Upcoming GTX 750 Ti tested

[Beavermatic]

I guess im still trying to determine if the 750ti is even Maxwell at all. Is there any confirmation of this other than some random leaked pics that may/may not be spoofed?

 

[Cloudfire777]

There is none GPUs that share the same specs as the GPU-z screenshot above.
960 cores, 16 ROPs and 128bit.

"GPU: 1380", does that one exist? Does that number mean anything or is it the "Device ID" thats important? Strange that he didnt reveal that ID...

 

[sushiwarrior]

I guess im still trying to determine if the 750ti is even Maxwell at all. Is there any confirmation of this other than some random leaked pics that may/may not be spoofed?

This is definitely the 750Ti, definitely Maxwell too, other benches will show that is has the integrated CPU (compute performance in certain apps goes way up, texture compression helps 128b bus at high res in certain games, etc.)

 

[Beavermatic]

There is none GPUs that share the same specs as the GPU-z screenshot above.
960 cores, 16 ROPs and 128bit.

"GPU: 1380", does that one exist?

Im not trying to sound like an idiot, but isn't that the case with any new card? I mean is that info enough to establish that its the Maxwell architecture and not just a revised kepler of some sort? I mean "any" new card is going to vary a bit, and I see a lot of kepler similarities there in those charts as well.

I guess im just a little skeptical of a 7xx series card (not to mention a entry level one) being released as the intro to a new architecture. That would be a rather "comes in with a whimper" introduction for the Maxwell lineup, which I presume would be under a new 8xx series. I've never known Nvidia to cross architecture lines in a series outside of some instances of mobile (where the current series in mobile is actually based on the last series desktop equivalent architecture)

 

[Cloudfire777]

Im not trying to sound like an idiot, but isn't that the case with any new card? I mean is that info enough to establish that its the Maxwell architecture and not just a revised kepler of some sort? I mean "any" new card is going to vary a bit, and I see a lot of kepler similarities there in those charts as well.

I guess im just a little skeptical of a 7xx series card (not to mention a entry level one) being released as the intro to a new architecture.

You are correct, the GTX 750 above and various Kepler cards does indeed share a lot of similarities. No doubt about that.
Its just that the specific ROP/memory bandwidth like that, it doesnt exist and why would Nvidia make a new Kepler when Maxwell is so close?

They could do a GTX 660, remove 64bit from the memory bus which they have done on countless GPUs before, but I`m not sure if they can just disable 8 ROPs..?
I also agree that it seems strange to release a Maxwell in the already exisiting Kepler 700 series.

But its Nvidia. Always chaos with their rebranding and their plans.

 

[Beavermatic]

why would Nvidia make a new Kepler when Maxwell is so close?

I want Maxwell ASAP too, but that's kinda what I'm getting at. There's so much bouncing around out there (its 2 months ahead of schedule -or- its not gonna be released until mid summer) and a lot of speculation with no concrete info to back it up. In fact, Nvidia has remained completely silent on it. The fact they are not talking about it is more proof... They're still focused on the 7xx and Titan series. When they start talking publically about it, then in the past, its a sign its right around the corner.

It doesn't help with the fact they just recently released the 780ti's, and now talks of Titan 2's and 790's and whatnot, which are also all just speculation and rumor (and if they do exist, likely still based on Kepler chips). Just a lot to throw into the mix of confusion and what architecture and when. I mean they could have made some drastic modifications to kepler, as a year and a half later, and we are now just getting releases with unlocked SMX's and such.

Also keep in mind... why release the Maxwell series when your current recent lineup is the fastest out? Especially only 2 months apart. Would just cannibalize sales of your current marketing effort and lead in the market and makes no sense from a financial aspect.

Just too many similarities in the comparisons with kepler for me to make a judgement call on what it is just yet, personally.

 

[Skurge]

I can confirm those benches are legit.

 

[Sohaltang]

I guess im still trying to determine if the 750ti is even Maxwell at all. Is there any confirmation of this other than some random leaked pics that may/may not be spoofed?

I agree. Looks like they just created a GHZ edition 650TI.

 

[Techhog]

I can confirm those benches are legit.

Disgusting

 

[ams23]

These purported GTX 750 Ti specs would mean that CUDA core count and mem. bandwidth is exactly 5x more than the Kepler.M GPU in Tegra K1 (and presumably texture and pixel fillrate would be similarly high in comparison).

The TDP of this card should be WAY lower than GTX 650 Ti.

 

[Arkadrel]

This card is ment to compete against the 260x ?
Both are small 128bit bus cards.... this looks slower though, so I guess it ll compete on price.

Thats gonna be good for consumers then, it ll drive down the 260x price.

 

[OCGuy]

I can confirm those benches are legit.

Can you tell us how?

 

[OCGuy]

They are obviously about to go public with the mobile 8 series, which is just a re-branded Kepler, as the rumors said. Messing around in the driver section I see this. In that GPU screenie, doesn't it list an official driver? Wouldn't that be odd for such a new arch?

