New details about Nvidia`s Maxwell

Videocardz spoiled the details today.

Here is how GM107 look like


Maxwell now features 128 CUDA cores per SMX. Kepler had 192 CUDA cores per SMX.

Best part is that its looking to be really power efficient despite running on 28nm.

Exclusive: NVIDIA Maxwell GM107 architecture unveiled | VideoCardz.com

 

"Maxwell is said to provide twice the performance per watt of Kepler." :thumbsup:

So, it's not delayed until the end of the year. I knew they wouldn't wait that long, it just wouldn't make sense. It looks like another late Q2 release for the high end GPU's.

 

This I found interesting in that article too:

[copy & pasted from article]
"As opposed to previous leaks, the die size of GM107 is even smaller, not 156 but 148mm2. Compared to GK107 the density of CUDAs per mm2 has increased roughly by 30%. The density of transistors increased by 15%. Remember, this is all on the same fabrication process."

So, Cloudfire, you worked out in another thread that each core was about 20% more efficient than a Kepler core on the basis of performance per core. In the above quote "Cores per mm" have increased by 30%, while transistor density increased only 15%, which to me indicates that each Maxwell core is simpler with less transistors than Kepler. If that's the case then Maxwell have been able to get more performance per transistor while at the same time increasing transistor density by 15%. I think this spells really great improvements in performance per mm squared when comparing Maxwell against Kepler - I might even hazard a guess at something like 30% more performance per mm squared. If that stands with the same or less Watts per mm squared then that's really a big achievement.

 

I think you're referring to 20nm Maxwell, the article was referring to 28nm Maxwell (that's how I read it). Maxwell will provide twice the performance of Kepler when shrunk to 20nm, not at 28nm.

 

TDP of GTX 750 Ti is 60W. TDP of GTX 650 Ti is 115W. GTX 750 Ti is atleast 20% faster than GTX 650 Ti. Incredible!
You are absolutely right. To be able to increase density which should normally increase heat, increase core performance while decreasing TDP is nothing but amazing.

I can`t wait to see what GM108, aka GT 840M will be able to do. According to many sites, Maxwell go live February 18th. So fingers crossed.

Also, this is pretty important:
I got my eye on that GM204

 

WOW! Now I'm impressed after you let me know those TDP's, because that was the missing piece in the puzzle for me! I think as long as all these leaks are accurate that we're basing this information from, then Maxwell is going to be amazing, even in 28nm! The numbers seem almost too good to be true, so it does make me question the validity of the leaks - but we shall see! Good find on all the leaks Cloudire!

Oooo yeah! GM204 looking pretty good!! (You watch me stick with my GK104 until some Volta goodness though!) GV304 FTW!!

 

I am extremely eager to see what the Lenovo Y50 will be able to do now.

 

More smoke and mirrors. The fact that Maxwell is not 20nm is disappointing enough. I'll just hold out and wait and see what happens. This has been a long road, and the light is seeming more of a dim candle than the blazing sun.

 

It is promised to have significant performance/watt improvement already on 28nm die and if that is true, that is super. And according to Fud, another 30% improvement coming with 20nm... I'm actually very happy to Maxwell, unlike AMD's new Radeons, these not re-names.

 

Wax On...Wax Off! Yes, I enjoyed your last sentence, it made me laugh, not in a bad way...tech can be so...spiritual sometimes. ;-)

Wouldn't a move to 20nm mean more like a 100% improvement? Theoretical calculations suggest that you could fit nearly twice the number of transistors per mm squared going from 28nm down to 20nm, so if FUD say that a move to 20nm will only give a 30% improvment in performance then I think they're way out.

 

That's assuming that the die size is the same though. When the initial Kepler chips came out, they were smaller than their Fermi counterparts.

 

Yep, that taken into account, then yes.

 

I present to you, GTX 860M

Two scenarios here:
Either GTX 860M is a Kepler rebrand. Note: We don`t have any 1152 core GPU with 128bit bus and 16 ROPs. The only GPU that is close is GTX 760 with 1152 cores, 192bit bus and 32 ROPs.

Or GTX 860M is based on Maxwell. GTX 750 and GTX 750 Ti are both Maxwell and run a 128bit bus.
Note that GPU-z is incapable of detecting the correct core count if this is Maxwell because it thinks 1 SMX = 192 cores, but with Maxwell 1SMX = 128 cores.

So GTX 860M have in total 768 cuda cores if based on Maxwell. Compare that with GTX 660M which had 384 cuda cores.

 

I love you Cloudfire. Where do you even get all that info from though?

 

The internet

-----------------------------------------------------------------------

Here is GTX 870M.
All credits goes to: 慕容蛛蛛

It have the core count as GTX 680M but does only have a 192bit bus. What the hell...

 

GM 200 is 20nm.

It should be available in May/June time for desktop in very short supply. Ak, Q3 release for full availability.

PS: GM 107 is a Kepler reminiscent, Kepler could have been tweaked to archive less power/watt; Maxwell GM 107 can disregarded.

 

GM107 far from Kepler. It is clearly a brand new architecture.

