1. Welcome to TechPowerUp Forums, Guest! Please check out our forum guidelines for info related to our community.

NVIDIA GeForce GTX 750 Ti (Maxwell) Review

Discussion in 'Graphics Cards' started by DGLee, Feb 18, 2014.

  1. DGLee

    Joined:
    Jun 9, 2011
    Messages:
    32 (0.02/day)
    Thanks Received:
    111
    Location:
    South Korea
    Today, NVIDIA introduced its next-generation graphics architecture "Maxwell", first as a replacement of entry-level product of current generation, GK107 based on Kepler architecture, namely GM107. (M in its codename refers Maxwell as well as K in GK107 means Kepler) Recently (for more than a couple of years) NVIDIA has launched its new architecture every even year, for example, Tesla for 2008, Fermi for 2010 and Kepler for 2012, and this biennial rule made it more predictable that the company will release post-Kepler product sooner or later. Many enthusiasts have been questioning how fast the new flagship GPU compared to GK110. This question may, however, not be answered today. Since today we only got a baby Maxwell.

    In fact, release of GM107 is not that typical though the biennial rule is held, since on the other hand the company's product differentiation strategy goes somewhat differently than previous generations : GM107 itself is the symbol of NVIDIA's first bottom-up approach within a generation thank to delay of 20nm fabrication process by TSMC. In 2008, NVIDIA launched GTX 260/280 based on GT200 GPU, the biggest silicon ever made upon Tesla architecture, as the first Tesla chip at the same time. Smaller variations such as GT215/GT218 appeared a year later than GT200. In Fermi era, the same footage is repeated : GF100 -> GF104 -> GF106/108. These all can be categorized as top-down approaches. What made this kind of approach actually works is evolution of fabrication process. 55nm process was matured when Tesla was introduced as well as 40nm was matured when Fermi came, so was 28nm for Kepler. For Maxwell, however, there was no matured 20nm fabrication process and thus making a big chip based on Maxwell is virtually impossible as of today. This is why NVIDIA changed its differentiation strategy from top-down to bottom-up.

    [​IMG]

    As a first Maxwell variant, GM107 features followings :
    - Different core config : 192 CUDA / 16 TMU per SMX -> 128 / 8 per Maxwell SM (SMM)
    - Higher energy & performance efficiency, mainly derived from smaller SM design
    - Less susceptible to graphics memory bandwidth thank to enlarged L2 cache : 256KB -> 2048KB

    Many of above are highly anticipated that we can see them in bigger Maxwell as well. For short, Maxwell features "(relatively) smaller die size, higher efficiency" in compared to Kepler. Well, let's taste these out by testing our very first GM107 SKU, GeForce GTX 750 Ti.

    Test system I used consists as follow.

    [​IMG]

    I ran all games 3 times repeatedly for each resolution and choose the median as the represent value for the game under each resolution. Graphics quality of each game was the highest possible in-game settings except anti-aliasing. Further details will be noted in test result chapters.

    Before we go for tests, I want to give you some theoretical backgrounds. Let's see this table.

    [​IMG]

    GM107 seems lie between GK106 and GK107 in terms of its backend units : CUDA core count is more like GK106 than GK107 while the contrary is more likely for anything other than CUDA - TMU count, ROP count and memory interface of GM107 is more like GK107 than GK106. Based on that, GM107 is expected to perform 'a bit' faster than GK107 and 'a double bit' slower than GK106 if there is no architectural change in sub-SM unit level. In other words, If actual GM107 (GTX 750 Ti) would performs better than provided below, it means there are some primitive improvements in Maxwell architecture. See below.

    [​IMG]

    The table is derived by VGA calculator* (* = see this : http://iyd.kr/612) I designed. By assuming that all units among SKUs are vertically equivalent(that is, 1 CUDA core of GTX 650 Ti is same as 1 CUDA core of GTX 660, 1 TMU of GTX 650 Ti is same as 1 TMU of GTX 660, and so on), 'conceptual' GTX 750 Ti will have overall performance just one-digit % faster than GTX 650 Ti and about 30% slower than GTX 650 Ti Boost. Let's keep that in mind and compare the virtual results derived in Performance Summary chapter. Then we know whether Maxwell can be distinguishable from Kepler or not.

