispredict penalties will only come with a restructuring of their pipeline and cache access. The fact that it's slow and the clock speeds are low doesn't matter too much on the server end (though it definitely makes a difference), it's the performance hits on the desktop where the L2/L3 heavy apps rely on fast access and proper predictions with low penalties where it hurts (gaming, for instance). The added cache helps but it doesn't overcome the issues.

I thought they do but it's placed post decoder and not pre?



Err, you're right.

What's worse is that the write-coalesce cache that AMD added to serve between the L1 and L2 is hampered by the L2 cache's problems. I'm not sure we'll see a write-through approach unless AMD can iron out the issues with the misprediction penalties and L2 cycles. Although they might not even need a WCC if they increase their 4-way to 8-way access? I don't know

At the end of the day, the truth is Bulldozer isn't particularly good at anything in it's current form. With low IPC, tons of branch mispredictions, and low clock speeds, combined with high power consumption it misses the mark for either the consumer desktop or server markets.

The difference is, the architecture has potential with changes for the server market. For the desktop market, not so much.

 

[pelo]

Dan, it makes a bit more sense if you look at AMD's roadmaps. They're either server-focused or laptop-APU focused with nothing in between. On the server, high number of integer coars with highly parallel workloads make a lot more sense, especially if you're going to be compiling for a particular ISA/architecture (HPC mostly). So its shortfalls, though they cripple it there too, are sort of excusable because you can tell what their intentions were.

If you look at APUs and HSA it makes sense as well. Half the number of FPUs because in the future AMD will rely on GPGPU to make up for that disadvantage and turn it into a massive advantage. Therefore you only need a higher number of integer cores compared to a lower number of relative FPUs (relative because you can't have all FP-related workloads done by the GPU but you only need it for some [and that too depends on the software at hand]). So the modular approach makes a lot of sense.

The desktop is stuck in this weird middle ground for them at the moment but only for the enthusiast. AMD shipped a lot of APUs on the desktop but that doesn't really appeal to the enthusiast. With server workloads becoming more and more parallel the desktop typical workloads are taking a back seat. While desktop workloads often cross over with laptop workloads (it's why SB/IB do well on both), the laptops benefit from the on-die GPU much more significantly (also why SB/IB have an on-die GPU).

So, yea, AMD's chips suck ass for the desktop but they actually make a lot more sense in their intended markets.

 

[sirmonkey1985]

i think the architecture its self has a ton of potential its just getting to that point. you can't really change a processor architecture as drastic as bulldozer is over night. none the less bulldozer is still a step in the right direction, they just need to fine tune it some more and fix the obvious problems before moving on. in the end i think the modular design is where all processors will be in the next 2-3 years. intel can only go so far with hyperthreading so i wouldn't be surprised if they themselves have already started working on a similar modular design.

Dan, it makes a bit more sense if you look at AMD's roadmaps. They're either server-focused or laptop-APU focused with nothing in between. On the server, high number of integer coars with highly parallel workloads make a lot more sense, especially if you're going to be compiling for a particular ISA/architecture (HPC mostly). So its shortfalls, though they cripple it there too, are sort of excusable because you can tell what their intentions were.

If you look at APUs and HSA it makes sense as well. Half the number of FPUs because in the future AMD will rely on GPGPU to make up for that disadvantage and turn it into a massive advantage. Therefore you only need a higher number of integer cores compared to a lower number of relative FPUs (relative because you can't have all FP-related workloads done by the GPU but you only need it for some [and that too depends on the software at hand]). So the modular approach makes a lot of sense.

The desktop is stuck in this weird middle ground for them at the moment but only for the enthusiast. AMD shipped a lot of APUs on the desktop but that doesn't really appeal to the enthusiast. With server workloads becoming more and more parallel the desktop typical workloads are taking a back seat. While desktop workloads often cross over with laptop workloads (it's why SB/IB do well on both), the laptops benefit from the on-die GPU much more significantly (also why SB/IB have an on-die GPU).

So, yea, AMD's chips suck ass for the desktop but they actually make a lot more sense in their intended markets.

server and APU markets are where the real money is right now and the future, the desktop market is slowly on its way out as the majority market move to cloud computing and mobile devices. the desktop probably has another 5 maybe 6 years left it in which at that point it'll mostly just be diehard pc users and enthusiast still using them. AMD has a pretty big lead on Intel when it comes to the APU market.

 

[Dan_D]

Dan, it makes a bit more sense if you look at AMD's roadmaps. They're either server-focused or laptop-APU focused with nothing in between. On the server, high number of integer coars with highly parallel workloads make a lot more sense, especially if you're going to be compiling for a particular ISA/architecture (HPC mostly). So its shortfalls, though they cripple it there too, are sort of excusable because you can tell what their intentions were.

If you look at APUs and HSA it makes sense as well. Half the number of FPUs because in the future AMD will rely on GPGPU to make up for that disadvantage and turn it into a massive advantage. Therefore you only need a higher number of integer cores compared to a lower number of relative FPUs (relative because you can't have all FP-related workloads done by the GPU but you only need it for some [and that too depends on the software at hand]). So the modular approach makes a lot of sense.

The desktop is stuck in this weird middle ground for them at the moment but only for the enthusiast. AMD shipped a lot of APUs on the desktop but that doesn't really appeal to the enthusiast. With server workloads becoming more and more parallel the desktop typical workloads are taking a back seat. While desktop workloads often cross over with laptop workloads (it's why SB/IB do well on both), the laptops benefit from the on-die GPU much more significantly (also why SB/IB have an on-die GPU).

So, yea, AMD's chips suck ass for the desktop but they actually make a lot more sense in their intended markets.

In the future yes. Like Netburst was for Intel, I see Bulldozer as a stepping stone for AMD and a foundation for great things down the road. And yes, aspects of the architecture and many of it's features are good for their intended markets, but its not enough to overcome their failings in today's market, in each segment with today's applications.

AMD needs to turn a profit today. Intel can afford to play long term games of chess with the CPU market and AMD can't. One of the reasons why Intel is so successful is that their products work well for the time they are released, and rarely if ever take a step back. Architectural issues which may not pay off immediately, are generally not implemented in a big way if it will compromise performance too much at the time of release. Intel realizes you have to be adaptable in the current market while looking ahead. AMD seems to only look off into the distance day dreaming of possible futures while tripping over things in their path at the time. Intel has screwed up once and learned valuable lessons by doing so. The Netburst architecture had features which eventually paid off, but at the time they lost the performance crown in nearly every area. They had to rely on their reputation and brand name to continue to sell units (as well as unscrupulous business tactics) and in order to survive during this time. Intel wasn't nearly as adaptable at the time and while their IPC was adversely effected by their brute force approach, their tactics would have worked if the architecture actually scaled the way it was intended to. We'd be looking at processors from Intel which would probably more closely resemble Bulldozer than Ivy Bridge and run at 3 times the clock speeds of today's fastest Ivy Bridge CPUs if it had.

But again, Intel made mistakes, acknowledged them and learned not to repeat them. They could stand to lose out during that generation because they had enough cash and still turned a profit despite having the inferior technology. AMD simply can't do that.

Products which are too far ahead of their time often fail to sell because they aren't particularly useful in the present or because their drawbacks paint them in a negative light. Few people truly understand the technology they use and don't care about what it will turn into. No one wants a car engine that you can't easily fuel, and performs like shit today even if it will one day power every car on the road and every air craft in the sky if it isn't as least as good as what they already have now at a similar price point. Consumers are only concerned with the here and now. Thinking about what could be and what might be is a task best left to the scientists and engineers who will one day bring us these new technologies in a form which is palatable at the time.