FEATURED Bulldozer rumors

ChanceCoats123 said:

I don't think I understand. All you have to do in order to run multiple instances of super pi at the same time... is to open the program as many times as you want. Unless you set the priority yourself, they would all be assigned to core 0 and your times would be horrible.

Now if you're talking about benchmarks that use all available core/threads, you'll want to see Wprime.

If you're talking about assigning two cores to do the work of one in Super pi, then we discussed this a few pages back, and how it's physically limited by the hardware. It's physically impossible (as of now) to split the instructions such that two cores could do the work of one at the same time. While that would be the absolute best, right now we are stuck with using a law which Dolk brought up: http://en.wikipedia.org/wiki/Amdahl's_law.

Right now, multithreaded work doesn't end until the slowest thread finishes. So if you have four threads, 1, 2, 3, and 4. 1 finishes in 25s, 2 finishes in 35s, 3 finishes in 20s and 4 finishes in 30s, the time to complete this whole task will be 35s since the other threads physically can't go back and help the other. When we can start doing that, the times will drop insanely.

Click to expand...

Yeah... thats a law all scientist hates. Its a legit one at that. The way cores are designed they simply read out the voltages that has been taken in by motherboard and bios I/O and simply calculate everything. If the voltages were to somehow cross you risk shorting out the CPU or simply ending process then and there.

If the software was smart enough to learn when the other cores have finished work it would subdivide the remaining work on the final core to the others for parallel processing. The flaw is, it is possible to actually extend the time because of that hardware limitation of the slowest core always ending the process, the division will continue to happen till you either hit the return of zero, or your computer crashes from the infinite loop. So it could drag on to ages which hopefully the software is designed to terminate such a process before it kills the system. You still have that hardware limit...

It would be an incredible advance in computer science for someone to destroy that limit.

So "we" want a code that can divide/share-in the "ongoing" work on a core But also this should be for idling cores right. Scary.

Theocnoob said:

If this is legit that's pretty great. $320 CPU performing about 85% as well as $600> cpu in gaming. Not to mention the mobo comes in 30% cheaper for the same feature set.

So is it correct what people have been saying? at identical clocks, 1 BD core would be slightly inferior to 1 Bobcat core?

Click to expand...

Well you have to understand that $1000 cpu is... Not Sandy Bridge LGA 2011

Bobcat is K14
Bulldozer is K15

So, Ask AMD lol

Bulldozer is meant to compete against Nehalem though...So, if this came out in 2009 or when the 980X EE came out well..... look who will be boss now lol

@Tangil and Chance, Its not the software. Well it is, but not fully.


Quick look into how CPUs operate on the Fetching Process. Think Highway for multi-thread applications, and streets for single-thread applications.

Now for your Single-thread application, the Fetch cycle of the CPU will grab one instruction at a time, and send its way down to a core to be processed. Now it is not a difficult step to imagine that a BD module would be able to start helping out with a single-threaded application. All the equpiment is right there, except for the Fetch cycle. The Fetch cycle does not only assign a core an instruction but also tells the decode cycle what registers may be needed. The Decode stage will actualy reserve all the registers and information needed to process the instruction. In order to share this single-threaded instruction, the other cores would have to recieve the fetch and decode information at the same time, so they can reserve the same registers to access later.

But why not just call this multi-threading? Because thats all it is. Multi-threaded instructions allow the multiple cores to share the same instruction. Each core gets the same information and reserve the registers as needed. Granted there are adjustments to the multi-threaded apps these days, that allow the CPU not to use as many registers for each shared instrcution.

Now moving on to Admalh's law, and Gustafson's Law, which both clear state that there is a limit the speed up, and it can either be determeind by the number of cores or the frequency of the CPU. Each time a new CPU comes out, the Architects are figuring out a new way to process those same integer, FP, memory, etc. instructions. Now it could be that they increased the internal speed of that specific area, which in some cases do not need it. For example, almost all ALU instructions can be completed in 1 cycle instead of taking multiple cycles in the past. The plentiful amount of registers and the core speed of the CPU help accomplish this. What actually has been increasing the CPU speeds these days is not figuring out these small speed ups of certain areas, but rather the front end. For you see what wastes the most time, in cycles, for the CPU is the front end prediction.

