AMD Details Upcoming Bulldozer Architecture

Vigile writes "AMD is taking the lid off quite a bit of information on its upcoming CPU architecture known as Bulldozer that is the first complete redesign over current processors. AMD's lineup has been relatively stagnant while Intel continued to innovate with Nehalem and Sandy Bridge (due late this year) and the Bulldozer refresh is badly needed to keep in step. The integrated north bridge, on-die memory controller and large shared L3 cache remain key components from the Athlon/Phenom generation to Bulldozer but AMD is adding features like dual-thread support per core (but with a unique implementation utilizing separate execution units for each thread), support for 256-bit SIMD operations (for upcoming AVX support) all running on GlobalFoundries 32nm SOI process technology."

Comments

Re:Nobody cares.

[Anonymous Coward]

And why, exactly, should/do we not care? This is akin to the announcement of i7 or Sandy Bridge. Maybe if you don't care you shouldn't be reading this story.

Re:Sweeeeet nectar

[Cornelius the Great]

No pricing nor benchmarks. The article is purely a discussion about the architecture.

Not much new information

[kg8484]

Compared to such articles as AnandTech's coverage of this in November 2009, I don't see much new information. Perhaps the key bit, and this is glossed over but you can tell from the slides AMD gave them, is the difference between the bulldozer and bobcat cores. The bulldozer cores contain the two integer units that have been revealed before, but the bobcat core only has one but it still implements hyperthreading.

Re:Not much new information

[Rockoon]

Just to be clear, when you say "integer units" you mean "integer schedulers" and not actual integer execution units, of which even the old Athlon's had 3 per core (and that hasnt changed since then.)

Unlike Intel design, with highly asymmetric execution units, AMD's have had 3 symmetric integer execution units per core since the original Athlons. Its actually a pleasant breeze to write hand-optimized integer code on AMD's.

This new design looks (in the diagram) like it actually has 4 symmetric integer execution units per integer scheduler, with the bulldozer having 2 schedulers per core while the bobcat only having 1 per core (I would guess that the logical cores are alternated on rise-and-fall states of the clock on the bobcat, and the diagram certainly makes it look like that is the case.)

Each seem to have two wide floating point execution units, so the floating point performance of both bulldozers and bobcat's are probably equivalent.

What I think AMD has done here is that with the bulldozer, in integer performance it is going to behave like it has 2x the number of real cores. So an 8 core (16 thread) chip will perform much like an 8 core CPU in floatng point work, but much more like a true 16-core CPU in integer work. This should give it a large advantage over Intel in integer work in equal-core comparisons, but the floating point performance will still lag behind Intel.

Re:Not much new information

[imgod2u]

I believe Bobcat's 2 FPU paths are 64-bits wide. For a total of 128-bits. It initially will not support the 256-bit AVX instructions that are coming with Sandy Bridge and Bulldozer.

Its ALU's also appear to be significantly different than Bulldozer. With only one of the integer units can support multiplies and only two of them can support arithmetics. Two others (using a different scheduler) are load/store units. Bulldozer doubles the ALU resources (but not the number of schedulers) compared to Bobcat. So each scheduler has access to two AGU's, one ADD/DIV ALU and one ADD/Mult.

I was never a big fan of the 3x symmetric ALU's in the Athlons. When it comes to integer intensive code, having a ton of independent ADDs or MULs that I'd need that kind of parallelism for was rare. And the latency (compared to a sane design like Core at least) were significantly higher due to the units being multi-purpose. In either case, with the introduction of SSE2, one could use SIMD if one had a throughput heavy workload anyway.

Bobcat and Bulldozer appears to have moved in the right direction here. I really do like Bulldozer's approach to multi-core and think that with some extension, this could make into very interesting CPU/GPU hybrids as well. Although you could argue it's just another version of SMT similar to Hyperthreading, only with a wider back-end intended specifically for multi-threaded processes.

Re:Not much new information

[Rockoon]

I was never a big fan of the 3x symmetric ALU's in the Athlons. When it comes to integer intensive code, having a ton of independent ADDs or MULs that I'd need that kind of parallelism for was rare. And the latency (compared to a sane design like Core at least) were significantly higher due to the units being multi-purpose.

In the Phenom II design the latency of most of the register to register integer instructions is exactly 1 cycle just like the i7. The units being multi-purpose is not a latency sacrifice at all, although maybe the original Athlons had poor latency for another reason and Agner Fog's reference actually indicates that most of the register to register integer instructions on even the early K7's also had 1 cycle latency.

