Richland & Kabini rumours

[SiliconWars]

I "got" it. It is not a difficult concept to comprehend. You are still however making an apples to oranges comparison to try to show AMD in the best light.

And I never said the pentium plus 7750 would be cheaper. But lets look at the numbers. Pentium is about 80.00, 7750, OK, I will assume 100.00 which is a very high estimate. A10 maybe 110.00. So 70.00 difference or maybe 15 to 20 percent of the cost of a system, IF you can find a well priced A10. In gaming I would estimate at least 50% improvement. So 20% more cost for 50% improvement. Seems like a good trade-off to me.

Yes but you're also getting lower cpu performance and higher power draw (especially at idle), while basically having no game future-proofing because of the dual core. Suddenly it seems like a really bad tradeoff and you should have spent another $30 and gone with the A10. You can always play the "add a little more money for something better" game.

 

[AtenRa]

Except the power consumption over time graphs are identical. There is no meaningful difference between those program versions.

There is a big performance/power difference between HD 4.0 with Windows 7 vs HD 5.0.1 with Windows 8.

Also, techreport only measured 60secs of the benchmark. Since the benchmark performance and power consumption fluctuates all the time, the measurements they took(60sec) doesn't reflect the power consumption each system consumed to finish the entire benchmark.

You can see from AT graph that 3570K power usage picks at more than 100W when in Techreport graph the 3570K stays close to 90W the entire 60secs.

Even if we assume that you're correct about efficiency (but let's not forget, as I've just demonstrated, that you're not), you've still ignored the other half of his argument involving performance in a given die space.

One SandyBridge core(32nm) is almost the same size as a Bulldozer Module(32nm). I will say that Core die size/performance/power usage in MT loads is almost the same (depending on the application).

Since CPU dies are so much different, SB has iGPU when Bulldozer has 8MB L2 Cashe, we cannot directly compare them.

We can compare the Intel Core i7 3820 with FX8350 with a die size of 315mm2.

Core i7 3820 is a quad core + HT with 1+10MB L2+L3 Cache and 294mm2 die size at 32nm and 130W TDP

FX8320 is a Quad Module with CMT with 8+8MB L2+L3 Cache and 315mm2 die size at 32nm and 125W TDP

I would really like to see a Performance/power comparison between those two.

 

[AtenRa]

I "got" it. It is not a difficult concept to comprehend. You are still however making an apples to oranges comparison to try to show AMD in the best light.

And I never said the pentium plus 7750 would be cheaper. But lets look at the numbers. Pentium is about 80.00, 7750, OK, I will assume 100.00 which is a very high estimate. A10 maybe 110.00. So 70.00 difference or maybe 15 to 20 percent of the cost of a system, IF you can find a well priced A10. In gaming I would estimate at least 50% improvement. So 20% more cost for 50% improvement. Seems like a good trade-off to me.

Then i will say spend an extra 20-30 bucks and get the A10-5800K + the HD7750 and have better gaming performance than Pentium + HD7750.

Have a look how your Pentium gets crushed by the A10-5800K + HD7970 or GTX680.

http://uk.hardware.info/reviews/371...rce-gtx-680-tested-with-10-cpus-dirt-showdown

Get used to it, modern games need minimum a quad core CPU 😉

 

[frozentundra123456]

According to Anandtech's review, it runs it at 47.6fps. Likely different settings though.

http://www.anandtech.com/show/6332/amd-trinity-a10-5800k-a8-5600k-review-part-1/5

With that said, you're thinking in too much of a gamer's mindset. Having the graphical horsepower of a Radeon 7750/7770, while nice, simply isn't necessary to actually enjoy a game. Even new games can be played on lower level hardware. I don't primarily game on my laptop (which sports a Geforce 540m), but when I do, I can get the job done at a reasonable, enjoyable level.

