Kitguru : Nvidia to release three GeForce GTX 800 graphics cards this October

[jpiniero]

Date 9-5-2014

Geforce GTX 980 Alleged Benchmark and TDP Surfaces – Just 170W for ~110% 780 Ti Performance

http://wccftech.com/geforce-gtx-980-alleged-benchmark-tdp-170w/#ixzz3CXlnf766

That sounds about right.

Part of the reason I think the full die is going to be 1920 and not 2560 is that it'd be too big for mobile. The 980M is probally 1600 cores @ 800-900 Mhz and 100W or so.

They could still of course later release a 990 if they do a Big Maxwell part on 28 nm or even fake 16 nm.

 

[Grooveriding]

Didn't mean to sound hostile. You just can't resist upgrading, even if not immediately, when there is something better out. Even if 980 is a wash performance wise, you could run 3 980's for better performance and equal power consumption... plus you'd have new toys to tweak and play with.

Again, this really doesn't make sense either. Someone could run 3 780tis right now then. Power consumption is an irrelevant metric when it comes to high end cards, you don't upgrade in the high end to use less power for the same performance... Personally I'll be busy playing with my new X99 setup anyways.

The more that comes out about this upcoming gm204, the more it looks like it's a not a card worth bothering with if you are already running a top card from the current 28nm lineups. Same performance with less power consumption... probably slower @ 4K and possibly 2560x resolutions. gtx 680 managed to outdo the 580 even with anemic specs because of the advantage of a new node. This card is stuck on the same node.

All I see to look forward to with this launch is seeing what nvidia's marketing spin is going to be. How it attempts justify labeling this card as a new flagship if all it brings to the table is a performance increase @ 1080p you can't notice outside of running canned benchmarks.

 

[Fastx]

They could still of course later release a 990 if they do a Big Maxwell part on 28 nm or even fake 16 nm.

lol ^ that was funny and assume you mean due to 20nm/16mn combined.

 

[MathMan]

Power consumption is an irrelevant metric when it comes to high end cards, you don't upgrade in the high end to use less power for the same performance...

Few people care about power as a stand-alone metric.

But it's important because it requires a less beefy cooling solution, so lower cost. And if gm204 has a drastically lower power requirement, it can be used in PCs with less beefy power supply as well. (Something Nvidia didn't forget to point out for GTX750Ti.)

And it's important for the highest-end because it effectively dictates the maximum possible performance.

If gm204 is as efficient in perf/mm2 and perf/W as gm107, we have a good indication that Maxwell scales with size. That means we'll be looking forward to a gm200 with a huge leap in performance.

 

[jpiniero]

lol ^ that was funny and assume you mean due to 20nm/16mn combined.

Well, Samsung and TSMC inflated the naming convention. What was 20FF is now "16 nm"; what was TSMC 16nm? is now "10 nm". Samsung's "14 nm" is also comparable to TSMC's "16 nm".

 

[BrightCandle]

We care about power consumption realistically only in how it relates to how many transistors they can put on a GPU and how many of them can work at a time towards the goal of rendering a game. We have a realistic upper limit of 300W and a ergonomic limit for sound around 200 Watts. Above 300W and you are outside the PCI-E slot specification, above 200Watts and the cooler is both expensive and loud.

A few years ago the issue was power consumption. The consumption was not dropping at the same rate as transistors were being added to the designs and thus performance was being thermally limited. We are still largely thermally limited, instead the chips now throttle down, run at lower peak clock speeds than they could and in addition have areas of silicon that isn't used at the same time (dark silicon), all in the pursuit of saving power. Its only going to get worse by the looks of things, the thermal limit we hit years ago hasn't really been beaten yet, we just worked out how to delay the problem for a few more years and mostly at idle.

 

[tviceman]

Again, this really doesn't make sense either. Someone could run 3 780tis right now then. Power consumption is an irrelevant metric when it comes to high end cards, you don't upgrade in the high end to use less power for the same performance... Personally I'll be busy playing with my new X99 setup anyways.

It's not an irrelevant metric. It affects overclocking, the components around it, the type of PSU needed, fan noise, and performance scaling. Like I said, if it overclocks noticeably better and it's 10% faster at stock settings vs. a 780 TI to begin with, then many people will find it more than worth it. You've gone from Titan's, to 780's, to 780 TI's. None of those were particularly noticeable upgrades, but now you're suddenly saying GM204 will suck and isn't worth upgrading although a power user (like you) will likely be able to run 3 GTX 980's at the same power consumption as two 780 TI's and have ~40% noticeably better performance.

Other people out there running 290's or 290x's on stock coolers who are vendor agnostic will trip over themselves to get their hands on these. They'll actually be able to hear anything besides the tornado sitting next to their desk and their home AC unit won't have to run 24/7.