Even in mem,reg operations, the Phenom II beats out the I7 in latency on many operations (ex: xor [ebx], eax .. 6 cycle latency on i7, 4 cycle latency on Phenom II) .. where the I7 really shines the most is in reading directly from L1 into registers (2 cycle latency vs 3 cycle latency), with a massive 50% advantage on one of the most common operations..

Mmm

[elsurexiste]

Call me whatever you want, but the only reason AMD is still alive and well is because they've been innovating and building good products for a while now. Itanium, anyone?

Re:Mmm

[Amouth]

the first AMD64 CPU shipped in April 2003

the first Itanium Shipped June 2001

So the AMD was 23 months late - all they did was tack on to existing x86 where as Intel was trying and did develop a whole new architecture.

almost all of the complaints for the Itanium being slow was due to it having to emulate x86 for software that was not written specifically for the IA64 - Code that was and is written for IA64 runs fast as hell and there is a reason why they are still used today - just is specific applications.

Intel's failure was due to them trying to jump to a whole new computing architecture and expecting programmers to go with them - instead programmers resisted and AMD jumped on that by just extending the existing x86.

Development on what became the IA64 started in 1989 by HP and Intel was brought in in 1994 and the first implementation was in 1998 - hell it is the reason we don't see Alpha's anymore.

AMD64 started in 1999.

So in computing terms AMD had many generations to watch Intel actually Innovate - and then take the short cut to market. Please note I'm not putting AMD down for AMD64, I'm just pointing out that you can not compare the success of it VS the Itanium because they are not the same by a long shot.

Also if you want to learn something new - read up on why IA-64 is so different form x86 and you will see why it is worth investing in. Not for the current project but rather for the knowledge gained by doing it. You would be surprised how much of the R&D that went in to the Itanium is currently running in your newer computers and servers.

Re:Mmm

[petermgreen]

Itanium, anyone?
Yes some time ago intel was screwing arround with itanium (which hardly anyone wanted because it ran x86 code so badly) and netburst (which was slower per clock than a P3) while AMD was pushing ahead with the hammer architecture.

However since core 2 and especially with nahelm (where intel moved to a point-point architecture from a shared FSB architecture) intel has gradually regained the lead starting with the single sockets and gradually moving up to larger platforms. AMD is resorting to throwing cores at the problem in a desperate effort to make up for thier poor performance per core but the trouble is that typical desktop workloads can't really load up four cores, let alone six.

Re:Mmm

[MechaStreisand]

Bullshit. The Itanium is slow as shit because Intel didn't bother to give it out-of-order execution like every other modern processor has. As a result, it is only fast on DSP-like operations and slow at everything else. Out-of-order execution is essential because the compiler can't know at compile time exactly where everything is going to stall: it's provably impossible.

Remember, this is the same company that designed the P4 without a barrel shifter.

Re:Mmm

[imgod2u]

There are plenty of things to learn from Itanium, specifically, what not to do if you want a good general purpose processor. For one, you don't make processor performance so incredibly reliant on instruction scheduling that the biggest compiler team on Earth (Intel's compiler group) couldn't make it run fast on anything except a small subsection of problems.

Secondly, when attempting to gain ISA adoption, making it an exclusive ISA that only you have control and rights to use is a big no no. Sure, it'd be heaven for Intel to be the sole supplier.

And lastly, process and iterations mean more for performance than any fancy ISA. Itanium is consistently one or two process generations behind its x86 counterparts and consistently one or two micro-architectural iterations slower (it takes 2 revisions of the Core micro-arch before Itanium comes out with one).

You can have as clean and fancy of an ISA (which IA-64 was not, btw) as you'd like but implementation matters far far more.

In the end, it wasn't fast enough (the best it ever did was match its x86 counterparts) and it didn't have any other advantages to warrant the switch.

Now, ARM on the other hand....

Re:Mmm

[coredog64]

Development on what became the IA64 started in 1989 by HP and Intel was brought in in 1994 and the first implementation was in 1998 - hell it is the reason we don't see Alpha's anymore.

The reason we don't see Alpha anymore is that Intel coerced HP to buy up Compaq and kill it off by offering to assist HP in porting HP-UX to Itanic.

AMD's stagnant?

[iamhassi]

"AMD's lineup has been relatively stagnant while Intel continued to innovate with Nehalem and Sandy Bridge (due late this year) and the Bulldozer refresh is badly needed to keep in step."

AMD just came out with Six-Core processors for $200, how is that stagnant? Intel's only 6-core processor is still $1000

Re:AMD's stagnant?

[TrisexualPuppy]

Likely another Intel fanboy trying to spread FUD about the company that he doesn't like and at the same time getting his username posted on the front page.

AMD may not have the resources that Intel does, but it isn't as though Intel is walking AMD around on a leash. This mindset gets annoying after a while.