Read the first post again.

 

If this is true, we could see thinner laptops with top end gpus. After all if 750M was 60 watts, and these are also, then the cases that held a 750M will hold these.

 

GM 107 is Kepler reminder. Read my post carefully.

 

Maybe you're both right? It's new and not new at the same time. Maxwell is an evolution of Kepler, the same way that Kepler is based off of Fermi, which is based off of Tesla, etc. Nvidia hasn't had a completely new uarch since, what, G80?

 

You are right.

In terms of features, the "real" Maxwell will be an ARM core orientated(more changes made than gm 107 offers). gm 107 is squeezed into Maxwell, but it is not a proper presenter of it.

 

The post in the first page should have cleared this up although someone here insist on mention Kepler to spread confusion.

GM107 is a denser chip than GK107 despite both on 28nm.
GM107 have TDP of 60W against GK107 with 115W, despite both on 28nm.
More performance/core
New SMM to replace SMX with less cores/SMM
More instructions per clock cycle, which is the corner stone of a new architecture

How is it not a new architecture?

If they almost sliced TDP in half (although I`m sceptical), its looking pretty promising for mobile graphic cards where TDP is everything

 

It is a new architecture. And 'someone' said that Maxwell is reminiscent of Kepler. Which is correct. Even the article in the OP says "You can find here references to both Fermi and Kepler."

Also, where are you getting the TDP for GK107? The GTX660 has a TDP of 64W and the GT650 has a TDP of 65W, with some other GK107 cards going as low as 50W.

 

meaning its just a rebrand of the 760/765m

 

Remember the 680m had 900mhz ram and the 870m if true with 192bit 1250mhz would just be at a rop disadvantage to the 680m placing it between the 770m and 680m.

 

Cloudfire, If the GTX 860m IS maxwell and IS 768 cores then won't that mean that it isnt GM107 and hence, might be 20nm ??
Considering the fact that videocardz claims the full GM107 is 640 cores...

 

You are probably mixing here the mobile versions TDP, but we do not have any details of mobile Maxwell power consumption yet. Instead we know the desktop GM107 consumes only 65W, the desktop GK107 115W. As Cloudfire is saying it is huge improvement!

 

gm 107 is a middle ground between Kepler and Maxwell(which is to come soon).

 

GTX 660 have a TDP of 140W. I compared GTX 750 Ti (60W) against GTX 650 Ti(115W), since benchmarks have revealed that 750 Ti is beating 650 Ti by 20%. I wrote GK107 but 650 Ti is actually GK106, but that doesnt matter. What is amazing here is the reduction of the TDP by that much and its beating the 650 Ti.

GTX 650 do have a TDP of 65W but GTX 750Ti score over 2x as much in 3DMark11 (2711 vs 5611). Which is why I didnt bother to mention it lol

It is Maxwell.
What tells you that this is not Maxwell? Lacking ARM processor?

No it couldnt be further from a GTX 765M.
If GPUz is reading the GPU wrong, it have 128 cores/SMM, meaning its Maxwell not Kepler.

If GPUz is reading the GPU correctly, it is Kepler, but it have 1152 cores vs GTX 765M 768 cores.

Good spotted, if it is Maxwell it means it have 1 SMM more than GTX 750ti. Which Videocardz said was a full GM107. If it is Maxwell then yeah it must mean its a different chip. Maybe 20nm like you say although Im sceptical. But who knows. All mobile Keplers was based on 28nm when they came. They didnt do any Kepler in the old 40nm

But lets be open to that it could be Kepler too.

 

However it turns out it is probably best to wait a little longer for 20nm. If what I read is true than a 100% jump in performance is worth ignoring a mere 20% increase on 28nm. Impressive as it is a 28nm maxwell appears to simply offer the opportunity to throw money away on it! The same applies to anything amd try to foist on us!


Sent from my iPod touch using Tapatalk

 

I agree with you King. 20nm should def be everyones primary choice.

----------------------------------

We should get correct GPU-z screenshots from GTX 750/750Ti/GT840M/GTX860M any time now.
GPU-z just released a new version with correct detecting.

GPU-Z 0.7.7 changelog

GPU-Z 0.7.7 Released With Improved NVIDIA Maxwell Support | Geeks3D

 

Holy crap guys. Its confirmed, Maxwell will be amazing.

From Nvidia APAC conference:

First of all, we are looking at 2x the performance/watt compared to Kepler. With the first generation Maxwell!


Second, GTX 750 Ti is confirmed to be a 60W GPU.


For those who don`t know, GTX 750 Ti is beating a GTX 650 Ti by 23%. The best part is, GTX 650 Ti is a 115W card while GTX 750 Ti is now a confirmed 60W card....!!!

Hello beastly Maxwell mobile graphic cards!

