[wccf] NVIDIA Maxwell Based GeForce Graphic Cards Reportedly Shipping In March 2014

[MrK6]

If it's not 20nm I'm not interested.

 

[f1sherman]

A senior TSMC executive revealed recntly that the company will begin 20nm production in the first quarter of 2014, contributing to the company's revenue in the following quarter.

Industry sources said TSMC's 20nm production capacity has been booked up with orders from industry giants including Apple Inc., Qualcomm Inc., Xilinx, Altera, Supermicro, NVIDIA, MediaTek and Broadcom Corp.

:hmm:

http://focustaiwan.tw/news/atod/201312050035.aspx

 

[NTMBK]

A senior TSMC executive revealed recntly that the company will begin 20nm production in the first quarter of 2014, contributing to the company's revenue in the following quarter.

Industry sources said TSMC's 20nm production capacity has been booked up with orders from industry giants including Apple Inc., Qualcomm Inc., Xilinx, Altera, Supermicro, NVIDIA, MediaTek and Broadcom Corp.

:hmm:

http://focustaiwan.tw/news/atod/201312050035.aspx

20nm SoC... remember, NVidia also makes Tegra.

 

[f1sherman]

ya... I'm guessing Maxwell will debut at 28nm. Scratch that - I am sure of it.
Last thing Nvidia needs is another supply/capacity nightmare.

 

[dangerman1337]

I think Mobile Maxwell will debut on 28nm next quarter and then possibly following Q2/Q3/Q4 we'll see 20nm Maxwell Desktop and then 20nm Maxwell Mobile in Early 2015.

 

[Silverforce11]

20nm production is very limited, eventually the first 20nm Maxwell need to be put into Teslas to fill the HPC demand. Anyone expecting 20nm Maxwell in consumer GPUs in 1st half 2014 is very very optimistic. Its not NV that I doubt, it's TSMC.

 

[blastingcap]

Competition is good for consumers no matter what brand you prefer.

 

[Vesku]

I could see Nvidia releasing a 28nm Maxwell large die product for Tesla and top tier GPUs. Competition in the HPC market is getting quite fierce.

 

[TechFan1]

It would be sweet to see what kind of cards they could make using Intel's fabs.

 

[ams23]

the likelyhood of the first batch of (presumably) mid-range Maxwell's overtaking the 780 Ti and R9 290X is very small. There will be no killer 680-type card this time around.

You forget that Maxwell will have superior energy efficiency compared to Kepler (and Kepler already has better energy efficiency than Hawaii). GTX 880 will be a killer card and will surely be the world's most energy efficient discrete GPU when it comes out. That said, most people with GTX 780 Ti may be on a two year upgrade cycle and may go with an even higher performance, larger die future Maxwell variant.

 

[Stuka87]

You forget that Maxwell will have superior energy efficiency compared to Kepler.

Source?

 

[Grooveriding]

I think overall 20nm cards will be a huge disappointment after what we've seen on 28nm coming off 40nm. TSMC is on record as saying 20nm will not deliver the degree of improvement over 28nm that 28nm did over 40nm.

 

[krumme]

I think overall 20nm cards will be a huge disappointment after what we've seen on 28nm coming off 40nm. TSMC is on record as saying 20nm will not deliver the degree of improvement over 28nm that 28nm did over 40nm.

We have to look at this in perspective. 20nm gives huge size savings. That translate into better products for us especially when the first crazy run from mobiles on 20nm is over.
I think huge gpu will get good benefits compared to other products.
Same size. More transistors. Just lower freq and the tdp target is there.
Actually i think phones and smaller cpu have far more need for 16nm performance because they are prone to leaking and high perf tl get both good battery life and high top perf. To a far higher degree than big desktop gpus.

 

[SiliconWars]

You forget that Maxwell will have superior energy efficiency compared to Kepler (and Kepler already has better energy efficiency than Hawaii). GTX 880 will be a killer card and will surely be the world's most energy efficient discrete GPU when it comes out. That said, most people with GTX 780 Ti may be on a two year upgrade cycle and may go with an even higher performance, larger die future Maxwell variant.

Kepler isn't more efficient than Hawaii and this will be seen once the aftermarket 290X's are out.

 

[Grooveriding]

Kepler isn't more efficient than Hawaii and this will be seen once the aftermarket 290X's are out.

What ? Have you seen 780ti vs 290X perf/w ?

Obviously any statements here about Maxwell here are fantasy, as nothing is known until you see the hardware in action, but 780ti is much better perf/w than 290X.

 

[blackened23]

Kepler isn't more efficient than Hawaii and this will be seen once the aftermarket 290X's are out.

LOL

Hawaii is fast in the right conditions but this? No it isn't. Kepler is a larger die and is still more efficient than Hawaii.

Oh but the..."just you wait!" conditional argument. Alrighty. We'll see I suppose.

