Comments on: AMD: 'Huge, monolithic' chips not our style

Company's ATI graphics chip unit isn't trying to match Nvidia by building large chips, according to an AMD exec. But an Nvidia exec says smaller isn't always better or more efficient.

[i_am_still_wade]

I think AMD's approach is better. NVidia says it isn't better because NVidia doesn't know how to do what AMD knows how to do, that is multi-core chips. So NVidia could say "hey, we are behind" or "hey, our way is better, trust us". Which do you think the stockholders want them to say?

The AMD-Ati merger will pay off in the end, first for the Ati, then for the AMD side. AMD still does multi-core better than Intel, although that doesn't make the Phenon better. It won't be long before AMD puts out a quad-core GPU.

[sonymaster101]

correction, it would be a quad gpu card, not quad core. gpus technically have hundreds of cores, called stream processors.

and about the cpu thing, it doesnt matter how good one processor is compared to another, it matters how much money you can make. but I think amd is in a dangerous spot these days. intel is pulling ahead fast in prosessor design and efficiency, for instance

-atom
-nehalem

amd was touting their phenom for a long time, but it turned out to be a bust in performance compared to intels much more developed quad core, even though is wasnt a "purebred" quad core.

the nehalem is going to really hurt amd unless they can either put out a better product or market the phenom differently.

[smokified]

the only part about your statement that cannot be shut down is the part where you say YOU THINK AMD's approach is much better. The rest of what you say is based on blind ignorance.

nVidia makes WAY better video cards. ATI/AMD has been trying for the last couple of years to compete with the next-gen cards being released by nvidia and as soon as this highly acclaimed "nvidia buster" comes out, it is quickly realized that their approach is not working.

When they say more energy efficient I think of the hybrid car. 2 engines. Slow as crap.

People are not switching over to nVidia in droves because their cards suck and are more expensive. They are switching to nvidia because their cards work extremely well.

It sounds to me like you bought the ATI card to save a few bucks and you are now looking to justify that decision by making false statements.

[Mam00th]

@i_am_still_wade.

Well sorry to tell you that but the 8800GT is the best quality/price ratio you can get at the moment... I don't think Nvidia doesn't know how to make good graphic card or good multi-core chips, each chip already have more then a hundred cores (I know it's a bit different but still...)... Don't get me wrong, I admire ATI/AMD as well, they were clearly the winner in the last generation, just don't say Nvidia don't know how to do their job...

[sonymaster101]

amd's gpus are still worse than nvidia's, no matter how many ways you put it. nvidia didnt even have any competition for about a years time with the 8 series, and the 9 series topped that.

"huge", "monolithic" chips will always be faster, because the chips dont have to communicate with each other over a relatively slow bus(compared to the chip). and, um, i_am_still_wade, if amd is better at their job, how come they have to use 3 times as many stream processors and still have a slower card?

[ralfthedog]

My guess is that Nvidia will take the top performance prize while AMD/ATI will win on a price/performance scale.

In the end, not even the top people at AMD or Nvidea know how this will turn out.

[ralfthedog]

My guess is that Nvidia will take the top performance prize while AMD/ATI will win on a price/performance scale.



In the end, not even the top people at AMD or Nvidea know how this will turn out.

[darkr]

i_am_wade i'm not sure where you are gettting that AMD is better at multi core then Intel
but you better check your facts first

Intel Core 2 Duo series beat AMD's x2 series on performance/power consumption/ value except for the high end opterons but thats a different chip series you gotta compare opterons to zeons

perhaps you are thinking about the pentium D series which i would aggree x2 beats the pants off it.

though AMD did put 4 cores on 1 die faster phenom then intel a 4 cores on 2 die

http://techreport.com/articles.x/12091
http://www.extremetech.com/article2/0,2845,2203595,00.asp

[AppleSuxLeo]

But...Will it blend ? Polygons I guess ?

[cubicleslave1]

I've been a chip design engineer for 22 years, and one has always held true: the more input/output pins you have on a chip, the more power you burn. That's because the capacitive load of a board trace can dwarf that of an on-chip wire by up to 100x. In addition, the IO driver circuitry itself burns a lot of power. So breaking a chip into two chips and communicating over the board is very costly in power consumption. It is also a lot harder to drive signals as fast over a board, due to signal integrity issues. However, the advantage of splitting a chip into two pieces is that each piece is smaller, and that makes it easier to get good yields in the fab. The bigger the chip, the harder it is to get good yields because defect density can only be held so low. It's like trying to hit the broad side of a barn compared to hitting a squirrel with a machine gun (think of a bullet as a defect site). The barn has a much higher probability of being hit. I think Desai hit the nail on the head, AMD just doesn't have the money to invest in a big chip and get the defect density low enough to get good yields. Desai was also being diplomatic, it is not only "simplistic" to say that two chips is better for performance, it is plain wrong. So the bottom line is, AMD is unable or cannot afford to develop their process to the point of yielding well for a large die for the GPU market. Probably their "good" fab capacity is all taken by Barcelona and they have to make-do with their lower-tech fabs for ATI. "Not our style" is just a lame attempt at media spin.

