AMD Greenland Vega10 Silicon To Have 4096 Stream Processors?

#5250927 Posted on: 03/28/2016 06:08 PM

---

I understand what you mean, but clocks are still important, not alone, but in the context of the design.

Im not saying nvidia will clock higher than amd or vice versa, I'm saying I care about the throughout, whether it's achieved through a big slow design, or a fast small design doesn't affect me in the slightest.

The only way I can justify the 8x thermal density is if 14nm transistors consume double the 28nm ones, which would be pretty funny. 4x transistor density, 2x power = 8x thermal density. How I actually think it is : 4x density 0.5x power

Just saw your comment Fox, okay yeah if you're talking 150% higher clocks then it makes sense, it's just that... That's not gonna happen
That clock can't happen because we do not have cooling for it. And it was to support idea, that each 14 and 16nm will hit certain clock at given point in time and technological advancement, and OC will be very difficult to achieve.
And that meant both AMD and nVidia are going to end up with much more similar GPU clocks than they did on 28nm. 25% clock difference is very improbable to happen again. (28nm limitation for both AMD and nVidia came from transistor density which limited stability, 14nm clock limitation will come from cooling)

Now what I meant by nVidia having to do more work on die shrink.
980Ti: 8B transistors, 1350MHz clock
Fiji: 8.9B transistors: 1050MHz clock
We can say that 980Ti performs 5% better in average.
Therefore: 980Ti perf = 1.05 * Fiji perf *
8 * 1350 = 1.05 * 8.9 * 1050
1.1 =
In other words, if GM200 and Fiji had same clock and were limited to same transistor count, Fiji would deliver 10% higher performance. And that's while Fiji is well rounded GPU with no significant weak points like low compute performance, or pixel shader length limitation.

Taking Fiji has higher performance per clock per transistor and having more transistors than GM200. Doing just die shrink and hitting similar clock for both chips would make 14nm Fiji much better. And as I stated before, that's why both companies continue to improve. AMD knows they have certain inherited advantage which they can play. You can see their confidence at presentations. But if they sit on their hands, nVidia will take that advantage from them as it is not that big.

#5250935 Posted on: 03/28/2016 06:27 PM

---

That clock can't happen because we do not have cooling for it. And it was to support idea, that each 14 and 16nm will hit certain clock at given point in time and technological advancement, and OC will be very difficult to achieve.
And that meant both AMD and nVidia are going to end up with much more similar GPU clocks than they did on 28nm. 25% clock difference is very improbable to happen again. (28nm limitation for both AMD and nVidia came from transistor density which limited stability, 14nm clock limitation will come from cooling)

Now what I meant by nVidia having to do more work on die shrink.
980Ti: 8B transistors, 1350MHz clock
Fiji: 8.9B transistors: 1050MHz clock
We can say that 980Ti performs 5% better in average.
Therefore: 980Ti perf = 1.05 * Fiji perf *
8 * 1350 = 1.05 * 8.9 * 1050
1.1 =
In other words, if GM200 and Fiji had same clock and were limited to same transistor count, Fiji would deliver 10% higher performance. And that's while Fiji is well rounded GPU with no significant weak points like low compute performance, or pixel shader length limitation.

Taking Fiji has higher performance per clock per transistor and having more transistors than GM200. Doing just die shrink and hitting similar clock for both chips would make 14nm Fiji much better. And as I stated before, that's why both companies continue to improve. AMD knows they have certain inherited advantage which they can play. You can see their confidence at presentations. But if they sit on their hands, nVidia will take that advantage from them as it is not that big.

Pretty sure 980Ti is 1200mhz stock, and again, it makes no sense to me that you compare at clock parity when a Fiji cores caps out at 1150 and maxwell can do 1500. Everything about next gen clocks is conjecture, we just assume they will be higher than 28nm, and Im agreeing that thermals will be an issue, I just have no idea how to follow the logic that leads you to making statements about Fiji vs maxwell. Yeah Fiji is faster per clock, it also has 1300 more SPs... And runs 400mhz slower

#5250945 Posted on: 03/28/2016 06:50 PM

---

SP is just way AMD/nV invested transistors (same could be said: that GM200 has 32 more ROPs = 50% more than Fiji). Comparison is performance per transistor per clock, because those are limiting factors of manufacturing technologies.
And I can be wrong, maybe GPU frequency will not be limited by heat. Because Heat can be adjusted partly by transistor density. But that equals cost.

Edit: as for clocks a lot of 980Ti owners claimed here that their card boosts to 1350~1450 without them doing anything. But that's probably per manufacturer. Some shops list official nVidia material value, other list those higher boost clocks. Maybe you can disable OC and see how your card boosts.

#5250947 Posted on: 03/28/2016 06:57 PM

---

Taking Fiji has higher performance per clock per transistor and having more transistors than GM200. Doing just die shrink and hitting similar clock for both chips would make 14nm Fiji much better.

Hitting similar clocks! Just like that...

Does that mean that Nvidia gets ****load of shaders, bazillion TFLOPS, async compute, True Audio and whatnot... because clean slate!?
Or is clean slate reserved only for clocks, and everything else remains relatively the same?

If hitting same clocks with GCN was that easy, AMD would do it on 28HPM ya know...




Now what I meant by nVidia having to do more work on die shrink.


Nvidia having more to more work on die shrink? This is obviously wrong.
Considering that not even using high-end HBM was enough to catch up with 980Ti, obviously AMD had much bigger homework to do.
Kind of what NV has already done Kepler -> Maxwell.


980Ti: 8B transistors, 1350MHz clock
Fiji: 8.9B transistors: 1050MHz clock
We can say that 980Ti performs 5% better in average.
Therefore: 980Ti perf = 1.05 * Fiji perf *
8 * 1350 = 1.05 * 8.9 * 1050
1.1 =
In other words, if GM200 and Fiji had same clock and were limited to same transistor count, Fiji would deliver 10% higher performance.


Exactly. There is only that little
Taking design that clocks poorly, relatively speaking, and saying that
all they need to do clock it same as that high-clock design, is beyond ridiculous.

#5250948 Posted on: 03/28/2016 06:59 PM

---

SP is just way AMD/nV invested transistors (same could be said: that GM200 has 32 more ROPs = 50% more than Fiji). Comparison is performance per transistor per clock, because those are limiting factors of manufacturing technologies.
And I can be wrong, maybe GPU frequency will not be limited by heat. Because Heat can be adjusted partly by transistor density. But that equals cost.


I see what you mean, but it still makes no sense not to account for clocks when talking about relative performance. By that token a 390X is miles faster than a 980, yet it's consistently outperformed by it, especially at a moderate 1500mhz

Clocks matter.


Hitting similar clocks! Just like that...


Taking design that clocks poorly, relatively speaking, and saying that
all they need to do clock it same as that high-clock design, is beyond ridiculous.

Something else I want to point out regarding the popular view that AMD somehow better support for future apis, even in ashes of the singularity, AMD's best case scenario, a 980Ti outperforms a fury X when overclocked, despite 'async' being detrimental to performance in this game. Clocks matter.

AMD has gotten a lot better in many respects recently, doesn't say anything at all about nvidia