But with that said, if you're buying a 5800k with the mindset to game on it full time, then you're better budgeting your money elsewhere. People who buy the 5800k, whether they realize it or not, are going to be purchasing it with the mindset to get a cheap, capable computer. Buying an additional $100 piece of hardware not only defeats the purpose, and contradicts the purchasing environment.

I mean, let's be honest. The largest market for something like a 5800k is going to be HTPCs, all-in-ones (which are limited by space anyways), and people who go to Best Buy and purchase a cheap desktop computer. Not gamers.

You are right about the discrepency in the benchmarks. I dont have any explaination for it. Toms did a series of articles on gaming with the igp, but I could only find one. That is where I took those numbers. To be honest, Anand's numbers do seem more reasonable, but OTOH,
there are some graphically intense games that are basically unplayable on an igp except at extremely low quality and resolutions. (e.g. Metro 2033 even in Anands test)

I will agree with you on a laptop. I think trinity is an attractive choice for lite gaming on the go, where you are constrained by thermal/power issues and cant really add a discrete card.

On the desktop, it just doesnt make sense. Even if a discrete card costs 100.00 more, and you use the computer for 3 years, that is less than the cost of buying one game per year.

But I give up. I am outnumbered by those insisting an APU is a "gaming" chip. So be it. To me it is a very limited solution that can be easily beaten for a minimal increase in cost. I am done arguing with those who want to accept the compromises of gaming on an APU.

 

[frozentundra123456]

Then i will say spend an extra 20-30 bucks and get the A10-5800K + the HD7750 and have better gaming performance than Pentium + HD7750.

Have a look how your Pentium gets crushed by the A10-5800K + HD7970 or GTX680.

http://uk.hardware.info/reviews/371...rce-gtx-680-tested-with-10-cpus-dirt-showdown

Get used to it, modern games need minimum a quad core CPU 😉

Now you are bringing up another scenario, but yes, I agree, an A10 plus 7750 would be faster than the pentium plus 7750 in most games, and both would be faster than an igp. The point I have been trying to make is that any igp is very limited for gaming. Again this scenario is irrelevant to that contention.

 

[Torn Mind]

I "got" it. It is not a difficult concept to comprehend. You are still however making an apples to oranges comparison to try to show AMD in the best light.

And I never said the pentium plus 7750 would be cheaper. But lets look at the numbers. Pentium is about 80.00, 7750, OK, I will assume 100.00 which is a very high estimate. A10 maybe 110.00. So 70.00 difference or maybe 15 to 20 percent of the cost of a system, IF you can find a well priced A10. In gaming I would estimate at least 50% improvement. So 20% more cost for 50% improvement. Seems like a good trade-off to me.

100 dollars for a 7750 is not a high estimate at all. It is slightly above the cheapest price of $90 AR on pcpartpicker.com. http://pcpartpicker.com/parts/video-card/#c=80&sort=a5

 

[Rvenger]

Use the A8-5600k + HD7750. If you are using the HD7750 there is no point in going to the A10 since the A8 is a quadcore and you can get 4ghz on stock volts.

 

[shady28]

I will have to disagree with that. Have a look at the link bellow. The review is with a single GTX670. A10-5800K(FX4130) is on par or better than Core i3.

http://pctuning.tyden.cz/hardware/g...esoru-na-vykon-ve-hrach-od-phenomu-po-core-i7

On par is maybe accurate; I wouldn't say the A10 is better as a CPU.

But again, I think trying to grade the A10 against other chips with discrete GPUs is pointless and has nothing to do with its purpose.

What would be a really accurate statement is that the A10-5800K is currently the fastest combined CPU/GPU single chip solution in the world insofar as 3d graphics and gaming goes. I am pretty sure that is an accurate statement.