A CPU is able to predict what recourses it will need for each instruction. If the prediction is right (and a good estimate for this success rate is around 77%), an instruction will take a estimated 30% increase in performance. If the prediction is a miss, than this can cause a core to clear out a lot of its contents and reperform the instruction; this wastes many cycles. For every new CPU that has come out since the P4, on both AMD and Intel's side, we have seen a greater rise in pushing for a better prediction system.

Now this is where we can actually compare Intel and AMD. But this is for a different post.


Want to learn more? Keep an eye out for my upcoming article In Depth Look at CPU Architecutre. It will cover the basics of how CPUs work.

@Everyone looking at the benchmark SS

That benchmark SS does not really show the performance of BD, its showing how well the 990FX works with SLI and XFire compared to the Intel's chipset.

@Sernox
The new sandy bridge 2011 is intel's value core. It does not perform better then the 1000 dollar one otherwise the I7 extreme edition's price would have dropped. Which it hasn't for years. Because a newer core is released from a Micro-processing company does not mean that it will be faster then the previous, unless the company has intended so. Also... that does not compare anything about cores... as stated by Dolk

tangletail said:

@Sernox
The new sandy bridge 2011 is intel's value core. It does not perform better then the 1000 dollar one otherwise the I7 extreme edition's price would have dropped. Which it hasn't for years. Because a newer core is released from a Micro-processing company does not mean that it will be faster then the previous, unless the company has intended so. Also... that does not compare anything about cores... as stated by Dolk

Click to expand...

Sandy Bridge E you'll see your "Intel "Value" core." crushing Westmere by a huge whooping 50% increase in performance

This not accounting for the change in interface, either,
with that it's a 75% increase in performance

Dolk said:

@Everyone looking at the benchmark SS

That benchmark SS does not really show the performance of BD, its showing how well the 990FX works with SLI and XFire compared to the Intel's chipset.

Click to expand...

It's comparing Bulldozer SLI to Westmere SLI

It's not a Phenom II test

Bulldozer is not even out yet... so till then I would not trust the benchmarks. The difference in frame rates if that is the case is hardly noticeable so I would not even bother buying a processor with a price jacked far beyond 1000. So where exactly have you gotten this chart from?

Intel's cpu's get insanely hot, even at idle. That 50% increase will come at a heafty price. (Yes I have an I7 desktop I am planning on selling, and even with a 3rd party cooler. It heats up like crazy. But I admit it is fast, just unnecessary because a few fps gains of a difference does not change anything)

So.... hm... a $1000 core rated at 3.9GHz max with 6 real cores 12 logical cores (NOT GOOD FOR CG) and a very noisy stock cooler OOOOORRRR

a 340 dollar 8 real cores, 4.0-4.4Ghz with a noisy stock cooler. What is the better buy...

But alas... the final question is... does it have a badge.

tangletail said:

Bulldozer is not even out yet... so till then I would not trust the benchmarks. The difference in frame rates if that is the case is hardly noticeable so I would not even bother buying a processor with a price jacked far beyond 1000. So where exactly have you gotten this chart from?

Intel's cpu's get insanely hot, even at idle. That 50% increase will come at a heafty price. (Yes I have an I7 desktop I am planning on selling, and even with a 3rd party cooler. It heats up like crazy. But I admit it is fast, just unnecessary because a few fps gains of a difference does not change anything)

So.... hm... a $1000 core rated at 3.9GHz max with 6 real cores 12 logical cores (NOT GOOD FOR CG) and a very noisy stock cooler OOOOORRRR

a 340 dollar 8 real cores, 4.0-4.4Ghz with a noisy stock cooler. What is the better buy...

But alas... the final question is... does it have a badge.

