Upcoming GTX 750 Ti tested

[Magic Carpet]

ARM is most likely Telsa only if any at all. No ARM in GM107.

Not that it mattered anyways. And I dont think I have seen anything official related to Maxwell dGPUs supposed to have ARM chips. Might simply be a misunderstanding with the Tegra chips.

I see. Well, the idea of having ARM co-processors to remove drivers overhead is more robust (compared with Mantle). Can't wait for the first products with this philosophy behind.

I dont think so. They want to push CUDA so i doubt this. Its quite sad because they probably could match AMD's performance or atleast have somewhat respectable performance except they don't choose to. The drivers are just not even touched for openCL.

Yeah, for OpenCL AMD is a better buy. Also, as it appears, 1st Gen Maxwell is still lacking AMD's ZeroCore Power equivalent tech, which is quite sad since its aimed to maximize power savings at all levels. I love how both of my 270's shut down their fans when I leave my PC unattended. Happy to see nVIDIA retaining VGA output for legacy purposes. Not being able to SLI is about the only serious drawback of this card (two of these sipping power in SLI would make for a brilliant power-efficient gaming system). For HTPCs and such, this card is overpowered and overpriced, when modern APUs have taken over this market anyway.

 

[frozentundra123456]

Ok, lets calculate. Numbers from AT Bench, Crysis 3.

power
750 Ti: 184W
640: 152W

fps
750 Ti: 60,6
640: 22

perf/w
750 Ti: 0.33
640: 0.14

Yes, 750 Ti has 228% efficiency of GT 640, so it actually has more than double perf/w.

But how about GTX 650? That's also GK107.

power
750 Ti: 184W
650: 176W

fps
750 Ti: 60,6
650: 38.8

perf/w
750 Ti: 0.33
650: 0.22

750 Ti has 149% efficiency of GTX 650, not quite the 200% we are looking for.

What this mostly shows is that GT 640 is just pretty inefficient.

How about the card it's closest to in perf, GTX 650 Ti Boost?

power
750 Ti: 184W
650 Ti Boost: 236W

fps
750 Ti: 60,6
650 Ti Boost: 72.9

perf/w
750 Ti: 0.33
650 Ti Boost: 0.31

750 Ti has 107% efficiency of GTX 650 Ti Boost.

Yes, the numbers are skewed because of the high power usage of the test system. But it's not going to turn the 149% against GTX 650 into 200%.

This calculation needs to be done based on the power drawn by the gpu, not total system power. Obviously, the power drawn by the rest of the system is a bigger part of the total power usage for the low end cards, so will make their performance per watt look artificially low.

 

[coffeejunkee]

I would if I had the numbers. Also notice the last sentence. But ok, I should say the '750 Ti system' vs the '640 system' etc.

Still, I think it gives a good indication that GT640 is just not very efficient. It's easier to improve your perf/w by a factor 2 if you compare to your least efficient card. That's what the marketingslide is showing, but many people will say 'oh look, maxwell is twice as efficient as kepler'.

Don't get me wrong though, I think it's still a very good job by Nvidia engineers.

 

[Magic Carpet]

This is just a wild guess, but I think... Maxwell is like Haswell, very efficient at low to medium loads but loses its efficiency as it gets "bigger". Also, explains why nVIDIA decided to start with this chip instead (to show off and make some good PR). First impression is very important

 

[24601]

This is just a wild guess, but I think... Maxwell is like Haswell, very efficient at low to medium loads but loses its efficiency as it gets "bigger". Also, explains why nVIDIA decided to start with this chip instead (to show off and make some good PR). First impression is very important

Haswell loses no efficiency by "getting bigger"

It loses efficiency at high clock speeds/voltages.

Saying Haswell loses efficiency by "getting bigger" is basically saying that the 10-12 core Haswell will be less performance per watt than the 4 core Haswell, which is the opposite of the truth.

The same result (increase in efficiency) will happen when Maxwell is scaled up, as that's how GPUs work.

 

[Magic Carpet]

Haswell loses no efficiency by "getting bigger"

It loses efficiency at high clock speeds/voltages.

That's what I meant. Sorry for the confusion.

The same result (increase in efficiency) will happen when Maxwell is scaled up, as that's how GPUs work.

Time will tell, if and when scaled up Maxwell will maintain =>200% performance per watt efficiency over its replacing appropriate counterparts. Obviously, they needed to close the gap in the 700 series, but this is a very strange release otherwise. Especially, when the most significant power savings are to be made with the bigger power envelopes. Or perhaps, they want to milk us, the consumers, a bit more, with the older tech, while annoy the AMD leadership at the same time. Very smart indeed.

 

[f1sherman]

Time will tell, if and when scaled up Maxwell will maintain =>200% performance per watt efficiency over its replacing appropriate counterparts. Obviously, they needed to close the gap in the 700 series, but this is a very strange release otherwise.

How so?

