These days, gamers like their graphics cards beefy. Double-slot coolers and fancy fan shrouds are typically what elicit Tim Allen-style grunts and knowing nods of approval. After all, high frame rates require complex GPUs. Billions of transistors cranking away at Battlefield 4 get hot. And all of that heat needs to go somewhere.
So if you're coming to the table with a short, naked PCB, it'd better have a trick or two up its figurative sleeve.
Yet, Nvidia, perhaps trying to prove a point, shipped out its reference GeForce GTX 750 Ti on a less than six-inch board. With no auxiliary power connector. Sporting a little bolted-on orb-style heat sink and fan. It's pretty much the same size as GeForce GTX 650 Ti. But without the big cooler, GTX 750 Ti is daintier than even a lot of sound cards we've tested.

Nevertheless, Nvidia claims that its first Maxwell architecture-based product targets gaming at 1920x1080 in the latest titles using some pretty demanding settings. Could this be the graphics world’s Prius?
Maxwell In The Middle
Maxwell’s story is intriguing, partly because of what it means to the company’s design approach moving forward, but also because Nvidia is keeping more architectural details to itself than usual. Let’s start with the design.

Back in December of last year, we were in Santa Clara learning about Nvidia’s Tegra K1 SoC. We already knew that K1’s graphics engine was Kepler-based, essentially a single SMX with notable changes to the structures connecting various subsystems in a bid to optimize for power. But Jonah Alben, senior vice president of GPU engineering, also made it clear that every new architecture, from Maxwell onward, would be built with mobile in mind. Engineers would optimize the fabrics between GPU components based on performance targets and power budgets. However, the fundamental building blocks would stay common between segments, and efficiency would guide the important decisions.
Tegra: Whence It Came
The impetus for Maxwell comes from Nvidia's effort in the smartphone and tablet space. Read Nvidia Tegra K1 In-Depth: The Power Of An Xbox In A Mobile SoC? to learn more about that architecture.
This is clearly good news for the Tegra family, which continues clawing around for a more meaningful slice of market share. K1-based devices aren't even here yet and we're already thinking about Nvidia's claim that Maxwell offers two times the performance-per-watt of Kepler, and what such sizable improvements could mean to mobile gaming.
A renewed emphasis on efficiency should be good on the desktop too though, providing the company's retooled architecture continues scaling up from single- to double- and triple-digit power ceilings.
Fortunately, you won't have to wait long for an answer. The GeForce GTX 750 Ti launching today should demonstrate what Maxwell can do (at least at a 60 W TDP). Nvidia says its more effective design pulls power consumption way down and nudges performance up, even in a GPU featuring fewer CUDA cores. Knowing that it wouldn’t have a new process technology node to lean on, Nvidia had to make its improvements to Maxwell with 28 nm manufacturing in mind. In other words, it needed to gets its GPUs working smarter, since simply tacking on more resources wouldn’t be an option.
The Maxwell Streaming Multiprocessor
Company representatives tell us that Maxwell’s biggest gains come from a redesign of the Streaming Multiprocessor, now abbreviated as SMM.
In Kepler, each SMX plays host to 192 CUDA cores, four warp schedulers, and a 256 KB register file. There’s also 64 KB serving as shared memory and L1 cache, a separate texture cache, and a uniform cache, plus 16 texture units. The big jump in CUDA core count and control logic helped Nvidia overcome losing Fermi’s doubled shader frequency. But the SMX apparently proved difficult to fully utilize in this configuration.
Maxwell attempts to address that by partitioning the SMX into four blocks, each with its own instruction buffer, warp scheduler, and pair of dispatch units. Kepler’s 256 KB register file now gets split into four 64 KB slices. And the blocks have 32 CUDA cores each, totaling 128 across the SMM (down from Kepler’s 192). The previous architecture’s 32 load/store and 32 special function units carry over to Maxwell. However, double-precision math is further pared back to 1/32 the rate of FP32; that was 1/24 in the mainstream Kepler-based GPUs.
| GM107 SMM (Left) Versus GK106 SMX (Right) | |||
![]() | |||
| Per SM: | GM107 | GK106 | Ratio |
| CUDA Cores | 128 | 192 | 2/3x |
| Special Function Units | 32 | 32 | 1x |
| Load/Store | 32 | 32 | 1x |
| Texture Units | 8 | 16 | 1/2x |
| Warp Schedulers | 4 | 4 | 1x |
| Geometry Engines | 1 | 1 | 1x |
Every pair of blocks is tied to a 12 KB texture and L1 cache, adding up to 24 KB per SMM. Block pairs are also associated with four texture units, meaning SMMs come armed with eight. That’s half as many texture units compared to Kepler’s SMX. And the table above makes it look like GM107 actually gives up some ground to GK106. But don’t freak out about bottlenecks quite yet. Remember, the architecture is supposed to get more done using less resources.
Lastly, there’s a 64 KB shared memory space for the SMM, which carries over from Fermi and then Kepler, but is no longer called out as L1 cache for compute tasks. It used to be that this space could be configured as 48 KB of shared space and 16 KB of L1 and vice versa. Now that's not necessary, so all 64 KB is used as a shared address space for GPU compute.
As you might imagine, cutting 64 CUDA cores and eight texture units from the SMM means that each building block consumes significantly less die size. Meanwhile, Nvidia claims that it’s able to hold onto ~90% of the multiprocessor’s performance by keeping cores busy in a sustained way. If you’re contemplating what that might mean to a tablet, you’re not alone. But in a desktop application, Nvidia’s simply able to pack more SMMs into a set amount of space. The GeForce GTX 650 Ti this card replaces employed four SMX blocks, while GeForce GTX 750 Ti incorporates five SMMs.
Constructing GM107
This is the first time we’ve seen Nvidia introduce a new architecture on a decidedly mid-range graphics card. With Fermi, it was the full-force GF100. Even the Kepler-based GK104 was an impressively-fast way to meet that architecture. So, the messaging is quite a bit different with GM107 leading the charge. Of course, that’s because GeForce GTX 750 Ti has to slot into a portfolio still dominated by Kepler, rather than simply ascending a throne.
And so it does so using a fully-enabled implementation of GM107, composed of five SMMs in a single Graphics Processing Cluster with its own Raster Engine. GM107 can set up one visible primitive per clock cycle, which is just behind GK106's primitive rate of 1.25 prim/clock and double GK107's .5 prim/clock.