More slides here:
NVIDIA Maxwell GeForce GTX 750 Ti and GTX 750 Official Specifications Confirmed, 60-Watt GPU - GeForce 800 Series Arrives in Second Half of 2014

 

That's pretty cool Cloudfire, but something is not fitting for me. NVidia roadmaps show Maxwell to be twice the performance of Kepler per Watt, but in those roadmap slides they were referring to 20nm Maxwell (roadmap slides that came out months ago). Now, how can it be 28nm Maxwell that is showing twice the performance per Watt of Kepler, because then if they shrink that to 20nm then it would 4x the performance per Watt of Kepler. This info doesn't add up with the NVidia roadmap projections. All seems kind of strange, and also I just don't know how they can get twice the performance per Watt of Kepler while staying on 28nm - I just don't think that can be possible. If those slides you were posting were referring to 20nm Maxwell then I could believe it. It just all doesn't all add up to me. Something tells me I'm only going to believe it when I see detailed reviews of these Maxwell 28nm cards from the likes of Guru3D.

 

Yeah, those have to be referring to 20nm Maxwell.

 

Didn`t you guys notice my earlier post?

GTX 650 (non Ti), TDP = 64W
3DMark11 score: 2759

GTX 750 Ti, TDP = 60W
3DMark11 score: 5611

Thats over 2x the score while having lower TDP.
Its already happening.

Look at the first slide I posted. They state its from first generation. First generation Maxwell is GM108/GM107 I think. Second generation is GM206/204/200 which was listed in Nvidia`s driver a week ago. 20nm GPUs.

Its pretty simple: They don`t have to go down to 20nm to fit more cores in the chips. GM107 is already pretty small (144mm^2) and have only 640 cores. When we get up to 3000+ core count (GM20x series), they will need 20nm because we are talking a lot of transistors on a small die. We have all seen how hot R9 290x became when they fitted more transistors on a small (ish) die. It was designed to run at 94C...

 

I know, I remember your previous post, I'm still sceptical, I just can't believe that they can double performance per Watt through architecture alone without shrinking the process down to 20nm. It also doesn't fit with NVidia's Roadmaps that they released. I think it's great that you're digging up all this stuff, it's really interesting, and I like talking about it, but I'm sceptical I'm afraid.

 

A reseller from UK tested GTX 750 Ti
NVIDIA GeForce GTX 750 and GTX 750 Ti preview leaks out | VideoCardz.com

TDP of GTX 750 Ti is listed here:
http://i.imgur.com/s4Y3LiZ.jpg

GTX 650 3DMark11 score can be found here:
2012 Budget Mini-ITX Desktop PC System Build Guide - Page 9 of 10 - Legit ReviewsBenchmarks, Temperatures & Power Consumption

TDP of GTX 650 is listed here:
AnandTech | NVIDIA Launches GeForce GTX 650: GK107 With GDDR5


Launch is like in 3 days anyway. We will get definitive proof then but its looking pretty promising. No scratch that. Amazing
Since GTX 750 Ti is 60W, one can expect pretty low TDP of GT 840M.

 

Oh right, great, I'd be interested to read the official reviews when it comes out next week then - I will be amazed if it actually turns out to be twice the performance per Watt while at 28nm, but if it is then that's surely a good thing anyway!

 

Wow I scratch my previous comment if they can release maxwell at 28nm at more 2x the performance per watt of Kepler!! 20 nm won't be needed for another year even for high end mobile chips. I wonder how and will respond. I like to see a little competition in the market to drive prices down ;]


Sent from my iPod touch using Tapatalk

 

That's it, my next computer will be a desktop.

 

Maxwell is so huge improvement already on 28nm, which hard to believe, it is like Nvidia did something impossible. Anyway it seems we are looking forward to very awesome mobile GPUs, cannot wait for 840M and other Maxwell mobile GPUs. I just wonder AMD has any answer for this or they going to loose GPU battle too?

 

from wccftech courtesy tinhte.vn

 

Considering Kepler is 2 years old (gcn too) I think a large improvement is expected.


Sent from my iPod touch using Tapatalk

 

Just wait for 20nm for the difference.

These 2 marketing slides, not sure why people treat it properly.

780 Ti according to NV is 30% then 290x .when tested by reviewers, only 13%

 

Indeed, they always kind of take the best case gains and show them to the press

 

I know: 640 vs 750Ti and GTS 450 vs GTX 750? What the heck are they trying to show?

 

That you can get 2.5x as much FPS while going down 5W.
That both GT 640 and GTX 750Ti can run on PCI-e which is why they compared them. And therefor be used in HTPC, but you get way bigger performance with Maxwell against Kepler.

 

Did anyone notice the slide where they said Maxwell give 2x the performance/watt but also said performance/core is +35%?

That means if you compare a 1500 core Kepler against a 1500 core Maxwell, you get 35% better performance and the wattage go down 50%.
This means they can do 28nm on small dies, but to increase performance over 35% they will need to go up in core count. Which means they will need to go down to 20nm to fit those transistors in to the die sizes we have today.

Which is why Nvidia is only doing 28nm on GM107 and GM108 which is small die's.

 

The 2x perf/watt should already include the 35% extra perf/core. Otherwise they would say 2.7x perf/watt, no?

 

Perfect! Time to buy a new GPU for my desktop