Click to expand...

DIE SHRINK. nothing more to be said.

Don't care. Besides that destroys the CPU's architecture. Some programs, and games may have issues with this.
Aren't you an Intel person <____<

bmwbaxter said:

DIE SHRINK. nothing more to be said.

Click to expand...

LOL, you know what else had a die shrink? The GTX 580. That shrink did a lot for heat output and power draw.

Been out of town for a while. Any update on a release date? Well I guess I will wait till my BD for BD.

tangletail said:

Don't care. Besides that destroys the CPU's architecture. Some programs, and games may have issues with this.
Aren't you an Intel person <____<

Click to expand...

I am only an intel person ATM. I want BD to be a bat outta hell and total wreck anything intel has on the marke or can put on the market. Intels performance in benchmarking is the only reason I like them more. I wasn't saying die shrink to start an intel amd. I was just stating that newer architectures usually get die shrinks which lower power consumption and tend to increase performance, not destroy cpu architecture. Also I don't know of one game or program that can run on a 45nm fabrication process chip and not run on a 32nm chip.

ChanceCoats123 said:

LOL, you know what else had a die shrink? The GTX 580. That shrink did a lot for heat output and power draw.

Click to expand...

sorry for the double post

but the 580 also has 512 cuda cores unlike the 480 with 480 cuda cores. Also there was a decent performance increase between the 2 cards and the 580 runs cooler than the 480. not much lower but a bit.

tangletail said:

Don't care. Besides that destroys the CPU's architecture. Some programs, and games may have issues with this.
Aren't you an Intel person <____<

Click to expand...

Games should love it

http://www.anandtech.com/bench/Product/294?vs=292
1600 VLIW5 SPs vs 1536 VLIW4 SPs

The shrink of the cores in a whole should allow for a higher efficiency

Phenom II Core 0(2/3s used) = Bulldozer Core 0(Fully used)

3 ALUs vs 2 ALUs

But that is just the core

Bulldozer Core 0&1 are 150% die space
Phenom II Core 0 is 100% die space
Phenom II Core 1 is 200% die space

so there is a 25% core shrink in components



But this isn't taking the account of the effects of 32nm and HKMG has on Bulldozer and what the whole architecture does in performance

Edit:
1GB 5870 to 1GB 6950
1600 VLIW5 SPs to 1408 VLIW4 SPs
http://www.anandtech.com/bench/Product/294?vs=331

ChanceCoats123 said:

LOL, you know what else had a die shrink? The GTX 580. That shrink did a lot for heat output and power draw.

Click to expand...

http://www.anandtech.com/show/4008/nvidias-geforce-gtx-580/2

It wasn't just a die shrink

I think I saw that bench sheet also for 1100t. Lol someone's playing with us.

Archer0915 said:

Been out of town for a while. Any update on a release date? Well I guess I will wait till my BD for BD.

Click to expand...

Still looking likely to be Q3 but when still remains a mystery

Seronx said:

http://www.anandtech.com/show/4008/nvidias-geforce-gtx-580/2

It wasn't just a die shrink

Click to expand...

Okay then, I give you the 480.... Either way you look at it, not all die shrinks mean less power and less heat.

ChanceCoats123 said:

Okay then, I give you the 480.... Either way you look at it, not all die shrinks mean less power and less heat.

Click to expand...

Well...

Going from 65NM/1400M(280)TDP 236
to 40NM/3000M(480)TDP 250

It's not as hot as you think

65nm to 40nm isn't a 2x shrink

But, ya with a die shrink doesn't always mean less power and less heat

Die shrink you can expect:
More power, Less power, Less heat, More heat
More efficiency
More cheaper
Higher overclockability

The italic list is not actually die shrink based but more quality based on the die shrink

It's an excuse to increase price. Either way you put it, the power needed will only go down so far, in fact it is possible Die Shrinks will do nothing but increase cost. Also I would not say higher overclock ability. Smaller parts will sometimes not last as long.