AMD Bulldozer and Fusion

[nbohr1more]

Since Jaxa has stirred the topic, I thought I would share my thoughts on AMD at length.

 

Before the Bulldozer released, AMD was repeatedly emphasizing that this processor would NOT be remarkable at traditional x86 workloads. Their assertion has been that 99% of those workloads are already serviced by half-decade old processors and the future of CPU workloads is "parallel". AMD even acknowledged that their upcoming design would be short on Floating Point processing because their strategy was to move Floating Point workloads to the GPU.

 

Then when the processor was released all the Benchmark sites sensationalized the "shocking" benchmark results for benchmarks that AMD already had admitted would be weak areas for the new architecture.

 

The biggest contention point: "CPU heavy games that need lots of Floating Point math".

 

It's simply tragic that all these hardware pundits have lost the thread.

 

AMD is moving to further integrate their GPU products into their CPU. When that comes to fruition, AMD CPU's will have MOUNTAINS of Floating Point power. It will be the most media rich CPU since the heyday of the PowerPC and it's AltiVec instructions. Right now, the AMD CPU might be a sliver shy of Floating Point but you can buy a GPU to remedy that deficiency. All it will take is for software developers to make more strides towards GP-GPU and the modest advantage that Intel has in Floating Point will be irrelevant. When you upgrade your Bulldozer to a Fusion version, you've got a damned super-computer on a chip.

 

The only poor decision by AMD in this grand plan is that they sacrificed the cost of hand-tuning the transistors on their latest CPU in favor of using software automated circuit layouts (thus wasting MILLIONS of transistors).

 

Sadly, is seems that the bad vibes from all these reviews are rubbing off on the flippant management over at AMD and I fear that they will scale back or cancel this grand engineering vision. The next process node was just cancelled (in favor of a delay to the one after) and I hope that Fusion wasn't scrapped along with it.

 

I hope that AMD survives this brutal period of poor esteem in the hardware enthusiast community and can at least get one valiant foot in the door for a true heterogeneous parallel core x86 CPU (that takes IBM and Toshiba's "Cell" processor approach to it's ultimate conclusion). If at least a DX11 low-mid-range GPU becomes the new "minimum" amount of on-CPU floating point power, then games can become wildly more rich in physics and AI. We can only hope that this trend would keep on escalating.

 

Let's not even think of how poorly the Opteron line is doing. (Though, I'd still like to see how it fares in massive parallel HPC benches,)

[lost_soul]

I too have been reading about Bulldozer. AMD is important for budget-minded people like me, because we can get a Phenom 2 955 for $115, or a 1100t for $170. Problem is, the benchmarks on Bulldozer failed to set the world on fire. It even fell behind the 1100t in a few benchmarks. This is a problem if they expect people to spend 60-70% more for the new chip. Sadly, it isn't the 90s anymore and we aren't seeing massive jumps in CPU performance with every new architecture. Those were great times. I remember playing Descent on a 386 40 and then a 486DX2 machine for example and being blown away by the difference.

 

My backup machine Phenom 2 955) compiled Doom 3 in about 45 seconds. I was pretty impressed with that, because I was expecting it to take 5 minutes or so.

Edited December 2, 2011 by lost_soul

[nbohr1more]

That's just it though. If the Fusion plan bears it's fruit, there will be a colossal jump in Floating Point math resources. Intel's modest architectural changes and process changes are not offering anything close to that potential.

[lost_soul]

One area where things still need to improve is efficiency. If I'm sitting here listening to music and reading articles on-line, I don't really need SIX CPU cores running simultaneously. This is one area where Intel has AMD beat. I like to have a computer on 24/7 for doing a variety of tasks and sometimes I choose my netbook because it doesn't have a 600 watt PSU and it won't use much energy at all.

 

This is where software efficiency and optimization is key. I worked on someone's Athlon 64 4xxx system (1 GB RAM) today and you know what? The Athlon 64 3400+ I was given last week for free with a GB of memory runs circles around it. The one I was given has XP, the one I worked on today had Vista.

 

I'm interested in a Fusion netbook. There was a video on Youtube of an AMD netbook playing Doom 3 at 640x480. In a year or two, we should see netbooks that can play TDM. That will be when I abandon my Atom N270 based one. I'll stick to playing Q3a and GLQuake on that one for now.