    (※ Detailed results are skipped for image # limit. Full review is on my personal blog : http://iyd.kr/624)

    Here are performance summary graphs:

    [​IMG]
    [​IMG]
    [​IMG]

    As a whole, GTX 750 Ti is about a quarter % faster than GTX 650 Ti, almost twice than GTX 650 and 6% slower than GTX 650 Ti Boost. In comparison with Radeon counterparts, GTX 750 Ti lies almost at the middle between Radeon R7 260X and 7850. When it comes to a matter of actually "buy" a graphics card, one we must consider is performance-per-price. GTX 750 Ti is announced as $150, a price that almost exactly represent its relative performance among current generation competitors - Nothing blazingly good, nothing bad.

    So far we have covered how fast (or slow) GTX 750 Ti is. Do you think it's an attractive newbie on graphics card market? To answer this question, we must distinguish the GM107 itself and GM107-based SKU, GeForce GTX 750 Ti. Let's first discuss about the latter.

    Despite the fact that nomenclature of GTX 750 Ti resembles that of GTX 650 Ti and GTX 750 Ti is significantly faster (+25%) than its predecessor, GTX 750 Ti itself is not an eye-popping attractive product in terms of performance. It just lies between GTX 650 Ti and GTX 650 Ti Boost, both of them are currently tangible & affordable. In other words, GTX 750 Ti throws nothing special in that price-performance niche in graphics card market : existing-level performance for existing-level perf/price ratio. Does is make sense?

    When we talk about the former, however, the tone must be changed. As the first tangible silicon based on Maxwell architecture, GM107 features 66% more CUDA cores than GK107, 25% more TMUs and same amount of ROPs as well as same memory interface. Though we cannot compare those units apple-to-apple match, it is possible to be said that GM107 is almost twice faster than GK107 despite the fact that it only earns +0~66% of units in their own kind. Increment on performance definitely excess the increment on backend hardwares. In other words, NVIDIA succeeded to improve unit-level efficiency in Maxwell architecture. In addition to that, Maxwell's silicon efficiency is also impressive. Die area of GM107 is only 148 mm2 (square milimeter) while GK106 is 221 mm2 and GK107 is 118 mm2. Remind that GM107 is as twice faster than GK107 while die area is only bigger by 20% and reaches over 80% of performance of GK106 while GM107 occupies 60% of silicon of GK106.

    Just Imagine : "Maxwell's efficiency done within & fully utilized the size of current big chip."

    Well, the article is finally over.
    Thank you all for reading my words. Have a nice day!
     
    Mathragh and FX-GMC say thanks.
  2. ZetZet

    ZetZet

    Joined:
    Feb 27, 2013
    Messages:
    190 (0.15/day)
    Thanks Received:
    27
    Location:
    Lithuania
    Wow that's just terrible. Worse than 650ti boost.
     
  3. raschmidt

    raschmidt

    Joined:
    Jun 9, 2012
    Messages:
    28 (0.02/day)
    Thanks Received:
    5
    Location:
    Chengdu, China
    6% worse than a 650 ti boost is not that bad, IIRC that was close to $200 when it came out. It's a brand new chip and it's bottom of the bucket. What I think is more surprising is the gap between the GTX 760 and 750 ti. The 650 ti and 660 wasn't that big. I bought one of each when they first came out and while I could tell the 660 was better, I still own the 650 ti and I don't regret it. NVidia has almost a $100 gap between the 750 ti and 760, so does that mean we get a 750 ti Boost at around $180-$190 this summer?
     
  4. DarkOCean

    DarkOCean

    Joined:
    Jan 28, 2009
    Messages:
    1,719 (0.62/day)
    Thanks Received:
    443
    Location:
    on top of that big mountain on mars(Romania)
    what did you expected, miracles? it's just a low end gpu afterall.
     