Replacing big GTX 780 Ti and K40 which are only few months old would be strange.

 

[Magic Carpet]

How so?

Replacing big GTX 780 Ti and K40 which are only few months old would be strange.

That 780 Ti flagship is based on older tech, they could have re-badged a 650 Ti or Boost as 750 Ti, for example, the same they have done to 680 and 670. I suppose, this is how it's going to be from now on. Okay.

http://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units

The "big chip" had always been based on the latest tech and out the gates 1st, until this release. You don't find that strange, okay

 

[f1sherman]

Based on old tech... I did not even mention Titan Black. Because that's somewhat of a rebadge.

K40 and 780 Ti are fully fledged and meaningful products, not mention top dogs in its class... hows that for "Based on old tech"

 

[jpiniero]

Has anyone done any underclocking/undervolting results? Kind of curious what the clock speeds would be for a theoretical mobile part.

 

[ams23]

This is just a wild guess, but I think... Maxwell is like Haswell, very efficient at low to medium loads but loses its efficiency as it gets "bigger". Also, explains why nVIDIA decided to start with this chip instead (to show off and make some good PR). First impression is very important

The efficiency gains will filter through to the high end too because the perf. per CUDA core is way up, and the perf. per mm^2 is way up (not to mention perf. per watt too). And the crazy thing is that this is on 28nm HP fab. process (higher end variants will eventually use more advanced fab. processes). Anyway, as a "mobile first" GPU architecture, it makes sense that Maxwell would be first introduced as a lower end or midrange variant for desktop and mobile platforms.

 

[ams23]

On a side note, I noticed in a Newegg youtube video that GTX 750 Ti FTW (with ACX cooler) is actually faster than GTX 650 Ti Boost at 19x10 resolution with high details!

 

[tviceman]

Also, explains why nVIDIA decided to start with this chip instead (to show off and make some good PR). First impression is very important

How easily people everywhere are forgetting (or did not even realize) that GK107 was the first Kepler chip to hit the market.

 

[BlockheadBrown]

A friend has a 300w power supply and was looking to get a 750 Ti based on the numerous reviews and even Nvidia's site which states only a 300w is required. Yet, when going to purchase any of the available 750 Ti cards, all state a 400w requirement. Are manufacturers just being overly cautious or is there more to this?

 

[nwo]

A friend has a 300w power supply and was looking to get a 750 Ti based on the numerous reviews and even Nvidia's site which states only a 300w is required. Yet, when going to purchase any of the available 750 Ti cards, all state a 400w requirement. Are manufacturers just being overly cautious or is there more to this?

Yes, they are being overly cautious. A 750Ti uses less than 100W at full load (less than 70W while gaming without OC), even when OCed... . However, most 300-400W PSUs are garbage, especially if they are in OEM PCs or the cheap ones that come with a case.

For instance, my 500W Raidmax PSU that came with the case which I bought several months ago has only 22A on the 12v rail which is 264W. That means it's only able to provide about 50% of it's total output on the 12v rail, and even that is questionable. A good PSU should be able to provide at least 90% of it's total output on the 12v rail(s).

You need clean & stable power in order to run a gaming video card smoothly.

If your friend has a reputable 300W PSU (something like Antec) and has at least 20A on the 12v rail (which will be overkill but anyway...), he will be fine.

 

[el etro]

A friend has a 300w power supply and was looking to get a 750 Ti based on the numerous reviews and even Nvidia's site which states only a 300w is required. Yet, when going to purchase any of the available 750 Ti cards, all state a 400w requirement. Are manufacturers just being overly cautious or is there more to this?

I think any 300w power supply can drive the 750Ti. Nvidia asks for 20A(is more or less 240w or more at 12v PSU rails) on the 12v lines of the PSU that will drive the card.

 

[ams23]

This calculation needs to be done based on the power drawn by the gpu, not total system power. Obviously, the power drawn by the rest of the system is a bigger part of the total power usage for the low end cards, so will make their performance per watt look artificially low.

Yes, isolating GPU power is essential in order to accurately gauge GPU perf. per watt. When isolating GPU power, GTX 750 Ti is ~ 50% more efficient than GTX 650 Ti Boost in Anandtech's Crysis 3 benchmark (assuming that the GPU power consumption of the former is ~ 65w). And that is just one [bandwidth-limited] game. IIRC, Tom's Hardware sampled five games and found GPU perf. per watt of GTX 750 Ti to be 2x (!) better on average than 650 Ti Boost!

 

[coffeejunkee]

About 154% is what I get too if I calculate gpu power only vs 650 Ti Boost. I assumed 50W cpu power because it neatly fits with the cards tdp.

Obviously that's a great improvement but it's not close to 200%. So you think that other games perform so much better they do turn it into 200%? I can't really check which values Tom's used. What I do know is that you can't just take one game's power usage and then calculate perf/w for all games based on that like tpu does.