Edited December 2, 2011 by lost_soul

[jaxa]

Before the Bulldozer released, AMD was repeatedly emphasizing that this processor would NOT be remarkable at traditional x86 workloads. Their assertion has been that 99% of those workloads are already serviced by half-decade old processors and the future of CPU workloads is "parallel".

 

...

 

Then when the processor was released all the Benchmark sites sensationalized the "shocking" benchmark results for benchmarks that AMD already had admitted would be weak areas for the new architecture.

 

It may be the case that expectations were too high, but AMD marketing did appear to talk up Bulldozer/FX-8150 to a degree that was at least disingenuous. Where are you supposed to get your info then? Internal memo leaks? Or maybe some benchmarks? On release, FX-8150 was beat by lower priced Intel chips. I think some of these benchmarks do take into account discrete/integrated graphics as well, but remember that adding a GPU is also adding cost.

 

http://www.tomshardw...ting,13701.html

 

Following the Bulldozer criticism came 10% workforce cuts by AMD, which is just one reason I see it at risk. It's interesting that these jobs appear to be PR/marketing though: http://slashdot.org/...-pr-departments

 

One area where things still need to improve is efficiency. If I'm sitting here listening to music and reading articles on-line, I don't really need SIX CPU cores running simultaneously. This is one area where Intel has AMD beat. I like to have a computer on 24/7 for doing a variety of tasks and sometimes I choose my netbook because it doesn't have a 600 watt PSU and it won't use much energy at all.

 

So apparently Bulldozer is not as power efficient as it should be or as people expect from AMD. I'm not sure this is a failing of having 4-6-8-X cores though. Even if you claim to run 2-3 applications you're probably running a couple dozen processes... some of these can be run in parallel rather than waiting in line. But I think with the energy usage of two chips being equal and assuming an OS/programs better suited to multithreading, an 8-core could be more powerful than say, a dual-core. It just seems to me the FX-8150 is not that chip, or OS can't exploit what it offers. And I don't think there's any reason why it shouldn't be. The chips sitting at the top of benchmarks all have 4+ cores.

 

Here's some of the questions I have:

 

* Bulldozer is a new architecture and represents a strategic shift. Benchmark sites blasted it for poor marks, energy efficiency, you name it. Is this a result of:

a. Early jitters, early and not necessarily systemic design flaws, bad firmware / OS support to begin with? All indications show that Bulldozer will continue to see improvements as planned by AMD.

b. A disconnect between what the Bulldozer/Opteron line excels at in certain benchmarks and the real world applications (particularly gaming, considering Bulldozer was/is being marketed as a home consumer processor, although some of what I've read indicates that there was early confusion over Bulldozer's intended market segments). Could this be improved by further development of parallelism in applications?

c. Systemic problems with the design, poor decisions, or perhaps an inability to keep up with Intel in terms of cutting-edge research?

d. None of the above. Hardware benchmarking sites are poorly qualified to measure Bulldozer's capabilities and should be pairing Bulldozer with a GPU for more accurate results. Btw, am I correct in assuming that integrated graphics is a strategy purely targeting budget market segments and perhaps price/performance? i.e. discrete graphics will always exceed the performance of integrated graphics forever, but not necessarily in price/performance?

 

* Some of the Bulldozer criticism is centered around the 8 core claim (part of the marketing is that FX-8150 is the world's first consumer desktop 8-core processor). This is because each pair of cores shares some cache and other components. Is this design ultimately a good move or not for AMD? Let's assume that OSes become better able to use the Bulldozer architecture. What happens when Intel puts out a consumer 6-8 core chip? Is shared cache a liability, or is AMD just doing something else badly? For more info check out the text around "What is a core, anyway?": http://www.tomshardw...0fx,3043-3.html

 