Re:AMD's stagnant?

[SQL Error]

AMD are selling six-core dual-socket CPUs for $200 now. They're not quite as fast as the Xeon 5500/5600, but the price/performance is awesome.

Re:AMD's stagnant?

[blair1q]

"AMD's lineup has been relatively stagnant while Intel continued to innovate with Nehalem and Sandy Bridge (due late this year) and the Bulldozer refresh is badly needed to keep in step."

Likely another Intel fanboy trying to spread FUD about the company that he doesn't like and at the same time getting his username posted on the front page.

The facts in that quote were presented clearly. AMD is a generation behind on architecture, trying to get comparable performance by multiplying old cores, while Intel has been advancing architecture and multiplying cores at the same time. For about 4 years now, Intel has had 2-4 chips performing at levels above anything AMD could produce.

It remains to be seen if Bulldozer will put AMD anywhere near at-par on a performance/core basis, but it's not 2002 any more, and AMD has no hope of a performance lead.

Re:AMD's stagnant?

[Anonymous Coward]

Actually, AMD's has a great chance of beating Intel in the future. You fail to recognize that AMD has ATI now and they are going to be fusing CPU's and GPU's onto the same die in the future. They benefit from the experience and IP of ATI. Intels graphics capability so far has been a joke.

Re:AMD's stagnant?

[cynyr]

but not per $, which is the whole point. Sure i can build a screaming rig using a $1500 intel CPU, and a $400motherboard, and then toss in the ECC ram that board needs... and all of a sudden i could have bought a honda civic....

Or i could get 80-90% of that same rig, in certain loads 120-150%, for $500 from AMD.

Re:AMD's stagnant?

[h00manist]

Funny how so many people say competition is great, then turn around and start promoting the strategy of eliminating the competition, supporting whoever is "winning", etc.

Re:AMD's stagnant?

[ArcherB]

Intel's hexacore offering features hyperthreading technology, which allows each core to execute two threads simultaneously. This means that Intel's hexacore chips actually have twelve logical cores, while the AMD hexacore chips only have six logical cores.

I think you may be misunderstanding what hyperthreading is. A processor (or core) can only execute one instruction at a time, hyperthreading or not. All hyperthreading does is allow for two sets of instructions to be queued up, so if one thread (or queue) gets hung up for whatever reason, like waiting over a cache miss, the other instructional thread can proceed, rather than patiently waiting in line.

Think of it as one of those tumbling thingies you have to pass through to get into Six Flags or the subway. It's like that, but hyperthreading has two lines instead of one. If one moron has to stop to find his ticket at the front of the line, the other line may move until he finds it.

Your number of physical cores comparison is meaningless...

Um... no. I believe your "virtual" core comparison is meaningless. I'll take a quad core anything over a dual core hyperthreaded-anything-else any day, thank you. Virtual cores don't mean shit until a thread stalls.

and actual performance benchmarks show that the Core i7 980X is more than twice as fast as the AMD Phenom II X6 1055T. [1]

From the site you linked:

Intel Core i7 980X @ 3.33GHz: Score of 10,325 at $989.99*
AMD Phenom II X6 1055T: Score of 5,146 at $194.99*

Hmmmm... Twice the performance at over 5x the cost. Strange, I don't know why you chose that AMD chip. It's odd that you would choose the fastest Intel chip and a middle of the road AMD Chip. Why not this one?
AMD Phenom II X6 1090T: Score of 6,057 at $289.99*.

Oh, I know. Thenyou wouldn't be able to use the 2x faster line. I get it now.

Here, take a look at THIS chart and pay attention to the price/performace graph. You'll see that your chip performs about 2.5x less than the AMD Phenom II X5 965 when price is a consideration. Oh, and for nearly everyone that is not living off their mommy's credit cards, price is a consideration.

Re:AMD's stagnant?

[MBGMorden]

If you want fast graphics then you buy a discrete graphics card. If you're using integrated graphics you don't much care whether it's a crappy ATI chip or a crappy Intel chip because it won't run modern games at any reasonable speed either way.

That's conventional wisdom, but conventional wisdom doesn't always hold steady in the computing market. 15 years ago what you said there was true for both audio chips and network cards. Anybody who wanted one that was half-way decent bought a discrete unit because those performed well, and the hokey versions that you might find integrated were pretty much junk.

Today? All but a few holdouts and professional level users just use the integrated network and sound, because for your average user - even your average power user - the integrated stuff is plenty good enough.

I'd wager that in less than 8 years your statement of "If you want fast graphics then you buy a discrete graphics card." will sound just as outdated and clueless as "If you want to crunch numbers faster than you buy a dedicated math co-processor.".