 

[itsmydamnation]

Interesting. You are saying that single threaded performance went down as a trade off for better multi-threaded performance, and yet you don't correlate this with IPC. Care to explain your reasoning here?

no im not say that, what i am saying is:

bulldozer target single thread per module perf per clock = 1
bulldozer target dual thread per mudule perf per clock = 1.8
bulldozer delivered single thread per module perf per clock = 0.8
bulldozer delivered dual thread per module perf per clock = 1.3

im saying bulldozer missed its design targets, you can see this quite easily by PD having 10%-15% higher perf per clock with only limited changes. you dont see 10-15% in multithreaded because the module still gets bottleneck by instruction fetch/decode which wasn't changed for PD but is for SR.

In your opinion, what are the merits of the Core architecture and why does an Intel SMT core performs on par with AMD module at lower clock speeds?

1. bulldozer didn't hit its performance target
2. Intels core design as been about single thread perf with a mechanism (SMT) to increase peak throughput. if we look at the number of execution resources between a SB/IB core and a bulldozer module you get

             SB/IB / DB/PD
int ALU       3           4
SIMD/FPU      3         2+2 (two int SIMD, two FP SIMD, becoming 1,2 in SR)
AGU/store     3           4

So as you can see BD/PD module has more execution resources at a decreased complexity cost compared to: 1. putting that all in one core, 2. make the core SMT capable. Now when you compare it to SB/IB to BD/PD its simple to see that by the number of execution resources BD/PD should be able to (ignoring AVX-256 for the moment) beat SB/IB in threaded workloads quite comfortably, but it doesn't. why doesn't it? because its execution resources are being "starved".

in single thread SB/IB have a few advantages, more L1 cache for the single thread, more integer execution resources, depending on workload IB/SB L2 could help a lot. it should beat BD/PD in single threaded workloads but not by as much as it does.

Bulldozers implementation didn't go well, you have to factor that in when looking at high level architectures like http://www.realworldtech.com/bulldozer/10/

 

[mrmt]

no im not say that, what i am saying is:

bulldozer target single thread per module perf per clock = 1
bulldozer target dual thread per mudule perf per clock = 1.8
bulldozer delivered single thread per module perf per clock = 0.8
bulldozer delivered dual thread per module perf per clock = 1.3

What you are saying does not match AMD statement, as it mentions explicitly "Throughput advantages for multi-threaded workloads without significant loss on serial single-threaded workload components". Assuming that AMD was comparing Bulldozer with Thuban on that slide, it was more or less what happened from Thuban to Bulldozer, except that the single threaded performance loss was abysmal, not small.

And I have to ask you, how can they NOT lose IPC if they were going from 3 to 2 ALU and the decoder from 3-wide to 2-wide AND a longer pipeline AND more complicated cache management? Those are all actions that decrease IPC.

Interpreting AMD statement as you are, there isn't loss on IPC performance, it is always 1, or Thuban IPC. I cannot see how AMD could have taken the measures described in the paragraph above and not lose at least a bit of IPC.

 

[Tuna-Fish]

And I have to ask you, how can they NOT lose IPC if they were going from 3 to 2 ALU and the decoder from 3-wide to 2-wide AND a longer pipeline AND more complicated cache management? Those are all actions that decrease IPC.

By counterbalancing with other actions that increase IPC. For example, the branch predictor in BD is significantly better than what AMD has had before.

Unfortunately, it didn't pan out. The changes that cost IPC cost more of it than the changes that improved it.

 

[Phynaz]

IF AMD(GloFo) had the capacity they would. 😉

Are you really saying AMD's current sales issues are becuase GloFo can't supply Trinity CPU's?

That's why AMD paid them to not produce chips, because GloFo couldn't produce them? 😕

 

[Abwx]

the decoder from 3-wide to 2-wide

Decoder is 4 inst. wide.

 

[Homeles]

Decoder is 4 inst. wide.

Split between two ALUs, yes.

 

[Abwx]

Split between two ALUs, yes.

Only for pure single thread workload , otherwise it is even
more shared by 4 ALUs...

 

[Homeles]

Two cores. Whatever.