* nb you clearly saw the launch of Bulldozer in a different light than many hardware enthusiasts. I remember the publicity stunts like the Guinness world record overclock of the FX-8150 (over 8.4 Ghz) which was later pushed even further twice by some enthusiast. This was using liquid helium of course, not part of your average setup, and an amusing diversion. I think there were also some hints that the chip would be killing the i7-2600 and stuff like that. When we take a grain of salt, should we be swallowing it for the literally decimated AMD PR/Marketing people, or AMD as an entity? Did AMD believe their own hype or were there multiple lines of communication, one of which you picked up on? Examining a statement such as "this processor would NOT be remarkable at traditional x86 workloads"... that wouldn't be a surprise considering the novelty of 8 cores for most people and that parallelism is an ongoing transition in software design... I've seen this problems with my A6-3400M for instance. But does this statement do two things: a) does it hold water considering Bulldozer is a consumer desktop processor. Maybe we should call it a "high end budget" processor rather than a "gaming" processor? b.) does Bulldozer do anything particularly novel to the point at which we can forgive its comparatively weak showing? The statement "multicore is the future" is so self-evident today... what's so revolutionary about 8 core over quad core? Moreover, for the noob in me, when we look at clock speed we see that FX-8150 is 3.6 GHz per core, implying that 8 cores can run at that clock speed at the same time (turbo is 4.2 GHz, but some cores have to be off, I view turbo as an interim step for the transitioning state of parallelism in software). What is it about an i7-2600's 3.4 GHz x 4 cores that allows it to match or thrash the FX-8150?

 

* Without considering the multiple factors that go into picking a CPU or CPU/GPU, such as pricing and wattage, what do we like for benchmarks? What do you think of http://www.cpubenchmark.net/ ? Or is it about TDM FPS, boot up times, or something else.

 

* How do we address Bulldozer energy efficiency criticism?

 

* What is Intel doing so wrong in the budget segment for laptops and desktops that is allowing AMD to continue to thrive? Anybody else think that this could change very quickly? Could Intel's great mass of mindshare and cash allow it to outcompete on all fronts (ARM, AMD, VIA, everywhere)? Consider the costs of continuing Moore's self-fulfilling prophecy, and that in similar sectors like hard drives the competition is buying each other up.

 

I haven't brought up much related to moves towards the mobile market, etc., but that's not Bulldozer's area. I guess that's more of Fusion's area.

 

I'd like to discuss more on this later but I'll be busy for the rest of the day. I'm interested in exploring this topic further so everybody can learn something. See you on Saturday.

Edited December 2, 2011 by jaxa

[7upMan]

Just two quick answers: AMD is working together with Microsoft to release an updated version of the Windows scheduler, so that workloads will be handled more efficiently on a Bulldozer system. This will allow for up to 5 percent performance increase. "We are also working with Microsoft on a scheduler update for Windows 7 that will be available soon." While this will probably not do much to mend and IPC that's on par with the first Phenoms, it might bring some relief to this struggling CPU.

 

Also, Intel has yet to release a graphics system that can compete with anything that's already on the market. If you remember, the grahics drivers are universally so woefully bad that gaming is mostly next to impossible. AMD is way ahead with its APUs (Llano and Bobcat). (By the way, my mother has a 15" HP laptop with a Bobcat APU, and it works like a charm.) Intel's Atom plattform is sub-par if compared to AMD's Bobcat plattform. This proves (again) that throwing money at something that is not your core competence will not necessarily make you a leader in that segment.

[demagogue]

I'm not all that tech minded, but what I mentioned in my previous comment was I was going back and forth between these two computers, both of them an HP dv6 with 15.6" screen, and since part of it was an Intel vs. AMD issue I may as well mention it:

 

1. AMD A8-3500M (1.5GHz) + AMD Radeon HD 6620G - http://www.amazon.co..._pr_product_top

2. Intel i3-370M (2.4GHz) + ATI Mobility Radeon HD 5650 - http://www.amazon.co...C/dp/B0044XU4OG

 

I found a deal for both of them at $450, so price wasn't an issue. I went with #2 for two reasons... One because even with the lower number, the Radeon 5650 was getting consistently rated with higher FPS for games (I understand the tested computers were with an i5 430M. Even if the i3 has less performance, it's still not that much less),

http://www.notebookc...50.23697.0.html

http://www.notebookc...0G.54675.0.html

And I don't even play many modern games anyway. Both cards can run Doom3 at like 100FPS on Ultra setting, which is about the most I care about. But the difference in time for heavy-CPU tasks adds up, and I cared more about that... Like if I wanted to reencode a 1.5GB Fraps walkthru video to be a manageable size, the difference in time would be in minutes. And I just didn't like the sound of a 1.5GHz CPU (nor that they fudged in advertising it as 2.4GHz, but that's just when it's overclocked, or something. The stats page for the CPU itself says its base is 1.5GHz).

