Official Nvidia Maxwell slides

T

tviceman

Limp Gawd
Joined
Oct 11, 2007
Messages
398
Maxwell is officially BEAST mode. Can't wait for the bigger dies to come.


lpqJwdY.jpg

2djZi8x.jpg

bFsNWD4.jpg
 
I guess that comparison was to show the power pet watts right ? anyway looks sick
 
Too hard to really make sense of all of it. I'd rather see the difference between the 650 Ti not the 640. That's not really a fair comparison at all. TDP is nice though.
 
Liger88 said:
Too hard to really make sense of all of it. I'd rather see the difference between the 650 Ti not the 640. That's not really a fair comparison at all. TDP is nice though.
Click to expand...

Lol, why would they show the 750ti against the 650ti? That would mean the 750ti is only marginally faster...;)
 
Well, the main takeaway is the incredible efficiency. Supposedly the 750ti matches the 7790/6870 (or comes really close) while having a 60W TDP, which is incredible from an efficiency standpoint - and this is on 28nm.

Should have good implications for big Maxwell. GM200 on 20nm should be pretty beastly once it is scaled up. Essentially, GM107 is doubling the GT640's performance at a lower TDP - the GT640 is a 65TDP, while the 750ti is a 60W TDP. Over 2x performance with less power consumption....
 
Last edited:
Sodapopjones said:
Ok... I'm supposed to be excited right?
Click to expand...
Ionno if it folds nicely which i doubt it, but arch changes could mean changes there as well, might be a nice thing to shove a couple 20nm maxwell cards for efficiency. I guess it's mostly excising for laptops graphics. And depending on how it scales might work wonders for their flagship at 20nm after all more performance often comes in form of efficiency, as graphics card coolers can only dissipate so much heat and thus watts so improving performance per watt does more and more as you scale up.
 
So much for thinking the 750 Ti would be equal to the 650 ti boost it most likely will replace as it appears discontinued.
128bit memory bandwidth is going to choke down it's potential.

Not keen on running full power off the pcie either which is why I read recently for overclocking headroom some companies will probably add a 6 pin regardless.

They are noted to be able to SLI so if they are cheap and you combine two then maybe you have something.

3rd quarter before we see any type of 860 which should be slick.
 
Liger88 said:
Too hard to really make sense of all of it. I'd rather see the difference between the 650 Ti not the 640. That's not really a fair comparison at all. TDP is nice though.
Click to expand...

Well, if you want to compare Kepler vs Maxwell 650 Ti vs 750 Ti would not be ideal, that is GK106 vs GM107.

I think:
650 Ti is a GK106
650 non-Ti is a GK107
750 Ti is a full GM107
750 non-Ti is not a full GM107
640 GK207 or GK107

So actually 750Ti vs 640(GK207 version) or 650 non-Ti would be most interesting, if you want to compare Kepler vs Maxwell.
 
Nvidia supposedly confirmed at the presentation that the desktop 800 series with higher end Maxwell will be in 2H 2014.
 
2H, let's hope maybe around May then for a card you can actually push in recent games.
Guess it depends if they can get 20nm working quicker as a 28nm would just a stop gap.
 
tonyftw said:
55-60W tdp is impressive for the punch those two cards offer.
Click to expand...

Precisely. This has great implications for 20nm big Maxwell when it hits - Maxwell is designed as a scaling architecture just like Kepler was.

So in the end, there are a lot of rumors of how well the 750ti performs. I have heard from a few rumor sites that it scores 3000 on firestrike which puts it even with a stock 480; now I don't know if that is true , but if it is, it's pretty fucking amazing for a 60W TDP. Just for comparison sake, the 7850 is a 130W TDP and the new R7-270X and R7-265 are both a 150W TDP. That means, if that firestrike score is true, that it is achieving close to the same performance while using a 60W TDP compared to AMD's 130 or 150W TDP. That efficiency will be amazing given the relative performance level, IMO. Now i'm not sure if the firestrike score is accurate. But going by NV's slides here, NV did double performance per watt. They're getting 2.5 times GT640 performance in a 60W TDP, which has great implications for big Maxwell, as mentioned.

This fall will be exciting for sure whenever GM200 hits. I assume that GM200 is simply waiting on TSMC's 20nm, but who knows when that is - I hope sooner rather than later. Given how well these low end parts are doing in terms of efficiency - once scaled up to GM200, it should be straight up beast mode.
 
The 750Ti hitting GTX 480 performance level at 60W TDP is the most amazing thing about this. Why? I remember reading the power consumption of the GTX 480 and how high it could get at full load.

It is amazing, to me, how much of an advancement we've come so far in terms of efficiency and performance. Imagine bringing this forward in a few more years and we could get a 780Ti under 150W TDP in something like a future Geforce 950Ti for the price of $150 to $200.