Re:AMD's stagnant?

[blair1q]

That is how AMD stays in business, by cutting its prices well below the average market price for the performance rating.

But there is a large chunk of performance rating they can't even approach.

Here's last year's numbers (didn't see this year's in the first page of google results), which should give you an indication of why AMD went looking for more performance from each chip. I'm still not expecting Bulldozer to get AMD up to the top. They might match the second- and third-place chips from Intel, but they haven't come up with anything that can outright beat what Intel has fielded.

And AMD's pricing policy costs it a lot of money, since its production costs are much higher than Intel's. They've never had the kind of production efficiency Intel had, and now they have sold off their fabs and are contracting their parts out to GlobalFoundries, who charge a cash profit on each one and still aren't as efficient as Intel.

Re:AMD's stagnant?

[MBGMorden]

There's an infinite capacity to use floating point arithmatic too, but we abandoned the separate chip for it idea long ago. FPU's these days are still getting faster with each chip - no limit on processing power was hit. We simply got to a point where a completely capable FPU could be bundled in with the CPU and it's performance was sufficient for most users.

Imagine this scenario: the integrated solutions don't suck. Instead of being virtually useless for 3D graphics, they have performance about equal to the mid-line $150 to $200-ish cards of today (and let that scale for whatever cards meet that definition of the time). You can get better performance, but it's going to take huge full-length cards running SLI or the like, and it's going to take several hundred dollars to beat your standard integrated solution.

My wager is that 95% of the people who currently buy discrete chips would accept integrated at that point. The chips would still get faster over time, and there still might be a few extreme solutions available, but the average user wouldn't need them anymore. My guess is we'll get there quite soon. And if you're asking why the chip companies would want to sell us 1 chip where they previously sold 2? Simple answer: market competition. If AMD can push out a chip as fast or faster than Intels that also has an integrated GPU that rivals discrete solutions, then they'll take a lot of business from Intel. That's all the motive they need.

Re:AMD's stagnant?

[imgod2u]

No, what happened was that the most FP intensive tasks (rendering, 3D modeling for games) moved to another dedicated chip (the GPU). Bigger and better compute capacity there has not stopped being in demand ever since and shows no signs of slowing down.

The only thing left that were really compute intensive on the CPU were things like video transcoding and precise (production quality) 3D rendering due to the lack of double-precision support in GPU's as well as the difficulty of using them for compute (i.e. writing back to memory) purposes. For the vast majority of common consumer tasks, there really wasn't much demand for ever beefier FPU's.

Also, just as what may eventually be true for GPU's, there are advantages performance-wise of moving the FPU closer to the CPU. In the case of the FPU, it was much easier to tie in FP intensive tasks with control structures the CPU provided. I would suspect that in the case of the GPU, it will be removing the latency of the PCI bus and the need to copy memory between main and GPU memory, potentially allowing a lot more interaction between the physics and AI with the rendering process.

Bulldozer?

[Megahard]

Sounds like a slow-moving behemoth. Not the best choice for a name.

Re:Bulldozer?

[rotide]

Or it pushes everyone/everything else to the wayside. I guess it depends on your interpretation.

Re:Bulldozer?

[sideslash]

I think people's mental impression will vary; yours is typical of an office worker. To someone employed in construction or agriculture, a bulldozer is a symbol of getting huge amounts of work done in a very short time. This reminds me of an ad I saw some years back for an "object oriented database", where they showed a photoshopped race car with a tractor on the back end. Their marketing message was "Why do you have a sluggish RDMS on your web app's back end?" I found it hilarious, because my reflexive response was to ask, "Why is that totally useless racecar pasted on the front of that excellent looking tractor, the kind of vehicle that is used to grow all the crops that feed the world?" :)

Re:Bulldozer?

[Minwee]

Perhaps the team at AMD had been drinking heavily the night before.

At eight o'clock on Thursday morning Arthur didn't feel very good. He woke up blearily, got up, wandered blearily round his room, opened a window, saw a bulldozer, found his slippers, and stomped off to the bathroom to wash.
Toothpaste on the brush -- so. Scrub.
Shaving mirror -- pointing at the ceiling. He adjusted it. For a moment it reflected a second bulldozer through the bathroom window. Properly adjusted, it reflected Arthur Dent's bristles. He shaved them off, washed, dried, and stomped off to the kitchen to find something pleasant to put in his mouth.
Kettle, plug, fridge, milk, coffee. Yawn.
The word bulldozer wandered through his mind for a moment in search of something to connect with.
The bulldozer outside the kitchen window was quite a big one.
He stared at it.
"Yellow," he thought and stomped off back to his bedroom to get dressed.