[jaxa]

And I just didn't like the sound of a 1.5GHz CPU (nor that they fudged in advertising it as 2.4GHz, but that's just when it's overclocked, or something. The stats page for the CPU itself says its base is 1.5GHz).

 

Looks like you goofed. I have the A6-3400M which is a shade under yours at 1.4 GHz @ 4 cores. 2.4 GHz (2.3 for me) represents "turbo," a technology used by both AMD and Intel to automatically overclock one (maybe two?) cores when others are shut off. If you overclock yourself, turbo doesn't apply.

 

PassMark for AMD A8-3500M APU with Radeon HD Graphics is 3034.

PassMark for Intel Core i3 370M @ 2.40GHz is 2219.

 

These numbers aren't perfect (my A6 has a score of 3514, probably because of having many more samples taken) but when you just look into the specifics of each processor you immediately see that you're comparing quad and dual core cards.

 

http://ark.intel.com/products/49020/Intel-Core-i3-370M-Processor-%283M-cache-2_40-GHz%29

 

Also note that i3-370M doesn't use "Intel® Turbo Boost Technology". So despite the initial lower clock speed, right out of the box the AMD card can apparently match the performance of the i3 while parts of it are shut off. At this point you want to think about your applications' use of multiple cores. This answer probably would have been more favorable to lower core processors say, 2-3 years ago, but nowadays a lot of applications can take advantage of multiple cores, and CPU dependent ones (video conversion, Adobe XYZShop) and games are among the first to do so. I don't know what TDM is specifically capable of using, but that's set to change sometime now that source is a released.

[demagogue]

Oh well whatever. It's absurd to expect people to understand those damn charts when for every brand there's a million different CPU/GPU combinations.

 

And I'm getting like 100 FPS for Doom3 compared to what I was getting before, and all the applications I'm using are like 3 years old anyway. It's fast enough.

[MoroseTroll]

The Bulldozer is a server & workstation microarchitecture. This is where the problem is: many weak cores are very good for highly multithreaded software (the server &workstation markets), but the desktop and the mobile ones require strong single thread performance, where the Bulldozer failed. Of course, a higher frequency will partially compensate its poor single thread performance, but has its own price - energy hunger.

As for the Fusion concept I'd like to say that it is a nice one, but I'm sure that the OpenCL/DirectCompute infrastructure is not as mature and solid as we'd like to see. So, in order to see a real result of the Fusion, we'd wait for two-three years at least.

[nbohr1more]

Yes, the logical result of AMD's efforts is very awkward for marketing purposes:

 

AMD: We're gonna RULE the Server market!

 

Home User: What about us?

 

AMD: We'll release a home version.

 

Home User: Will it be any good?

 

AMD: Meh, it'll be fine for now, read these tech docs... Make sure to buy a GPU

 

AMD Marketing: Uh... Dude? It'll be awesome! (Hope they fall for that.)

 

Home User: Is that Marketing guy lying?

 

AMD: Pretty much

 

AMD Marketing: Wha! (I'm fired)

 

AMD: Fusion will be the awesome CPU, just get ready now

 

Home User: I'm nervous... Will you guys survive until Fusion is released?

 

AMD: We hope so.

 

So we wait and hope they survive. We also hope that OpenCL adoption goes up...

Edited December 5, 2011 by nbohr1more

[7upMan]

In the x86 world, AMD has been a technology leader in many regards, be it the integrated memory controller, the x64 extension or the merging of CPU and GPU. Hope they'll survive long enough to make everyone realize the benefits of the Bulldozer architecture - and to make the devs actually use the multi/many core potential that is still largely untapped.

[MoroseTroll]

Well, I'm quite sure that upcoming AMD's APU Trinity (with next to Bulldozer microarchitecture named Piledriver, 1Q12) will be very successful. It has powerful integrated graphics (up to 30% faster than the Llano's one) and more-less strong x86-cores (up to 20% faster than the Llano's ones). Of course, Intel's Sandy Bridge and Ivy Bridge have a much more powerful x86-cores, but their integrated graphics is way too weak (2-3 times). So, what would you choose?

[nbohr1more]

The question is of time-tables.

 

When will AMD "share" transistors from the GPU part as part of the FP unit in the CPU?