 

[Erenhardt]

LOL

LMAO

 

[BallaTheFeared]

LOL

Just you wait!!!




Wouldn't GM104 be nearly as big as GK110?

I don't think we'll see either GM104 or GM100/110 on 28nm.

 

[VulgarDisplay]

What ? Have you seen 780ti vs 290X perf/w ?

Obviously any statements here about Maxwell here are fantasy, as nothing is known until you see the hardware in action, but 780ti is much better perf/w than 290X.

He's referring to aftermarket coolers lowering 290x power consumption by 30 or so watts which brings it right in line with Kepler perf/w.

 

[blackened23]

He's referring to aftermarket coolers lowering 290x power consumption by 30 or so watts which brings it right in line with Kepler perf/w.

So what happens when you put an aftermarket cooler on a Kepler card? Same thing, right? Like I said Hawaii is a fast chip in the right conditions. But the conditional "but aftermarket will make it efficient!" clause is ridiculous because if that were true, it would apply to kepler also.

Aftermarket also doesn't necessarily have better efficiency. Most aftermarket designs are beefed up with slightly more than reference voltage. You'll find plenty of 290X aftermarket cards will actually use more power because of this reason, because there is literally NO REASON for these cards to exist unless they're factory overclocked. And if they're factory overclocked, what happens? Power consumption goes up.

I'm not a huge stickler for power consumption, I only care about secondary effects (eg noise, heat). But this entire argument is ridiculous because aftermarket cards will be factory overclocked, any stock clocked aftermarket card is worthless. Being that they are factory overclocked, power use goes UP. Not down. Furthermore, the "aftermarket" clause applies to Kepler as well. If Hawaii power goes down with aftermarket, so does Kepler. But that won't happen. Like I said - aftermarket cards are designed to be better and faster than reference through overclocking. Obviously these factory OC'ed cards will use more power than the reference with that being the case.

 

[BallaTheFeared]

So what happens when you put an aftermarket cooler on a Kepler card? Same thing, right?

Magic.

 

[ams23]

He's referring to aftermarket coolers lowering 290x power consumption by 30 or so watts which brings it right in line with Kepler perf/w.

No, he is flat out wrong and in a state of denial. Just to give an example, if one looks at Anandtech's review, measuring both performance and power consumption in Crysis3 at the same settings and same gameplay sequence, GTX 780 has nearly 15% better (!) perf. per watt compared to R9 290X. That percentage obviously fluctuates, but overall it is very obvious that big Kepler is more energy efficient than Hawaii (just as little Kepler Is more energy efficient than Tahiti).

The "aftermarket" cooling solution you are referring to is a watercooled system that can be used on both Hawaii AND Kepler GPU's. If watercooled system power consumption goes down with Hawaii, it would go down by a roughly equal amount with Kepler too, and Kepler would still be more energy efficient in comparison.

 

[Atreidin]

LOL

Hawaii is fast in the right conditions but this? No it isn't. Kepler is a larger die and is still more efficient than Hawaii.

Well, yeah, having a larger area helps it be more power efficient, so that isn't surprising. It makes it easier to dissipate the heat, at the expense of increased cost. Total size is a deliberate design decision.

I thought an Nvidia employee basically said as much with reference to Hawaii when asked for comment on this subject.

 

[SiliconWars]

That's with 13C cooler temps. As soon as the reference 290X cooler is changed to something decent you'll see 40W drop off that, easily matching any high-end Kepler card. You have to realise just how much more Watts that awful AMD reference cooler is adding.

 

[KompuKare]

Surely, the reason that Hawaii's power consumption might benefit more from aftermarket cooler is because like the classic IDC posts show, leakage at 95°C is a lot worse than at 83° or so where the reference GTX 780TI runs at. So yes, in theory both would probably benefit from aftermarket coolers but the 290X should get a bit more from it.

In theory that is: but as blackend23 pointed out, non-ref models tend to have fancier power delivery, run at different voltages and get to charge premium for being factory overclocked. Now in theory AIBs could bin parts for their OC's but most don't (still have a sore point with my Gigabyte Windforce 7950 forcing 1.25V mostly because Gigabyte can't be bothered to bin their chips - mine runs LTC fine at 1022/1500 with 0.95V for instance).

But anyway, 28nm TSMC is pretty much equal for GPUs isn't it? And once you take into account a few factors (GCN's higher DP performance vs GK104, die size of GK110 vs Hawaii) they two architectures have fairly similar perf/watt. Even the TPU chart shows a max of 70/100 spread which is hardly massive considering the chart has everything from 650/7770 to 290x/780Ti.

Still, since the temperature/leakage/power consumption thing should be very well known to anyone working in the industry, that makes AMD's decision to run with the old blower reference cooler even stupider (although I suspect a beancounter made the decision not an engineer).