If you don't know what that's from, get off the computer and go to a library.

Re:Bulldozer?

[RaymondKurzweil]

"Why is that totally useless racecar pasted on the front of that excellent looking tractor, the kind of vehicle that is used to grow all the crops that feed the world?" :

Maybe it's because the people that were selling that "object oriented database" were far more honest than you assumed.

Re:Will it be compatible with AM3?

[WiglyWorm]

It will be drop in compatable with AMD server boards. At home, it will be AM3+.

Almost certainly not AM3

[drinkypoo]

It's a whole new architecture and will almost certainly require a new socket. ISTR the article saying nothing about memory technologies as well. The good news is that a new architecture on the horizon which almost certainly requires a new socket makes it seem less likely that AMD will bring out another socket before then.

Hamster Farm Analogy

[lymond01]

Processor Speed: Very fast hamsters on well-oiled wheels
Multiple Cores: Many well-oiled wheels
On die memory controllers: dangled cheese
Cache: water trough next to the wheel
L3 Cache: Camelback packs for each hamster
Shared L3 Cache: This is where the real innovation comes in and won't be defined as patent is pending.

Re:Nobody cares.

[Jesus_666]

Well, it's a new computer part. It computes. The computer I already own computes perfectly well so I don't see how a new one is any different. You act as if computers were a field of interest or even an industry.

What's next? People getting excited about new cars?

Ghetto computer

[mark72005]

...which is all that the 99% of people who would be classified as "ghetto users" need anyhow.

Re:Nobody cares.

[hairyfeet]

Well I personally care because it affects my prices and thus the prices I charge customers. After the Intel douchebaggery came out, along with the Nvidia bumpgate fiasco I switched all new builds to AMD exclusively. The bang for the buck is just incredible, with OEM triples for $60 and quads for $99, and their IGPs frankly kick the dog snot out of Intel, with them even able to run games like Bioshock acceptably.

This new part is gonna be nice because it will allow for really nice SFF HTPCs, and with a Radeon onboard it will accelerate (I personally use Media Player Classic Home Cinema which works great with AMD GPUs) all the major video formats out there, thus taking the load off the CPU and allowing a nicer customer experience. I have found even the bottom of the line Athlon II duals thanks to the Radeon onboard give a really nice experience for the customer, and allows them to have all the Windows 7 Aero features without needing a discrete GPU. This will also be great for netbooks, although I haven't had any customer complaints from the new Amd Neo duals with Radeon onboard, this will lower power and thus lead to longer battery life.

So yeah, there are some of us nerd that actually care about such things. I like the fact that with AM3 being backward compatible I was able to replace my 7550 dual for a 925 quad without having to trash the box and start over, or that pretty much any of the 125w motherboards can have a 6 core dropped right in. This new Bulldozer will allow for PCs which use less power, generate less heat, while still giving a good user experience and great bang for the buck. While TFA says Bobcat is more for the low power I bet I end up using Bulldozer more, as folks like having that extra performance and the AMD "drag and drop" video transcoder makes it easy for customers to convert videos for their portable devices.

If you haven't checked out AMD in awhile you really should. Yes Intel has the absolute speed crown, but for everyday tasks and even heavy lifting like transcoding I've no complaints, and I certainly love having a fully loaded quad with 8Gb of RAM for under $650.

Re:Sweeeeet nectar

[Nadaka]

paying 1/3 as much for more than 1/3 the computing power is a viable strategy known as "value based judgment".

At any given price point where there exists an AMD processor, there are few if any intel CPU's with equal or better performance.

The i5 750 and i7 920 are among the very, very few intel chips that compete with AMD on value (performance / price).

Re:Sweeeeet nectar

[ElectricTurtle]

Benchmarks certainly testify that the AMD 3.2 ghz hexacore is not three times slower than the Intel. At three times the price even twice as fast is a rip off, and Intel isn't even that far ahead.

Plus, Windows Server support is over...

[BUL2294]

Even Microsoft has thrown in the towel with IA-64 given the scalability of AMD64 (err, x86-64) in Xeon & Opteron processors. Windows Server 2008 R2 is the last version to support IA-64...

http://www.microsoft.com/windowsserver2008/en/us/2008-IA.aspx

Re:And to some people

[blair1q]

If you're running any Windows since XP, you're using all of your cores all the time, and it's benefitting you. You may not get all of them working on the same task, but the fact that the other cores can do background and response tasks while your foreground task pegs one of the cores is always a bonus. If it doesn't show up in outright speed of completion, it hides a vast array of niggling little delays that make things jerky.