 

[Ventanni]

Back to topic, gentleman! We're getting off track! /claps like a football coach

Here we have Vizio's 11.6" Hondo (Bobcat core) based 4.6w TDP dual core tablet running Windows 8. Initial hands-on seems pretty positive according to Anandtech. Should be a good sign for Temash, should it not?

http://www.anandtech.com/show/6672/vizio-tablet-pc

 

[piesquared]

Back to topic, gentleman! We're getting off track! /claps like a football coach

Here we have Vizio's 11.6" Hondo (Bobcat core) based 4.6w TDP dual core tablet running Windows 8. Initial hands-on seems pretty positive according to Anandtech. Should be a good sign for Temash, should it not?

http://www.anandtech.com/show/6672/vizio-tablet-pc

Yeah reports from CES were pretty positive of this little beauty. Awesome screen with 1080p and enough horsepower to run it smoothly. At 1W SDP on 40nm it will no doubt have enough graphics performance to outperform even the 7W SDP ivy bridge on 22nm! Temash has 2x gaphics performance of Hondo.

 

[Homeles]

Stop using SDP. That term is killing me.

 

[Abwx]

At 1W SDP on 40nm

Its SDP is about 2W according to intel s new scheme.

 

[podspi]

Its SDP is about 2W according to intel s new scheme.

Source and/or calculations? Wouldn't that require actual testing to know?

 

[Abwx]

1.72W for average task including the FCH , otherwise using
the intel norm of CPU dissipation only it is 1.12W , so Piesquared
was spot on....

 

[itsmydamnation]

What you are saying does not match AMD statement, as it mentions explicitly "Throughput advantages for multi-threaded workloads without significant loss on serial single-threaded workload components". Assuming that AMD was comparing Bulldozer with Thuban on that slide,

I have the video of the hotchips presentation that the slide comes from, it wasn't in relation to STAR's in is in relation to a bulldozer core with dedicated not shared front end. Not one comparison was made again STAR's based cores in the entire presentation.

And I have to ask you, how can they NOT lose IPC if they were going from 3 to 2 ALU and the decoder from 3-wide to 2-wide AND a longer pipeline AND more complicated cache management? Those are all actions that decrease IPC.

because of flexibility, STAR's ALU's and AGLU operated in pairs which limited performance. Also because memory access latency is the single biggest performance limiter, the better a core is at prefetch and prediction the higher the core utilization.

The cache hierarchy isn't that different they basically have just tried to reduce probs hitting the L1D. Also given that integer IPC in X86 on average is not much over 1, having 3 ALU's is into diminishing returns, if HT adds on average 30% perf and IB/SB have 3 integer ALU's what does that tell you. ( yes i know that is a simplistic example, but the entire core is balanced for it).

Interpreting AMD statement as you are, there isn't loss on IPC performance, it is always 1, or Thuban IPC. I cannot see how AMD could have taken the measures described in the paragraph above and not lose at least a bit of IPC.

my numbers aren't IPC but relative performance and have nothing to do with STAR's they are only in relation to bulldozer. If you wanted to guess where stars was i would have put it about 0.9.

 

[AtenRa]

Are you really saying AMD's current sales issues are becuase GloFo can't supply Trinity CPU's?

That's why AMD paid them to not produce chips, because GloFo couldn't produce them? 😕

I was talking in general,
Dont forget that currently AMD uses both GloFo and TSMC and supplies in the channel is not as smooth as it should be.

 

[beginner99]

A HD7750 is $90-110 on Newegg- making the whole shebang far more expensive than the Trinity setup. And that Pentium is going to choke in a lot of games with only two threads. And it doesn't have AVX support enabled, unlike Trinity.

EDIT:

From here: http://techreport.com/review/23662/amd-a10-5800k-and-a8-5600k-trinity-apus-reviewed/10

Modern engines are being optimised for more than two threads, and that Pentium seriously struggles.

EDIT 2: And yes, that's with a discrete GPU.

nice cherry picking.

Now post the same thing for starcraft 2, one of the most popular games, and see how that looks.