Maxwell Gen 1

[John P. Myers]

The specs on the 750 and 750Ti are now official, and somewhat hard to believe.

GTX 750Ti:
640 cores
60W TDP
1020MHz base/1085MHz Boost

GTX 750:
512 cores
55W TDP
1020MHz base/1085MHz Boost

These both use the same 28nm GM107 chip.

Performance? Heh. During Nvidia's conference, they showed a slide comparing frames per second while playing Call of Duty: Ghosts.
The HD 4600 IGP scored 10 FPS
A GT 640 (65W version) scored 22 FPS
And the 750Ti scored 51 FPS

This worries me. The performance increase is simply too good in games. When Nvidia went from the 500 to the 600 series, gaming performance increased alot there too, but the 600 series was worse at crunching. Supposedly these new Maxwell chips in the 750 and 750Ti offer 35% more performance per core while using half the power. That's just too much. Something's wrong somewhere.

[DrPop]

When you say, "something's wrong here" do you mean they are trading off Some kind of raw crunching horse power for a certain optimization in games? I'm not sure what that could be, unless it's something like sacrificing details in a scene for higher fps?

[John P. Myers]

Trading off some kind of raw crunching power, yes. Sacrificing game details wouldn't happen though. It's possible Nvidia crippled their FP64 even more, and possibly their INT too. Maybe there aren't even any FP64 units in these cores at all and all INT and FP64 math is converted to FP32 then reconverted to INT or FP64 for output (essentially an emulator). In games you wouldn't notice this at all since they mainly focus only on FP32. In crunching though, it would be horrid.

When someone makes the correct slot configuration of the 750, i'll get one and test it out to see. I suppose there's some slim possibility it really is faster in all areas, i just find it hard to believe at the moment. There's been no die-shrink here. It's the same process used since the 600 series. Maxwell Gen 2 will have the die-shrink to 20nm and i'd be much more likely to believe these numbers then, but without the die-shrink claiming to have 35% more performance per core at half the TDP is a huge leap for a single generation. After the 20nm Maxwells start showing up, how much of an additional performance increase over Gen 1 are we looking at? lol it's just too good to be true.

[John P. Myers]

Here's a comparison - GK107 (Kepler) vs. GM107 (Maxwell)

The GM107 die is 26% larger, has 256 more cores, has 570 million more transistors, 10% more bandwidth, is rated at 494 *more* GFLOPS FP32 (a whopping 61% increase), but uses 6.25% less power to do all this. Oh and 400MHz faster VRAM too (an 8% increase).

So Nvidia is telling me i can crunch 61% faster using slightly less electricity in just 1 generation? I call BS

BTW these GPUs are launching in 3 days.

[DrPop]

...
So Nvidia is telling me i can crunch 61% faster using slightly less electricity in just 1 generation? I call BS

BTW these GPUs are launching in 3 days.

Yeah, maybe it is too good to be true, but man that would be nice.

[conf]

Not very good for crunching. In Europe the 750ti costs the same as the "old" 660GTX,
but the second one has 150% power related to the new maxwell.
For energy purposes the new one is better but both have low power consumption.

plus a question : if I have two ATIs running how do I add a Nvidia in third PCIe port for different projects like Einstein or GPUGrid,
tried this often before but it never worked as I had expected. Used a cc_config for these trials.

[John P. Myers]

plus a question : if I have two ATIs running how do I add a Nvidia in third PCIe port for different projects like Einstein or GPUGrid,
tried this often before but it never worked as I had expected. Used a cc_config for these trials.

No cc_config needed. Just install the Nvidia GPU, install drivers, then select the project you want to run and tell it to only use the Nvidia GPU, then you won't get AMD/ATI WUs from that project. The projects you want to run on AMD/ATI, don't allow them to use Nvidia.



After 3+ hours of digging and digging for real numbers and more info, i found 2 things important to note

First is that manufacturers (most likely Zotac and Gigabyte) have stated they will be making true single-slot versions of not only the GTX 750, but of the 750Ti as well. Game on! You can assume from this there will also be a LP version of at least the GTX 750 for Al

Second is it *does* appear compute has increased a good bit. I saw FAH Bench results that we proming (Folding at Home's benchmark). I also saw benchmarks in regards to alt-coin mining. Not what we're doing, but math is math. The results were very promising in both cases. When the tru single-slot versions come out, i will buy one that day and get some testing done.

Also thirdly, Nvidia seems to have underclocked these cards big time. People are reporting stable results at 1300MHz with a tiny 31mV bump. EVGA has a version of the 750 with a 1294MHz Boost out of the box (stock is 1085MHz). That's pretty significant.

[STE\/E]

I always thought the amount of core's was the most important for crunching, they don't appear to have that many cores for BOINC Crunching ...

GTX 750Ti:
640 cores
60W TDP
1020MHz base/1085MHz Boost

GTX 750:
512 cores
55W TDP
1020MHz base/1085MHz Boost

[John P. Myers]

I always thought the amount of core's was the most important for crunching, they don't appear to have that many cores for BOINC Crunching ...

No they don't have many cores as this is a lower-midrange GPU. There are a few reasons i'm really interested in this though. It's the first chip based on the Maxwell architecture and since it's not a high-end GPU, the performance of the 750Ti will actually dictate how the rest of the Maxwell GPUs (the GTX 800 series) perform. Efficiency always increases as you move to the larger more expensive GPUs and with the 750Ti's GFLOPS/W rating starting out higher than even a GTX 780Ti, it follows that the rest of the 800 series would completely obliterate anything else ever made.

Another thing about the GPUs is you could buy 8 of the GTX 750Ti's for about the price of a Titan but compute 2.32 times faster. Another reason i like looking at these lower-end cards is the lower power draw allows manufacturers to make them only 1 slot wide, and i like finding GPUs i can stick in empty slots just as fillers. Currently this is the fastest, most power efficient GPU that is only 1 slot wide: http://www.newegg.com/Product/Produc...82E16814125468 I was first to review it If the 750 and 750Ti are as good as Nvidia says they are, they will become my new #1 choice.

[Al]

So, Low Profile, which means it will fit in my servers. Looks like a decent upgrade to the 430s that are in there now.