Wishful thinking, but at this rate I wouldn't be surprised it would happen given how the 750Ti looks like so far.
 
It'll be interesting to see how Nvidia handles the 800 series as well as subsequent 900 series. Will they hold back the largest version of Maxwell like they did with Kepler? If not will 800 series be largely disabled? How will they refresh for the 900 series then before Volta?

Also what will the memory bandwidth situation will be. Volta is at the moment said to get a large bandwidth jump via stacked memory but currently GDDR5 is basically pushed to the limit now, it is highly unlikely they can raise memory bandwidth further simply via clockspeed increases (as with the previous few generations). Will mainstream performance finally move beyond 256 bits? Will Nvidia also go to 512 bit for the high end? Possibly GDDR6 ready?
 
limitedaccess said:
It'll be interesting to see how Nvidia handles the 800 series as well as subsequent 900 series. Will they hold back the largest version of Maxwell like they did with Kepler? If not will 800 series be largely disabled? How will they refresh for the 900 series then before Volta?

Also what will the memory bandwidth situation will be. Volta is at the moment said to get a large bandwidth jump via stacked memory but currently GDDR5 is basically pushed to the limit now, it is highly unlikely they can raise memory bandwidth further simply via clockspeed increases (as with the previous few generations). Will mainstream performance finally move beyond 256 bits? Will Nvidia also go to 512 bit for the high end? Possibly GDDR6 ready?
Click to expand...
A lot of good questions.

I have not heard anything about GDDR6 and I don't know what the next evolution of video card memory technology will be that'll provide more bandwidth and speed than current GDDR5.

And, many mainstream cards are still at 128-bit. If Nvidia did come out with a 950Ti with the performance of a 780Ti in a few years but gimp it with a 128-bit bus, I'd think that would be doing a disservice to the original 780 card series. It'd be great to see a sub-$150 mainstream card with a 256-bit bus at minimum.
 
This is the slide that is the key, and I am surprised wasn't in the OP:

8v51.jpg

(thanks Videocardz.com)

Nvidia is claiming first-gen 28nm Maxwell is 35% more performance per core, twice the power efficiency, *AND* that they can fit that performance in less die area. Add in that Big Maxwell (Maxwell second gen for GM200/204/206 confirmed through drivers) should be 20nm in the second half of this year (hearing rumors it'll be around July but they're just rumors) and we're looking at a monumental performance upgrade like one we haven't seen since the 8800GTX days over the 7900GT.

Think about it.... they're saying a 170% boost per watt, or in other words 2.7x the performance per watt used. And that's without 20nm yet. 20nm will allow a lot more transistors for the die size area, obviously, and the Maxwell SMM's are already smaller at 28nm than the Kepler ones were.

It's going to be one hell of a good ride for GPU's this year if GM200 shows, which it seems very likely that it will on 20nm. I'm looking forward to seeing what the competition brings too.
 
Last edited:
GoldenTiger said:
This is the slide that is the key, and I am surprised wasn't in the OP:

8v51.jpg

(thanks Videocardz.com)

Nvidia is claiming first-gen 28nm Maxwell is 35% more performance per core, twice the power efficiency, *AND* that they can fit that performance in less die area. Add in that Big Maxwell (Maxwell second gen for GM200/204/206 confirmed through drivers) should be 20nm in the second half of this year (hearing rumors it'll be around July but they're just rumors) and we're looking at a monumental performance upgrade like one we haven't seen since the 8800GTX days over the 7900GT.

Think about it.... they're saying a 170% boost per watt, or in other words 2.7x the performance per watt used. And that's without 20nm yet. 20nm will allow a lot more transistors for the die size area, obviously, and the Maxwell SMM's are already smaller at 28nm than the Kepler ones were.

It's going to be one hell of a good ride for GPU's this year if GM200 shows, which it seems very likely that it will on 20nm. I'm looking forward to seeing what the competition brings too.
Click to expand...

That's impressive honestly.

By my thinking, wouldn't that put something like a mid-range future GTX 860 series card somewhere around a GTX 780 performance at less than 200W given the new SMM cores from Nvidia at 20nm?
 
The thing about using a GT 640 as a comparison card is that it's all over the map. It handles some games well enough and handles other games really badly. I have no idea whether COD Ghosts is one of its better or worse games, so I have no idea what that slide is telling me.
 
evilsofa said:
The thing about using a GT 640 as a comparison card is that it's all over the map. It handles some games well enough and handles other games really badly. I have no idea whether COD Ghosts is one of its better or worse games, so I have no idea what that slide is telling me.
Click to expand...

The gt 640 has zip to do with the slide I posted....
 
octoberasian said:
That's impressive honestly.