[aferozpu]

"So breaking a chip into two chips and communicating over the board is very costly in power consumption"

What is the comparison of the energy lost due to capacitive load on the board traces compared to the energy lost by the transistors switching on a chip?

If the energy lost due to capacitive load on the board traces is much smaller, then your argument does not hold.

Just because AMD has another processor does not mean double the IO pins on the chip; this is an Intel Design. I agree there will be more power dissipated due to increased resistivity in the circuit, but how much power is dissipated has yet to be demonstrated.

?Desai was also being diplomatic, it is not only "simplistic" to say that two chips is better for performance, it is plain wrong.?

What is your measure of performance? Is it only power dissipation? Is it performance per Hz? Or is it a combination of these two factors and possibly more? I think Desai made a legitimate statement; ?it's too simplistic of an argument to say that two smaller chips is always more efficient." He never mentioned performance.

Your arguments seem biased.. what company did you say you work for again?

[aferozpu]

In my last post I meant this is NOT an Intel Design.

[aferozpu]

test

[aferozpu]

For more visibility I reposted; Sorry I am new to this forum.

To cubicleslave1:

"So breaking a chip into two chips and communicating over the board is very costly in power consumption"

What is the comparison of the energy lost due to capacitive load on the board traces compared to the energy lost by the transistors switching on a chip?

If the energy lost due to capacitive load on the board traces is much smaller, then your argument does not hold.

Just because AMD has another processor does not mean double the IO pins on the chip; this is an Intel Design. I agree there will be more power dissipated due to increased resistivity in the circuit, but how much power is dissipated has yet to be demonstrated.

?Desai was also being diplomatic, it is not only "simplistic" to say that two chips is better for performance, it is plain wrong.?

What is your measure of performance? Is it only power dissipation? Is it performance per Hz? Or is it a combination of these two factors and possibly more? I think Desai made a legitimate statement; ?it's too simplistic of an argument to say that two smaller chips is always more efficient." He never mentioned performance.

Your arguments seem biased.. what company did you say you work for again?

[cubicleslave1]

aferozpu, I'm trying to understand where you're coming from. Why are you considering the energy dissipation of a transistor switching? We are comparing two cases. Case1: a huge monolithic chip with N transistors. Case2: same chip broken up into 2 chips, each with N/2 transistors and the two pieces talking to each other through board traces. The number of transistors is approximately the same in both cases, although for case2, you have additional transistors due to the I/O circuitry. Case2 is AMD's case. Your question about the energy to switch a board trace compared with the energy of a transistor switching is irrelevant for the reason just stated. However, I will still address it. The capacitive load of a shorter board trace is in the multi-picofarad range. The capacitive load of a transistor input is only a few femto-farads. The typical on-chip interconnect is also in the same range, although a very long one can get close to 1pf (but those are very rare exceptions). Power consumption in watts is directly proportional to capacitive loading, so it's reasonable to say that the signals that got replaced with board traces end up burning 100x more power than before. We must also consider signal integrity issues. At the board scale, lines become extremely long, several milimeters long, compared with the chip-scale. That introduces transmission line effects, and you need special analog circuitry to deal with that, to address reflections, termination, and differential sensing, etc. Analog circuits are also power hogs and tend to dissipate static power (power that is burned even with no switching activity). Due to the transmission line effects, you can also only drive a signal over a board at a certain speed, nowhere near what you can drive at the chip level.

So in summary, we are considering the merits of breaking a chip into two pieces, and replacing the on-chip wires that used to go between the two pieces with board traces. Again, there is no benefit to doing that in terms of power consumption and performance. The only benefit is that the dice become smaller and make it easier to get good yields, bringing down the manufacturing cost. Counter-acting this benefit, however, is the additional packaging cost and taking up more board real estate.

My measure of performance is either raw performance or performance per watt, and the two-chip solution is inferior on both counts. Less raw performance due to the limitations of driving a signal across the board, and more total power dissipation due to the additional board traces and IO circuitry.

Again, it is clear that AMD does not have the fab capacity to make their GPU's on their best fabs, so they're spinning the press with "not our style" hooey. I believe this is what Desai was probably thinking, but he made his statements a bit more charitable. I would guess that all of AMD's available cutting edge capacity is being taken up by Barcelonas and Opterons. Other than this paragraph (due to the "hooey" remark), I'd like you to point out where my comments are biased. Everything I've stated is pretty factual, I think.

[aferozpu]

In my last post I meant this is NOT an Intel Design.