As in Nvidia architectures prior, ROP partitions and L2 cache slices are aligned. And like the GeForce GTX 650 Ti’s GK106 processor, GM107 sports two partitions with eight units each, giving you up to 16 32-bit integer pixels per clock. Where the two GPUs really diverge is their L2 cache capacity. In GK106, you were looking at 128 KB per slice, adding up to 256 KB in an implementation with two ROP partitions. GM107 appears to wield 1 MB per slice, yielding 2 MB of memory used for servicing load, store, and texture requests. According to Nvidia, this translates to a significant load shifted away from the external memory system, along with notable power savings.
Going easy on memory bandwidth is smart, since GM107 exposes a pair of 64-bit memory controllers to which 1 or 2 GB of 1350 MHz GDDR5 DRAM is attached. Peak throughput is, interestingly, exactly what we got from GeForce GTX 650 Ti: 86.4 GB/s. The memory is feeding fewer CUDA cores, but they’re managed more efficiently. So, the big L2 is supposed to play an instrumental role in preventing a bottleneck.

Indeed, a look at global in-page random cache latencies helps illustrate how Maxwell's memory hierarchy keeps the GPU busy more consistently.
Beyond the pieces of GM107 devoted to gaming and compute tasks, Nvidia also says it improved the fixed-function NVEnc block. That’s the bit of logic responsible for letting ShadowPlay encode your frag fest with minimal performance impact. It’s what enables streaming to the Shield. And it accelerates a few transcoding apps to get big movies onto your portable devices quickly. Whereas Kepler was capable of encoding H.264-based content ~4x faster than real-time, Maxwell purportedly achieves 6-8x real-time. H.264 decode performance is said to be eight to 10 times quicker than it was before, also. Nvidia achieves those gains, it says, by simply speeding up the fixed-function blocks.
| GeForce GTX 650 | GeForce GTX 650 Ti | GeForce GTX 750 Ti | GeForce GTX 660 | |
|---|---|---|---|---|
| GPU | GK107 | GK106 | GM107 | GK106 |
| Architecture | Kepler | Kepler | Maxwell | Kepler |
| SMs | 2 | 4 | 5 | 5 |
| GPCs | 1 | 2 | 1 | 3 |
| Shader Cores | 384 | 768 | 640 | 960 |
| Texture Units | 32 | 64 | 40 | 80 |
| ROP Units | 16 | 16 | 16 | 24 |
| Process Node | 28 nm | 28 nm | 28 nm | 28 nm |
| Core/Boost Clock | 1058 MHz | 925 MHz | 1020 /1085 MHz | 980 / 1033 MHz |
| Memory Clock | 1250 MHz | 1350 MHz | 1350 MHz | 1502 MHz |
| Memory Bus | 128-bit | 128-bit | 128-bit | 192-bit |
| Memory Bandwidth | 80 GB/s | 86.4 GB/s | 86.4 GB/s | 144.2 GB/s |
| Graphics RAM (GDDR5) | 1 or 2 GB | 1 or 2 GB | 1 or 2 GB | 2 GB |
| Power Connectors | 1 x 6-pin | 1 x 6-pin | None | 1 x 6-pin |
| Maximum TDP | 64 W | 110 W | 60 W | 140 W |
| Price | $130 (2 GB) | $150 (2 GB) | $150 (2 GB) | $190 (2 GB) |
All told, the GM107 GPU ends up with 1.87 billion transistors in a 148 mm² die. If you keep the comparison to GeForce GTX 650 Ti going, then the first Maxwell-based processor replaces GK106, which packs 2.54 billion transistors into a 221 mm² die. Before we get to our performance results, we have to assume Nvidia’s emphasis on efficiency is significant enough to let the company use less transistors on a smaller die, cut out a lot of CUDA cores and texture units, and still improve performance overall. At least, that’s what we’ll be looking for…
Alternatively, you can put GM107 up against the 1.3 billion-transistor, 118 mm² GK107, if your preference is a face-off of thermal ceilings. In that case, the Maxwell-based processor is more complex, larger, significantly faster, and yet it should still use less power.
If you’re building a graphics card, dropping a low-power GPU onto it gives you some options. You don’t get shoehorned into a dual-slot, actively-cooled beast of a board. For its reference design, Nvidia chose a PCB less than six inches long—it pretty much ends with the PCI Express slot connector. There’s a single-slot I/O bracket with two dual-link DVI connectors and a mini-HDMI output. However, Nvidia covers the GM107 processor with an orb-style heat sink and fan that eat up two slots worth of space, so you still have to budget accordingly.