By my thinking, wouldn't that put something like a mid-range future GTX 860 series card somewhere around a GTX 780 performance at less than 200W given the new SMM cores from Nvidia at 20nm?
Click to expand...

Yes. And that's at 28nm. And remember this is first gen maxwell! The big ones (gm204 and gm200) will be a further revision, on 20nm from the rumors. I imagine amd is going to be cooking up something delicious too if nvidia is pulling this off. All in all 2014 is going to be an incredible year for graphics from both performance and cost, on any given price point.
 
GoldenTiger said:
Nvidia is claiming first-gen 28nm Maxwell is 35% more performance per core, twice the power efficiency, *AND* that they can fit that performance in less die area. Add in that Big Maxwell (Maxwell second gen for GM200/204/206 confirmed through drivers) should be 20nm in the second half of this year (hearing rumors it'll be around July but they're just rumors) and we're looking at a monumental performance upgrade like one we haven't seen since the 8800GTX days over the 7900GT.

Think about it.... they're saying a 170% boost per watt, or in other words 2.7x the performance per watt used. And that's without 20nm yet. 20nm will allow a lot more transistors for the die size area, obviously, and the Maxwell SMM's are already smaller at 28nm than the Kepler ones were.

It's going to be one hell of a good ride for GPU's this year if GM200 shows, which it seems very likely that it will on 20nm. I'm looking forward to seeing what the competition brings too.
Click to expand...

July would mean it is already taped out, possible but not likely, and mean A1 silicon, extremely unlikely. Late Q4 is the more likely scenario with maybe a GM204 in July/August.

Edit- Forgot to mention you are doing your math wrong too.
 
Last edited:
LordEC911 said:
July would mean it is already taped out, possible but not likely, and mean A1 silicon, extremely unlikely. Late Q4 is the more likely scenario with maybe a GM204 in July/August.
Click to expand...

Yeah, as I said, simply rumors on the dates. I'll throw my guess in the pile though and say 860/870 in July (end), and 880 at the beginning or so of September or end of august. Again these are just guesses, based off of personal speculation. :)

Re taping out... Supposedly it taped out at the end of December. Edit and I heard the design has been ready and waiting on 20nm for a long time now.
 
GoldenTiger said:
Yeah, as I said, simply rumors on the dates. I'll throw my guess in the pile though and say 860/870 in July (end), and 880 at the beginning or so of September or end of august. Again these are just guesses, based off of personal speculation. :)

Re taping out... Supposedly it taped out at the end of December.
Click to expand...

Something did, not GM200.

Take a look at the Kepler timeline for why Q3 is overly optimistic for GM200.
 
LordEC911 said:
Edit- Forgot to mention you are doing your math wrong too.
Click to expand...
Huh? I think you are reading it wrong. I said 2.7x As fast by those numbers, or in other words 170 percent faster. Those aren't my numbers but Nvidia's claims of efficiency and power draw potential.
 
LordEC911 said:
Something did, not GM200.

Take a look at the Kepler timeline for why Q3 is overly optimistic for GM200.
Click to expand...

As I said, personal speculation based on no inside knowledge for that one. Q3 is my hope and guess. I'll admit it may be pretty damn optimistic but I think if so it'll be due to 20nm not design.
 
GoldenTiger said:
Huh? I think you are reading it wrong. I said 2.7x As fast by those numbers, or in other words 170 percent faster. Those aren't my numbers but Nvidia's claims of efficiency and power draw potential.
Click to expand...

But we know both manufacturers claim something completely different than what actually happens. I wouldn't expect that much of a gain, it would most likely just lead to disappointment.
 
This is excellent news for me. The aim of my next build would be to reduce the power and heat of my PC, don't want a system that dumps tons of heat into my room while I'm gaming.
 
GoldenTiger said:
Huh? I think you are reading it wrong. I said 2.7x As fast by those numbers, or in other words 170 percent faster. Those aren't my numbers but Nvidia's claims of efficiency and power draw potential.
Click to expand...

No, you are extrapolating performance with those undefined slides.

135% performance per core = 35% increase in SMX performance.

2x performance per watt = undefined. At what level? Shader? Core/SMX? ASIC? Architecture?
 
Looks amazing, but I wouldn't hold my breathe on gm100/110 in the first series unless AMD puts out.

Still hoping I can skip Maxwell altogether for Volta with G-Sync, but if the leap from a 900MHz 470 to a 1300 MHz 780 is dwarfed by what Maxwell is capable of than it will be very hard to pass it up.
 
Man, these die code names are getting confusing. All this GM200 talk I'd assume would be second generation Maxwell? Are people trying to say they expect GM1xx and GM2xx in the same year (800/900 Series), one on 28nm and the other on 20nm. Sounds unlikely if that's the case.
 
You must log in or register to reply here.
Back
Top