  5. ZetZet

    ZetZet

    Joined:
    Feb 27, 2013
    Messages:
    190 (0.15/day)
    Thanks Received:
    27
    Location:
    Lithuania
    750ti should be closer to 760.
    And the price is close to 660 atleast in my country.
     
  6. DarkOCean

    DarkOCean

    Joined:
    Jan 28, 2009
    Messages:
    1,719 (0.62/day)
    Thanks Received:
    443
    Location:
    on top of that big mountain on mars(Romania)
    judging by the name and price yes, but looking at the die size and specs it cannot be on 28nm.
     
  7. raschmidt

    raschmidt

    Joined:
    Jun 9, 2012
    Messages:
    28 (0.02/day)
    Thanks Received:
    5
    Location:
    Chengdu, China
    In the US it's a pretty big gap between the 750 ti and 660. I guess they'll keep the 660 around for awhile, it fills the price/performance gap between the 750 ti and 760. Kinda surprised they're still selling the 650 ti and 650 though, the price hasn't gone down on those much and they're well over a year old now.
     
  8. ZetZet

    ZetZet

    Joined:
    Feb 27, 2013
    Messages:
    190 (0.15/day)
    Thanks Received:
    27
    Location:
    Lithuania
    What do you mean the price is not close http://pcpartpicker.com/parts/video-card/#sort=a7&c=114
    http://www.amazon.com/dp/B00I9I8NFE/?tag=tec06d-20
     
  9. raschmidt

    raschmidt

    Joined:
    Jun 9, 2012
    Messages:
    28 (0.02/day)
    Thanks Received:
    5
    Location:
    Chengdu, China
    I think we might need to wait on the final performance too......I was checking out EVGA's website and they have 4 models of each. Two with the reference cooler and two with the ACX cooler, they're suggesting the ones with the ACX coolers (and I guess 6 pin connectors) are going to have a pretty good power boost.
     
  10. raschmidt

    raschmidt

    Joined:
    Jun 9, 2012
    Messages:
    28 (0.02/day)
    Thanks Received:
    5
    Location:
    Chengdu, China
  11. ZetZet

    ZetZet

    Joined:
    Feb 27, 2013
    Messages:
    190 (0.15/day)
    Thanks Received:
    27
    Location:
    Lithuania
  12. raschmidt

    raschmidt

    Joined:
    Jun 9, 2012
    Messages:
    28 (0.02/day)
    Thanks Received:
    5
    Location:
    Chengdu, China
    So? It was the same with 550 ti and 650 ti, NVidia does better with higher end. These chips are going to cause AMD some serious problems on mobile. The performance per watt is way better than anything that AMD has and NVidia clearly said Maxwell would be built as a mobile chip first. Gives me high hopes for what the Tegra's will have in a year or two in terms of graphics capabilities.
     
  13. Steevo

    Steevo

    Joined:
    Nov 4, 2005
    Messages:
    9,275 (2.35/day)
    Thanks Received:
    1,788
    Nvidia is taking a hard look at die space and efficiency, heat issues aside AMD kicked their arse in performance per mm of die size, and they are trying to reach and exceed the same point with a new process so are starting out with these low end chips. AMD has done it before and it worked well, helping them learn how to use gates and vias to stabilize core voltage under load. All that translates to higher clocks and less leakage at temperatures they will have to endure to attain the performance they need.

    It is impressive this performance comes out of this low power die.
     
    10 Year Member at TPU 10 Million points folded for TPU
  14. eidairaman1

    eidairaman1

    Joined:
    Jul 2, 2007
    Messages:
    14,721 (4.40/day)
    Thanks Received:
    2,257
    Yeah but for a 750 to perform worse than a 650? Why bother then less u need a card for htpc/youtube/videocalls,igp died
     
  15. raschmidt

    raschmidt

    Joined:
    Jun 9, 2012
    Messages:
    28 (0.02/day)
    Thanks Received:
    5
    Location:
    Chengdu, China
    750 way out-performs the 650.