The card’s 60 W ceiling is easily satisfied by a 16-lane PCI Express slot, which is rated for up to 75 W. That means you won’t find an auxiliary power connector on the PCB (even if there are holes for one). We’ve long been fans of cards fitting this profile because of how flexible they are. Previously, AMD’s Radeon HD 7750 was your best bet for upgrading an old decrepit box with too-little power output or not enough connectors for a decent add-in board. Now the GeForce GTX 750 Ti is gunning for that position.
Nvidia's GK106-based GeForce GTX 660 is the card we're recommending around the $200 mark; AMD's Radeon R9 270 is just too pricey in comparison. This board from Gigabyte, overclocked from the factory, is a favorite. Read the Full Review
Gigabyte GeForce GTX 660 OC 2GB
Upper-Mainstream Performance
Unfortunately, there’s also no SLI bridge connector. True to Nvidia’s mainstream approach, the $150 price point is right about where you lose the option to sling two boards together for higher performance. This is a competitive disadvantage; AMD’s alternatives in the same price range allow CrossFire configurations. It’s probable that Nvidia could achieve SLI over PCI Express, but the company says it doesn’t see much demand from enthusiasts looking to link $150 cards. If you feel differently, speak up. We’d be curious to see if a couple of GM107s could beat a GeForce GTX 770, for sure.

Nvidia plans to offer two versions of the GeForce GTX 750 Ti—one with 1 GB of GDDR5, priced at $140, and available later in February, and a 2 GB model that should be selling for $150 by the time you read this. Moreover, there will be a GeForce GTX 750 that ships later in the month at a price point of $120.
The initial round of partner boards includes a mix of cards eating up one and two expansion brackets, but they’re all dual-slot designs. Down the road, though, we’re told to expect dual-slot passively-cooled solutions. Single-slot configurations are also possible, though nobody seems certain that a low-profile fan can yield a pleasant experience.
MSI GTX 750 Ti Gaming OC
MSI´s GTX 750 Ti Gaming follows Nvidia's reference design, forgoing an auxiliary power connector. But it does include an oversized cooler, which probably could have cooled the 60 W GM107 GPU passively.
Even still, MSI's configuration doesn't have to get very loud to do its job well; the card never gets hotter than 51-52 degrees Celsius, and fan speeds top out between 32 and 33 percent. That also means noise under load is barely louder than at idle. In fact, our acoustic measurements don't even appear plausible because the differences fall within the tolerances of our hardware.
Dimensions, Weight, And Connectors
The dimensions we measure almost never match the figures you get from manufacturer specifications, which is why we present our own data using the distances shown in the following image.
| Auxiliary Power Connector | None | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| Connectors | 1 x DVI-D (Dual-link) 1 x HDMI 1 x D-Sub (analog) | ||||||||
| Form Factor | Dual-slot | ||||||||
| Pros | + Very cool + Very quiet | ||||||||
| Cons | - Relatively tall | ||||||||
| Measurements |
| ||||||||
| Weight | 533 g |
As usual, GPU-Z provides us an overview of the card's specifications:
Front and Back View
Side Views
Gigabyte GTX 750 Ti Windforce OC
In addition to the MSI card, we also received Gigabyte's GTX 750 Ti Windforce OC just before publication. The company similarly leans on Nvidia's reference PCB. However, it also exploits the solder points for a six-pin auxiliary power connector, yielding potentially better overclocking headroom.
Dimensions, Weight, And Connectors
The dimensions we measure almost never match the figures you get from manufacturer specifications, which is why we present our own data using the distances shown in the following image.
| Auxiliary Power Connector | 1 x Six-pin | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| Connectors | 1 x DVI-D (Dual-link) 1 x DVI-I 2 x HDMI | ||||||||
| Form Factor | Dual-slot | ||||||||
| Pros | + Cool + Quiet + Compact | ||||||||
| Cons | - Relatively tall (one heat pipe) | ||||||||
| Measurements |
| ||||||||
| Weight | 418 g |
As usual, GPU-Z provides us an overview of the card's specifications:
Front and Back View
Side Views
Zotac GTX 750 Ti OC
Zotac's GTX 750 Ti goes the same route as MSI's, forgoing the six-pin power connector in favor of a design driven exclusively by a 16-lane PCI Express slot. It is built on the same reference PCB though, and is the only model employing a single-fan cooler.
Dimensions, Weight, And Connectors
The dimensions we measure almost never match the figures you get from manufacturer specifications, which is why we present our own data using the distances shown in the following image.
| Auxiliary Power Connector | None | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| Connectors | 1 x DVI-D (Dual-link) 1 x DVI-I 1 x Mini-HDMI | ||||||||
| Form Factor | Dual-slot | ||||||||
| Pros | + Short and compact | ||||||||
| Cons | - A little louder and warmer than dual-slot cards | ||||||||
| Measurements |
| ||||||||
| Weight | 304 g |
As usual, GPU-Z provides us an overview of the card's specifications:

Front and Back View
Side Views
We tested all of the cards in today's story with each company's latest beta drivers: Catalyst 14.1 beta 6 from AMD and 334.67 beta from Nvidia, however the new GeForce GTX 750 Ti required a special 334.69 beta driver. We used medium- to high-detail settings at 1920x1080 to give the GeForce GTX 750 Ti and its competition a realistic workload, which should strike a good balance between image quality and performance for this class of card.
| Test System | |||||
|---|---|---|---|---|---|
| CPU | Intel Core i5-2550K (Sandy Bridge), Overclocked to 4.2 GHz @ 1.3 V | ||||
| Motherboard | Asus P8Z77-V LX. LGA 1155, Chipset: Intel Z77M | ||||
| Networking | On-Board Gigabit LAN controller | ||||
| Memory | Corsair Performance Memory, 4 x 4 GB, 1866 MT/s, CL 9-9-9-24-1T | ||||
| Graphics | Sapphire Radeon R7 260X 1100 MHz GPU, 2 GB GDDR5 at 1625 MHz (6500 MT/s) XFX Radeon HD 7850 860 MHz GPU, 1 GB GDDR5 at 1200 MHz (4800 MT/s) Sapphire Radeon R7 265 925 MHz GPU, 2 GB GDDR5 at 1400 MHz (5600 MT/s) Reference Radeon R9 270 925 MHz GPU, 2 GB GDDR5 at 1400 MHz (5600 MT/s) Reference GeForce GTX 650 Ti 925 MHz GPU, 1 GB DDR3 at 1350 MHz (5400 MT/s) Reference GeForce GTX 650 Ti Boost 980/1033 MHz GPU, 2 GB GDDR5 at 1502 MHz (6008 MT/s) Reference GeForce GTX 660 980/1033 MHz GPU, 2 GB GDDR5 at 1502 MHz (6008 MT/s) Reference GeForce GTX 750 Ti 1020/1085 MHz GPU, 2 GB GDDR5 at 1350 MHz (5400 MT/s) | ||||
| Hard Drive | Samsung 840 Pro, 256 GB SSD, SATA 6Gb/s | ||||
| Power | XFX PRO850W, ATX12V, EPS12V | ||||
| Software and Drivers | |||||
| Operating System | Microsoft Windows 8 Pro x64 | ||||
| DirectX | DirectX 11 | ||||
| Graphics Drivers | AMD Catalyst 14.1 Beta 6, Nvidia GeForce 334.67 Beta Nvidia GeForce 334.69 Beta for GeForce GTX 750 Ti | ||||

We've almost completely eliminated mechanical storage in the lab, and instead lean on solid-state drives to alleviate I/O-related bottlenecks. Samsung sent all of our offices 256 GB 840 Pros, so we standardize on these exceptional SSDs.

Naturally, discrete graphics cards require a substantial amount of stable power, so XFX sent along its PRO850W 80 PLUS Bronze-certified power supply. This modular PSU employs a single +12 V rail rated for 70 A. XFX claims that this unit provides 850 W of continuous power (not peak) at 50 degrees Celsius (a higher temperature than you'll find inside most enclosures).
| Benchmark Configuration | |
|---|---|
| 3D Games | |
| Metro: Last Light | Version 1.0.0.14, Built-in Benchmark |
| Grid 2 | Version 1.8.85.8679, Built-in Benchmark Scene D6 |
| Assassin's Creed IV: Black Flag | Version 1.05, Custom THG Benchmark, 40-Sec |
| Battlefield 4 | Version 1.0.0.1, Custom THG Benchmark, 90-Sec |
| BioShock Infinite | Version 1.1.24.21018, Built-in Benchmark |
| Far Cry 3 | Version 1.05, Custom THG Benchmark, 55-Sec |
| Arma 3 | Version 1.10.114.700, Custom THG Benchmark, 30-Sec |
We begin with Arma 3, a game that plays more like a military simulation than a first-person shooter. Regardless of the title's realism (or perhaps because of it), it tends to tax modern graphics hardware with cutting-edge features.


The GeForce GTX 750 Ti is significantly faster than AMD's Radeon R7 260X, but slower than the GeForce GTX 650 Ti Boost and Radeon R7 265. Nvidia's first Maxwell-based card fares well using the High detail setting with 4x MSAA enabled at 1920x1080. Its frame rate stays above 36 FPS through our benchmark.