The spec sheet on the 820m shows 11.2 support with no qualifier like the other cards though....hmmm.

 

[ams23]

Assuming that GTX 750 Ti leverages the Kepler.M architecture used in TK1, then the TDP should be way lower than R7 260X too.
,
On an aside, GTX 750 Ti appears to have a bit more GFLOPS throughput than the PS4 GPU (with lower TDP too of course).

 

[OCGuy]

780m features page, see DX11 -

820m -

http://www.geforce.com/hardware/notebook-gpus/geforce-820m/specifications

 

[boxleitnerb]

According to sweclockers both, 750 and 750 Ti perform about 25% better than the 650 and 650 Ti and both require no additional power connectors. If that is true, then Maxwell has some crazy energy efficiency, given that it is still on 28nm.
http://translate.google.com/transla...0-kraver-ingen-extra-stromforsorjning&act=url

Sweclockers were always very well informed iirc.

 

[f1sherman]

According to sweclockers both, 750 and 750 Ti perform about 25% better than the 650 and 650 Ti and both require no additional power connectors. If that is true, then Maxwell has some crazy energy efficiency, given that it is still on 28nm.
http://translate.google.com/transla...0-kraver-ingen-extra-stromforsorjning&act=url

Sweclockers were always very well informed iirc.

woof those are some impressive, but necessary gains in order to deliver on 20nm later on

its been a long-long time since we've seen what purely architectural gains can do.

 

[Cloudfire777]

They are obviously about to go public with the mobile 8 series, which is just a re-branded Kepler, as the rumors said.

Incorrect.
GT 820M is actually Fermi.

 

[Cloudfire777]

Disgusting

GT 750 Ti is looking to be around 25-30% faster than GTX 650 Ti.
That is actually pretty good for just moving to a new architecture without going down to 20nm. But this depends on wether the gains come from extra power consumption or if it stays the same as 650 Ti.

Imagine when we get 20nm Maxwell where they can shove in almost 2x as many cores as Kepler.

 

[ams23]

@ boxleitnerb: I'm not surprised by that because Kepler.M (in Tegra K1) appears to be super energy efficient while also using a 28nm fabrication process. The million dollar question is: is GTX 750 Ti a Kepler.M derivative with Kepler's feature set or with Maxwell's feature set? The former is certainly a strong possibility.

Note that, while this rumored GTX 750 Ti has exactly 5x more CUDA cores and 5x more mem. bandwidth than Kepler.M in TK1, it does appear to have 2x more TMU's per SMX and 2x more ROP's per 32-bit mem. channel too (not to mention much higher GPU clock operating frequencies too) compared to the TK1 variant.

 

[OCGuy]

Incorrect.
GT 820M is actually Fermi.

Those drivers don't support 820m, the DX info is incorrect, GF117 was on the 28nm process and was the base of the 630M, and 720M but is faster.
Just a couple of the GPU-Z inconsistencies. If that is correct it is just a re-badge of a die-shrink with a bigger memory interface and the same clocks....snooze.

 

[Cloudfire777]

^The only error there is that it should be 28nm instead of 40nm. The guys who make GPU-z promised an update several years ago which should fix that after several GT 620M complaints.

Other than that, GT 820M is exactly like described in that screenshot. ie old as hell

 

[el etro]

Performance/stream processor of Fermi was X.

Performance/stream processor of Kelper is much less than X.

Viewing this scores, seems than performance/stream processor of Maxwell will be lower than Kelper's.




Nvidia have learned from AMD's architectures(since R700) and maked every stream processor the lesser complex that they can built. As a result, they end with a less GPGPU capable architecture, but ends too with a higher perf/watt architecture.

 

[ams23]

First of all, it is "Kepler", not "Kelper". The GPU is named after a scientist, not a seaweed. Second of all, perf. per stream processor is somewhat irrelevant without taking into account area and power consumption (not to mention the fact that Fermi had double pumped stream processors running at roughly 2x base clock). Three, the stream processors in this purported GTX 750 Ti have the same performance as any other stream processors in any other Kepler GPU. Fourth, the key advantage with Maxwell is perf. per watt, which is a result of the low power work done in Tegra (with Kepler.M and Maxwell.M) and not related to ATI. Last but not least, Kepler (and Maxwell) are tremendously capable with GPGPU in the high end variants that enable higher DP throughput (and in fact, the ten most green supercomputers in the world all use Kepler GPU's). The SP throughput of all Kepler GPU's is quite good too (relatively speaking).

 

[el etro]

It seems legit that older claims about Maxwell big chip pack 6000 stream processors.
If maxwell needs 960 sps clocked at 1100Mhz to beat 768 Kelper sps clocked at 928+boost Mhz and consumes less at the same 28nm process, here's the magic.