    Mentioned the EVGA 750 ti FTW earlier since EVGA was boosting power, looks like it gives it about 5fps higher than reference 750 ti in most of the games he tested. Looks like it's pretty solidly ahead of the 650 ti Boost - EVGA website says $170 for the FTW and $160 for the 750 ti ACX (which I'm guessing is pretty close).

    link to Paul's Hardware review on youtube
     
  16. 15th Warlock

    15th Warlock

    Joined:
    Aug 16, 2004
    Messages:
    3,002 (0.68/day)
    Thanks Received:
    1,580
    Location:
    Visalia, CA
    750 does not perform lower than 650, where did you read that?:confused:

    Not only that but 750 is almost twice as power efficient as 650, this is an amazing achievement considering both cards are equal in terms of functionality and fabrication node (28nm)

    Me, I can't wait to see what Maxwell will do at 20nm, think of 125W flagship cards, 250W dual GPU cards and 15W or lower power decent performing tablet GPUs, the potential to harness such performance at such low power requirements boggles my mind ;)
     
    10 Year Member at TPU
  17. Steevo

    Steevo

    Joined:
    Nov 4, 2005
    Messages:
    9,275 (2.35/day)
    Thanks Received:
    1,788
    It performs better than the 650, but not the Ti Boost, and they should be coming with TI Boost editions of this card too I am sure, plus I have always hated Nvidias naming scheme, and again, it is their first real spin of this I am betting and what they are learning now will be the difference between a new monster card and a hugely hot monster card with clock issues.
     
    10 Year Member at TPU 10 Million points folded for TPU
  18. Slizzo

    Slizzo

    Joined:
    Aug 2, 2011
    Messages:
    632 (0.34/day)
    Thanks Received:
    223
    Look again, the 750Ti consistently outperforms the 650Ti, but not the 650Ti Boost.

    For what it is (Maxwell on 28nm) it's pretty damned good.
     
  19. Steevo

    Steevo

    Joined:
    Nov 4, 2005
    Messages:
    9,275 (2.35/day)
    Thanks Received:
    1,788
    10 Year Member at TPU 10 Million points folded for TPU
  20. Slizzo

    Slizzo

    Joined:
    Aug 2, 2011
    Messages:
    632 (0.34/day)
    Thanks Received:
    223
    So Mr. Semantics, why are you glossing over the 650Ti then? For what it is, the 750Ti is performing admirably well despite being "down on power" compared to previous generation designs?
     
    Steevo says thanks.
  21. james888

    james888

    Joined:
    Jun 27, 2011
    Messages:
    5,522 (2.93/day)
    Thanks Received:
    2,892
    125w flagship. Come on. Pls give me a 200w epic flagship.
     
    15th Warlock says thanks.
    Crunching for Team TPU
  22. xorbe

    xorbe

    Joined:
    Feb 14, 2012
    Messages:
    1,351 (0.82/day)
    Thanks Received:
    436
    Location:
    Bay Area, CA
    Talk about "violent agreement"!
     
  23. erocker

    erocker Super Moderator Staff Member

    Joined:
    Jul 19, 2006
    Messages:
    41,436 (11.23/day)
    Thanks Received:
    16,763
    Price makes me not consider it. Close, but no.
     
    10 Year Member at TPU
  24. CJCerny

    CJCerny

    Joined:
    Nov 25, 2008
    Messages:
    917 (0.32/day)
    Thanks Received:
    277
    Location:
    Akron, OH
    I suspect that Nvidia will sell a boatload of this GPU. Imagine all the Dell, Lenovo, Asus, HP, etc. desktops out there with an open PCI-E 16x slot and no extra power connectors that can be turned into low end gaming PCs by adding this GPU.
     
  25. Fluffmeister

    Fluffmeister

    Joined:
    Dec 22, 2011
    Messages:
    1,424 (0.83/day)
    Thanks Received:
    675
    Definitely bodes well for future Maxwell releases, great efficiency improvements and you can't even credit a node change this time.
     

Currently Active Users Viewing This Thread: 1 (0 members and 1 guest)