We encounter moderate frame time variance spikes from a number of different cards in Arma 3, though they don't have a detrimental impact on the game's playability. Curiously, the Radeon R7 265 demonstrates higher variance than the competition, despite its strong frame rate. This finding was confirmed across multiple runs.
Assassin's Creed: Black Flag is an Nvidia-sponsored game, so we're curious as to how the company's GPUs measure up to AMD's in this attractive-looking title.


The GeForce GTX 750 Ti is playable at the Normal preset, but it finishes our test in the back of the pack, just shy of the GeForce GTX 650 Ti Boost's average frame rate, but more significantly in front of the GeForce GTX 650 Ti it's destined to replace.


Frame time variance is low overall. However, the GeForce GTX 750 Ti does exhibit a couple of spikes. Of course, this is the Maxwell architecture's first showing, and we're not sure what driver work needed to be done to optimize for it. Hopefully Nvidia's engineers iron this sort of thing out over time.
The move to Battlefield 4 allows us to choose the Ultra preset, though we drop MSAA to 2x and ambient occlusion to the SSAO setting at 1920x1080.


The GeForce GTX 750 Ti is certainly playable, although its frame rate does dip below 30 FPS on a couple of occasions. Nevertheless, it comes within striking distance of Nvidia's GeForce GTX 650 Ti Boost.


Three cards suffer from frame time variance spikes at these detail settings, and the GeForce GTX 750 Ti is one of them. The other two are AMD's Radeon R7 260X and HD 7850, both of which have 1 GB of graphics memory.
You need little more than a mid-range graphics card to play BioShock Infinite at its Ultra preset at 1920x1080.


The GeForce GTX 750 Ti ties Nvidia's GeForce GTX 650 Ti Boost in our average frame rate measurement, although it achieves a significantly higher minimum, too.


The Unreal engine that powers BioShock is well-optimized, yielding low frame time variance figures compared to other games.


Nvidia's GeForce GTX 750 Ti falls in between the GTX 650 Ti Boost and Radeon R7 260X in our benchmark sequence.


Frame time variance spikes in a handful of cases. Experientially, though, the game is smooth, so we're not coming down hard on any of these results.
We know from reviews past that Grid 2 isn't particularly graphics-bound. Rather, it tends to reward upgrades to host processing and memory. As a result, we're able to run this game at 1920x1080 using its Ultra preset and 8x MSAA.


The GeForce GTX 750 Ti serves up playable performance, though it brings up the rear amongst these seven other cards. With that said, it also appears between the GeForce GTX 650 Ti and 650 Ti Boost, just as we'd expect.


We record a couple of frame time variance spikes, but nothing so severe that it affects the title's smoothness.
Our last game benchmark is Metro: Last Light. We're running at 1920x1080 using the DirectX 11 code path and the Medium quality preset.


Nvidia's GeForce GTX 750 Ti ties the Radeon R7 260X in average frame rate, but manages a slightly better minimum.


Measured frame time variance is low across the board, though, as we've mentioned previously, hiccups do register on a number of cards from both vendors.

When we average the performance of all of these cards across all of our benchmarks, the GeForce GTX 750 Ti appears to perform a lot like AMD's older Radeon HD 7850. Although that's a decent accomplishment in and of itself, since the 7850 launched at $250 (and the only card in stock on Newegg still sells for $200), it's made even more impressive when you consider that the Maxwell-based board doesn't even need an auxiliary power connector.
AMD recently cut the prices on many Radeon R7 260X cards, improving their value. Sapphire's implementation comes overclocked from the factory, includes 2 GB of GDDR5 memory, and can be found for about $130, making it a good upgrade option for gaming at 1920x1080. Read the Full Review
Sapphire Radeon R7 260X 2GB
Mainstream Performance
Digging further, we took the average performance of all cards across five games, and then recorded detailed power usage stats during the runs. Using those results, we calculated the average frame rate per second per watt:

This is the chart that matters when you want to fold a 60 W TDP into the performance discussion. Nvidia's GeForce GTX 750 Ti really shines, almost doubling the performance/watt of AMD's Radeon R7 265 and Nvidia's GeForce GTX 650 Ti Boost. The Radeon R7 260X fares better, but still trails the GM107-based card's efficiency by a significant distance.
Factory Clock Rates and GPU Boost
We already know that the base clock rate of Nvidia's GPUs isn't indicative of what those processors can run at under load thanks to the company's GPU Boost technology, so long as certain conditions are met.
Base and GPU Boost frequencies can be set in firmware by board partners. So, depending on the card and its cooler, you might see varying Boost ceilings, even between products with the same base clock. In the following chart, Gigabyte posts the most conservative ceiling. However, it's able to sustain a frequency that does not fluctuate. The other three boards, lacking auxiliary power connectors, do bounce around more.

Overclocking? Yes, But Only A Little
After long stability tests using the partner cards, we settled on ~1.3 GHz as the highest usable GPU Boost clock rate. At that point, the power limit was intervening fairly restrictively. Unfortunately, this limit is set to 100%. Until Nvidia fixes whatever is preventing higher power targets, it won't be possible to push the board harder (up to the PCI Express slot's 75 W ceiling).
In terms of performance, the Gigabyte and MSI boards fare similarly. The GTX 750 Ti Gaming OC does manage to push higher peak clock rates, but then has to dial back more often due to the TDP limit. Gigabyte's offering delivers more rounded behavior. It also operates at an incredibly-cool 46 °C, despite our overclocked settings.
First, we'll compare base clock rate settings to observed GPU Boost frequencies:

It's clear that as you push overclocked frequencies higher, the card steps in more often to dial back down. In reality, this means you don't get a ton of benefit in real-world gaming. Let's compare the increased frequencies to the corresponding frame rates in Metro: Last Light and Crysis 3, which we normalize and average together.