The more that comes out about this upcoming gm204, the more it looks like it's a not a card worth bothering with if you are already running a top card from the current 28nm lineups. Same performance with less power consumption... probably slower @ 4K and possibly 2560x resolutions. gtx 680 managed to outdo the 580 even with anemic specs because of the advantage of a new node. This card is stuck on the same node.

Not arguing that one as I won't upgrade regardless. But people on the ultra high end, like you, often don't upgrade because it's worth it. They upgrade because they can. And how do you know it will actually be slower at higher resolutions? Do you have secret moles telling you this? The node used to make a chip has no direct correlation with resolution scaling whatsoever.

All I see to look forward to with this launch is seeing what nvidia's marketing spin is going to be. How it attempts justify labeling this card as a new flagship if all it brings to the table is a performance increase @ 1080p you can't notice outside of running canned benchmarks.

The GTX580 wasn't noticeably faster than the 480 outside of canned benchmarks, yet was labeled the new high end. The same with GTX 285 over the GTX 280. The same with 780 TI over Titan. The same with HD6970 over HD5870. The same with HD4890 over HD4870. The same with 4790k over 4770k. It happens ALL. THE. TIME. Aren't you used to it by now to the point where it's the norm and justification or spinning by any of the guilty parties (Intel, AMD, Nvidia, Qualcomm, etc.) is just par for the course?

 

[Grooveriding]

Few people care about power as a stand-alone metric.

But it's important because it requires a less beefy cooling solution, so lower cost. And if gm204 has a drastically lower power requirement, it can be used in PCs with less beefy power supply as well. (Something Nvidia didn't forget to point out for GTX750Ti.)

And it's important for the highest-end because it effectively dictates the maximum possible performance.

If gm204 is as efficient in perf/mm2 and perf/W as gm107, we have a good indication that Maxwell scales with size. That means we'll be looking forward to a gm200 with a huge leap in performance.

I agree with that, obviously perf/watt allows a card to deliver more performance within the limits of a given power envelope. Given the perf/w of maxwell, whenever gm200 becomes available - hopefully sometime in 2015 - it's going to be incredibly fast. I'd expect it will be twice as fast as a 780ti.

Unfortunately if these performance rumours are true the 980; being stuck on 28nm, 256 bit, the same 7ghz gddr5 we have now, smaller die than gk110 and so forth is going to be probably the smallest flagship improvement ever. Just seems an epic fail to me, as the performance improvement is what is great about new parts.

I think it works out well for nvidia because they can get more dies out of a wafer using gm204 cores rather than gk110 cores, then transitioning from selling gk110 based cards to gm204 based cards that are cheaper to produce. It helps end users save a little money as well due to lower production costs translating to maybe cheaper prices. Mostly blah metrics though. If the card struggles to show any tangible gains over a 780ti/Titan I don't think it will go over very well though. 1080p has pretty much been crushed with the current cards on the market, it's high resolutions where we need real gains now.

I'm looking forward to the reviews and will only be interested in how the card does at high resolutions to see if it can do anything at all to improve running a 4K setup.

 

[MathMan]

Given the perf/w of maxwell, whenever gm200 becomes available - hopefully sometime in 2015 - it's going to be incredibly fast. I'd expect it will be twice as fast as a 780ti.

I don't think that's realistic if gm200 is also 28nm. If a gm204 is roughly the same performance as a 780Ti (give or take 10%), and a gm204 is 420mm2, then a gm200 would have to be on the order of 800mm2 to have double the performance in the case of perfect scaling. Not going to happen, especially if gm200 has extra DP hw like gk110. 50% faster than a 780Ti is a given. 75% is a stretch.

If gm200 is 16nm, all bets are off.

 

[tviceman]

I agree with that, obviously perf/watt allows a card to deliver more performance within the limits of a given power envelope. Given the perf/w of maxwell, whenever gm200 becomes available - hopefully sometime in 2015 - it's going to be incredibly fast. I'd expect it will be twice as fast as a 780ti.

If it's 28nm, no way it'll be 2x as fast. There just isn't enough die space left to occupy. If they skip 20nm and go to finfets, then it should be well more than twice as fast.

 

[Grooveriding]

I don't think that's realistic if gm200 is also 28nm. If a gm204 is roughly the same performance as a 780Ti (give or take 10%), and a gm204 is 420mm2, then a gm200 would have to be on the order of 800mm2 to have double the performance in the case of perfect scaling. Not going to happen, especially if gm200 has extra DP hw like gk110. 50% faster than a 780Ti is a given. 75% is a stretch.

If gm200 is 16nm, all bets are off.

I was hoping on 16nm. Given gm204 is estimated at over 400mm2, and looks to perform about the same as 780ti if rumours/leaks are true, I would assume gm200 would come on 16nm because there is not much more another 100mm2 of die will add imo. I hope so at least.

Gm200 on the next process will be twice as fast as gk110 I believe. The better perf/w and increase in transistors will be impressive.