When will they gain traction for x86 extensions that use those extra FP capabilities (akin to the 3DNow! stuff of yore)?

When will OpenCL start cooking in the mainstream?

 

If the trends keep flowing to their natural conclusion, computing will be split between Server big-iron and Mobile platforms.

 

AMD has the advantage of a better graphic part but they lack the key low-power characteristics. They are doing better on power per mm squared than Nvidia though.

The magic spice that might keep them in the game is their ties to IBM and collaborative works to address new lithographic challenges. Thus far, they have not demonstrated the fruits of

that alliance. Global Foundries did not produce the quality of lithography expected (and needed) for the 32nm node... They certainly wont be taking advantage of any magic if they stick

to software designed layouts.

 

Will future Bulldozer variants be hand-tuned?

 

Was this generation the sacrificial lamb so that they'd have a product to market quickly at low development cost?

 

What would a hand-tuned Bulldozer perform like?

 

What would a hand-tuned Bulldozer "Fusion" product perform like?

[MoroseTroll]

When will AMD "share" transistors from the GPU part as part of the FP unit in the CPU?

Not in this life. Technically, this is not a problem, but there some drastically changes will be needed for the x86 ISA. Of course, AMD can do this, but the software industry likes open, multi-vendor standards. So, in this case, Intel and VIA should be get involved in the discussion, and I'm quite sure that Intel would disagree with any good thing AMD would propose.

When will they gain traction for x86 extensions that use those extra FP capabilities (akin to the 3DNow! stuff of yore)?

No, I'm sure that OpenCL is the way AMD will drive all the way .

When will OpenCL start cooking in the mainstream?

Technically, OpenCL is already cooking, because some software already supports this standard. All we need is more this software.

 

If the trends keep flowing to their natural conclusion, computing will be split between Server big-iron and Mobile platforms.

I totally agree.

[

]Will future Bulldozer variants be hand-tuned?

I believe, yes. The Bulldozer successors (Piledriver, Steamroller, Excavator) will be hand-tuned, I suppose.

Was this generation the sacrificial lamb so that they'd have a product to market quickly at low development cost?

I think, yes.

What would a hand-tuned Bulldozer perform like?

A little die, slightly faster cores.

What would a hand-tuned Bulldozer "Fusion" product perform like?

A medium x86-performance, a best performance of the integrated graphics.

[lost_soul]

The funny thing is, the prices on the Phenom 2 chips are going up now. That's a pretty dumb move. One could justify a Phenom 2 because the prices were so low, but now you might as well get an I5 instead.

[nbohr1more]

Holiday pricing is in effect.

 

Plus... many users are suffering lesser AMD chips for the sake of having a unified platform to match their ATI GPU.

[stgatilov]

From the point of high-performance computing, AMD Fusion is conceptually better than CPU + discrete GPU. It is obvious that transfer bandwidth between CPU and discrete GPU is rather slow (compared to internal memory transfers and computing power). GPGPU suffers from low transfer speed in many computational problems. Ideally, memory spaces of CPU and GPU can be merged together to make it possible for the programmer to access particular data from both CPU and GPU without any transfers at all.

 

However:

1. Now CPU and GPU memory spaces in APU are separate. And it is VERY difficult to utilize the fast transfers between them properly. I sincerely hope AMD will solve this problem...
   
2. OpenCL is functionally cross-platform but not performance-wise. A single code can be run on all the architectures with OpenCL: Intel multicore CPU, NVIDIA GPU, AMD GPU. But it is impossible to tune the OpenCL code to work fast on all of them. For instance, fast transfers (zero-transfers / overlapping transfer + kernel execution) are implemented in completely different way on AMD and NVIDIA GPUs. Unless various performance features can be used in a simple and unified way on all architectures, they can't be widely used commercially.
   
3. There is a CPU vs GPU myth that states that modern GPU is about 10-20 times faster than modern CPU in practical high-performance problems. It is far from being correct=) The actual difference is about 2-4 times. Now recall that GPGPU programming is much more difficult than CPU programming.,,
   
4. AVX instruction set is replacing the old SSE set. The width of SIMD instructions is doubled => more throughput on CPU.
   

So as you see, I do not believe in the idea of unloading all the floating point computations to GPU. To my mind, CPUs have to conserve and increase floating point computational throughput in the nearest future.