Roughly 14%-higher base clock rate (11% GPU Boost) yields about 7% more gaming performance. Power consumption barely budges though, and the GeForce GTX 750 Ti continues respecting its TDP limit. Short-term peaks (102%) are countered immediately with a clock rate reduction. But by capping the power limit to 100%, we end up with fairly modest gains, regardless of whether more power is available through a six-pin connector or not. Nvidia needs to fix its software-imposed ceiling before we're able to push higher.
Single-Precision, Good. Double-Precision, Bad.
Nvidia's mainstream Kepler-based GPUs offered double-precision compute performance that was 1/24 as fast as its FP32 math. Maxwell is ever worse at 1/32. Of course, that's purely theoretical until we double-check it with real-world benchmarks.
The Folding@Home benchmark is particularly good for comparing graphics cards under OpenCL. We had to do without CUDA-based numbers this time around because the Maxwell-based card wasn't properly recognized. This is simply something we'll have to put together later.
How big is the difference between single- and double-precision, really? Our benchmark results indicate a 8:1 ratio between them on Nvidia's GeForce GTX 750 Ti (Maxwell). That's quite a bit weaker than the GeForce GTX 760's (Kepler) 4:1 in this metric (Explicit Solvent).
But GM107's comparatively strong single-precision performance is what sticks out; it's able to compete with much more potent graphics cards. By the time you get to its double-precision numbers, compute throughput ends up just below where we would have expected it.
Single-Precision Benchmarks (SP)


Double-Precision Benchmarks (DP)


We're not sure what to think about GM107's increasingly hobbled FP64 capabilities. You can either say double-precision performance is really bad, or the single-precision numbers are really good. Regardless, at the end of the day, artificial limitations meant to prevent cheap desktop cards from being viable workstation parts are no less irritating.
The following disciplines don't capture every aspect of general-purpose computing (there are still Photoshop, WinZip, and video editing benchmarks in our suite that we weren't able to run). Nevertheless, they give us a good idea of how Maxwell improves upon the Kepler architecture, and where Nvidia now sits relative to AMD.
BTC Mining

Although the company isn't going into depth on the improvements it made, we know that Maxwell handles hashing far better than Kepler, which is reflected by its victory over the GeForce GTX 680 and 770. At least in BTC mining, however, GeForce GTX 750 Ti still gets hammered by Radeon R7 260X and some of AMD's other more mainstream cards.
Of course, Bitcoin isn't the only cryptocurrency reliant on hashing. MaxCoin, for example, is a member of the SHA3 family, and it's supported in the latest version of CudaMiner. Curious as to how GeForce GTX 750 Ti sizes up to 650 Ti, we ran the following SHA2-based test in Sandra 2014:

There are big gains to be had from DirectX's Compute Shader, but throughput via CUDA is downright phenomenal. It's probable that Maxwell improves some of the integer operations that were slower on Kepler. Hopefully Nvidia opens up more about what the new architecture can do.
LTC Mining
Bitcoin mining is almost irrelevant to CPU and GPU miners these days, if only because it's impossible to compete with dedicated ASIC- and FPGA-based devices working so much faster. But at least for a short time still, Litecoins remain at least a somewhat viable option. The use of scrypt (a password-based key derivation function) in their proof-of-work algorithm, rather than Bitcoin's SHA-256, makes dedicated hardware more difficult to develop. So, GPUs still rule, even if increasing difficulties make the investment in equipment and power greater than current returns.
Historically, Nvidia's cards came up short against competing Radeons, which is why you see R9 290X boards selling for $700 and up. But the Maxwell architecture's improvements allow the 60 W GeForce GTX 750 Ti to outperform the 140 W GeForce GTX 660 and approach AMD's 150 W Radeon R7 265, which just launched, still isn't available yet, but is expected to sell for the same $150. On a scale of performance (in kH/s) per watt, that puts Nvidia way out ahead of AMD. Today, four GM107-based cards in a mining rig should be able to outperform a Radeon R9 290X for less money, using less power.
LuxMark 2.0

LuxMark is another bastion for AMD, where the compute performance of its GPUs typically dwarfs competing Nvidia cards. The Radeons are so much faster, in fact, that Nvidia typically avoids addressing our results, stating only that it doesn't optimize for compute workloads on its gaming cards.
Again though, for a 60 W board, the GeForce GTX 750 Ti does pretty well. It still falls behind the Bonaire-based cards in its price segment. But compare its score of 943 to the GeForce GTX 580's 893. That's an almost-250 W former flagship!
RatGPU

AutoCAD 2013: 2D Performance DirectX (Cadalyst)
Although our comparison cards finish close to each other, the Maxwell-based board has a slight lead. Once again, this shows that 2D functions don't access hardware directly anymore (besides blitting and stretching), creating a bottleneck dependent on the rest of the platform. Still, there's more than enough performance for every use case.

AutoCAD 2013: 2D Performance DirectX (Cadalyst)
The GeForce GTX 750 Ti's performance in the Cadalyst suite's 3D component is interesting. Improved single-precision compute performance is evident. After all, DirectX uses SP nearly exclusively in its libraries, and all vertices and their transformations profit heavily in this simple 3D presentation.

Autodesk Inventor 2013: DirectX
We're using our 1000-cube extreme workload again. Contrary to what we saw in Cadalyst, the GeForce GTX 750 Ti falls in where we'd expect it to, based on the board's market position.

Maya 2013: OpenGL
Finally, we have a benchmark that reflects OpenGL performance, but is also computationally intense and memory dependent. The fact that theoretical performance isn't indicative of real-world behavior, and that less powerful cards often achieve better results is particularly interesting.

Temperatures
Our thermal measurements come from a long loop of the Metro: Last Light benchmark. By the end of the chart, you should be looking at the highest temperatures you'd see in a real gaming environment. Of course, depending on the title, you could certainly be looking at lower readings.
Two of our labs ran tests in closed mid-tower cases, finding the various GeForce GTX 750 Ti cards to vary by one to two degrees. Even though these cards dissipate heat internally, rather than through their rear I/O slots, the fact that they use so little power means heat is hardly an issue.
| Model: | Idle | Gaming Load |
|---|---|---|
| GeForce GTX 750 Ti Reference | 28 °C | 65 °C |
| Gigabyte GTX 750 Windforce OC | 25 °C | 45 °C |
| MSI GTX 750 Ti Gaming OC | 26 °C | 54 °C |
| Zotac GTX 750 Ti OC | 27 °C | 63 °C |
The following chart demonstrates that each card's temperature curve is affected by the heat sink, fans, and BIOS-based fan profile.
Gigabyte cools its card very aggressively, and could have pulled back on fan speed to bring noise down even more (even though the card is hardly audible as it sits).

Noise
We record noise measurements for each card using a studio microphone calibrated for our PC audio tests. The mic is oriented perpendicularly to the graphics card's center, 50 cm away.
| Model: | Idle | Gaming Load |
|---|---|---|
| GeForce GTX 750 Ti Reference | 31.5 dB(A) | 34.1 dB(A) |
| Gigabyte GTX 750 Windforce OC | 31.9 dB(A) | 33.2 dB(A) |
| MSI GTX 750 Ti Gaming OC | 30.0 dB(A) | 31.9 dB(A) |
| Zotac GTX 750 Ti OC | 31.1 dB(A) | 33.0 dB(A) |
It really worked out well that we received so many GeForce GTX 750 Ti cards for our launch coverage, since we're able to take measurements from products with big, beefy coolers and others (like the reference board) with tiny orb-style heat sinks.
Nvidia's own design proves that it doesn't take much to keep GM107 cool. A small sink and fan are all you need, really. And in a closed chassis, you aren't going to hear any of these four samples.
With that in mind, MSI's GTX 750 Ti Gaming OC appears to offer the best compromise between quiet operation and high performance. It's just too bad that the MSI and Gigabyte boards employ such big, bulky coolers. The whole point of Maxwell is efficiency; we'd like to see vendors start introducing thermal solutions to match.
Benchmark System And Procedure
We collaborated with HAMEG (Rohde & Schwarz) to upgrade our power consumption measurement system. It’s now one step away from being complete and will be continuously refined.
We record all channels and the corresponding oscilloscope value/curves for our measurements. The very precise and (more important) fast current clamps yield 100 mV/A, making it easy to calculate power based on the voltage. We also record the supply voltage to multiply its value with the recorded amperage. Depending on the resolution we choose, this procedure yields a very detailed power consumption history. We generally set this to 1 ms, allowing us to capture all fluctuations attributable to AMD’s PowerTune or Nvidia’s GPU Boost technology.
| Measurement Procedure: | Non-Contact Direct Current Measurement at the PCIe Slot Non-Contact Direct Current Measurement at the External PCIe Power Supply Direct Voltage Measurement 3.3 V / 12 V |
|---|---|
| Measurement Apparatus: | Oscilloscope: Power Clamp: Voltage Divider Probe: HAMEG HMC8012 |
| Bench Table: | Microcool Banchetto 101 |
| Test Hardware: | Intel Core i7-3770K (Ivy Bridge), Overclocked to 4.5 GHz Corsair H100i Compact Water Cooling Solution 16 GB (2 x 8 GB) Corsair Vengeance DDR3-1866 Gigabyte G1 Sniper 3 + Modified PCIe Adapter with Current Loops |
| Power Supply: | Corsair AX860i with Modified Plugs (Pickups) |
Power Consumption: Gaming
We're particularly interested to learn how Nvidia improved the efficiency of its architecture. In terms of total power consumption, the card averages (in a warmed-up condition, after applying full load) just a little bit over the GPU's TDP. This measures the whole card though, including memory. Thus, Nvidia keeps its word and clearly sets a new standard for performance per watt.
For the first time ever, the average power figure isn't centered between minimum and maximum values, but rather closer to the bottom of the chart. Peaks are clearly rarer, but they're also more extreme. This is important because the card is driven by the motherboard's PCI Express slot, making the platform's quality more important than it might otherwise be.

Because the 170-second run is hard to display at the resolution we used, here's a 10-second break-out for additional detail:

The snippet illustrates how important fast measurements are to the average power consumption value. Of course, we can get even more precise, so this next graph represents just one second. Fluctuations in draw effectively demonstrate what your power supply has to be capable of handling.

But how does consumption change when you take into account more use cases?
Next to its competition, the GeForce GTX 750 Ti earns our respect when it comes to evaluating power consumption.
By the way, those two watts we shaved off on this page in the gaming measurement are the result of different benchmark settings. While we typically use the most demanding settings possible for better comparability in this overview, we stepped back to quality settings that allow playable frame rates. Interestingly, power consumption turns out to be a little higher under those conditions.




Taking each of these measurements into account, we have to congratulate Nvidia for its enormous efficiency improvement. The GeForce GTX 750 Ti indeed gets along without an additional PCIe power connector even under extreme load.
But it's probable that the 750 Ti operates close to the GM107's sweet spot. If true, efficiency would suffer in the face of significant overclocking, closing the PCI Express slot's slight headroom. The fact that some of Nvidia's board partners are adding six-pin power connectors suggests more serious overclocks stand to benefit from extra juice.
Last week, AMD made two announcements. First, it was dropping the price of its Radeon R7 260X to $120, effective immediately. Second, it previewed the Radeon R7 265 at $150. Today, the 260X remains a mostly-$140 card. And the 265 isn’t expected for a couple of weeks yet.
So, Nvidia’s GeForce GTX 750 Ti 2 GB, which it says will be available by the time you read this for $150, really does battle with AMD’s Radeon R7 260X and what’s left of the outgoing GeForce GTX 650 Ti. Based on our benchmarks, we know that the GM107-powered board is quicker than both slightly cheaper cards. At least from a gaming angle, Nvidia’s pricing seems appropriate, though not particularly value-oriented.
Then again, this is perhaps the most volatile we’ve ever seen the graphics card market. If AMD manages to ship its R7 265 at $150, it’ll have a much faster card—something better-suited to battling GeForce GTX 660, currently selling for closer to $200. Gamers with modest power supplies would likely find that their best bet in a pure breakdown of performance per dollar.
It’s difficult to make this story all about frame rates when we’re comparing 60 and 150 W GPUs, though. Surely, power and efficiency have to also come into play. We tend not to dwell on those figures too intently because gamers don’t lie awake at night, worried about how many cents per kilowatt-hour they’re spending. But power does come into play when you’re eyeballing small form factor cases, when you’re trying to upgrade an old tier-one box, and when you’re mining cryptocurrencies.

In each of those situations, the GeForce GTX 750 Ti turns the Maxwell architecture’s significant efficiency improvements into palpable benefits. The card is small, for starters. Although every implementation we’ve seen thus far chews up two expansion slots, it should still fit in spaces less forgiving of longer boards. And we’re hoping to see single-slot implementations, too. Of course, it helps that 60 W of heat is easier to dissipate in the confines of a compact gaming box.
Or how about an older hand-me-down system with integrated graphics and a 300 W power supply? Without the extra connectors to support a modern discrete graphics card, you’re forced to either buy a new power supply or settle for the highest-end card with no other dependencies. Previously, that was AMD’s Radeon HD 7750. Now, it’s the GeForce GTX 750 Ti.
Nvidia's GeForce GTX 750 Ti, based on the Maxwell architecture, offers solid performance at 1920x1080 and is only rated for 60 W, so it doesn't need auxiliary power. This makes it ideal in aging gaming PCs or small form factor enclosures. Read the Full Review
GeForce GTX 750 Ti
Efficient Mainstream Graphics
And then there’s the cryptocurrency mining story, which currently shoulders the blame for unreasonable prices on most of AMD’s line-up. In absolute terms, the GeForce GTX 750 Ti is just about as fast as Radeon R7 265. But again, that’s a 150 W GPU. The new GeForce uses just 40% of its power. You could throw four onto a motherboard for $100 less than one Radeon R9 290X, use less power, and achieve higher hash rates. Let’s just hope Nvidia’s Maxwell-based parts don’t similarly suffer from their own aptitude in this discipline.
In the end, there are a number of situations where Nvidia’s GeForce GTX 750 Ti cannot be matched by any card from either vendor. In other comparisons, the GM107-based card is strong, but simply priced fairly based on its competition. And then there is an as-of-yet unrealized match-up where AMD can deliver a stunning value-packed counter-punch, though we remain skeptical of Radeon R7 265’s price and availability.
Even if it is able to deliver plenty of 265s at its promised price, AMD will undoubtedly be facing other Maxwell-based part in the near future. And based on GM107’s showing, we’re excited to think about what Nvidia might do with this architecture and a 250 W power budget.




























