Testing Maxwell: Nvidia's Six-Inch Hammer

By Loyd Case

Nvidia recently unleashed its latest graphics architecture on the world, Maxwell. The first iteration is the GTX 750, a GPU that will be the core of a graphics card whose asking price will typically be under $150. Here's how it performs.

Nvidia recently unleashed its latest graphics architecture on the world, Maxwell. The first iteration is the GTX 750, a GPU that will be the core of a graphics card whose asking price will typically be under $150. Two variants of the GTX 750 will be shipping, the GTX 750 and the GTX 750 Ti. I’ll get into the differences shortly.

Once upon a time, the first iteration of a GPU architecture would show up as powerful, power-hungry beasts of a graphics card. That changed a bit when Nvidia introduced its first Kepler card aimed at gamers, the GTX 680. The GTX 680 took the gaming world by surprise, delivering leading edge performance, but was miserly on power consumption and low on fan noise, particularly when compared to Nvidia’s earlier GTX 580s and AMD’s Radeon HD 7970s.

Still, the GTX 680 was a high end card, even though it broke the mold a bit as to what a high end card should be. The GTX 750 Ti’s typical asking price is $150. The overclocked EVGA GTX 750 I’ll be looking at is $169, but it’s both overclocked and ships with a 2GB frame buffer. You can find GTX 750 Ti cards from several manufacturers running at reference clocks for $149. GTX 750 1GB cards can be had for as little $119. I’ll take a look at performance of the EVGA GTX 750 Ti SC and an Nvidia GTX 750 Ti reference card, which also has a 2GB frame buffer, but runs at standard core clocks.

But first, let’s look at Maxwell.

Meet the new boss, much like the old one

The new GTX 750 and 750 Ti chips are based on the same GM107 die. At the individual shader core level, Maxwell is pretty similar to Kepler. What is new is a substantial reorganization of the GPU layout. For example, Kepler has a monolithic control logic unit that managed scheduling for up to 192 cores. Maxwell now allocates a smaller, more efficient control logic unit for each block of 32 cores.

Nvidia maintains that smaller control logic units handling scheduling for fewer cores is more power efficient and also results in better performance.

This change in the scheduler, a larger L2 cache (2048MB versus 25K in the Kepler-based GTX 650) and a large number of smaller improvements allowed Nvidia to build 640 shader cores on a die, versus 384 on the GK107-based GTX 650. Die size has increased as well. (Note that the GTX 650 Ti was a different beast entirely, since it was a cut-down GTX 660 GPU, not based on the GK107 variant of Kepler.) Let’s take a quick look at the difference between Nvidia’s budget GPUs.

SpecGTX 650GTX 650 TiGTX 750GTX 750 Ti
GPU CoreGK107GK106GM107GM107
CUDA Cores384768512640
Base Clock1058MHz925MHz1020MHz1020MHz
Boost ClockN/AN/A1085MHz1085MHz
L2 Cache256K384K2048K2048K
Memory Speeds5Gpbs5Gbps5.4Gpbs5.4Gpbs
Memory width128-bit128-bit128-bit128-bit
Texture Units32643240
Rated Power64W110W55W60W
Power Connector6-pin6-pinnonenone
Transistor Count1.3B2.54B1.87B1.87B
Die Size118mm-sq221mm-sq148mm-sq148mm-sq
The GTX 750 and 750 Ti don’t require a power connector. Even some overclocked models lack a power connector.
The remarkable thing about the GTX 750 Ti is that it requires no power connector.

The remarkable thing about the GTX 750 Ti is that it requires no power connector, yet has nearly double the shaders, more transistors and a larger die size than the last generation. The real question is, how does it perform? I stacked it up against an XFX Radeon R7 260X, which typically costs around $140. Note that AMD also makes the R7 265, a $150-$160 card, but couldn’t obtain one in time for this review. The 265 has 1024 AMD GCN shader cores, or 128 more than the R7 260X, so it’s likely to about 10% faster.

I ran benchmarks using the same Core i7 3770K testbed I’ve used in past review, which has 8GB of RAM, a 240GB SSD and an 850W PSU. It’s likely this system will retired soon, moving to Haswell and Windows 8.1 in the future.

With these configurations and caveats in mind, let’s take a look at the cards, then dive into performance.

EVGA’s GTX 750 Ti SC

When I initially contacted EVGA, they first offered us a highly overclockable version of the GTX 750 Ti, with dual cooling fans and a 6-pin power connector. I thought this configuration downplayed the real strengths of the new Maxwell, so opted instead for a single fan GTX 760 SC. It’s still overclocked, running at 1176MHz base clock (versus the 1020MHz reference) and 1255MHz boost (as compared to the 1085MHz reference boost clock.) This card is still the short form factor version, and still requires exactly zero power connectors.

This EVGA GTX 750 Ti Superclocked card is just six inches long and lacks a power connector.

At six inches long, this card will fit in just about any case. Contrast that to the more expensive GTX 760, which is 9.5 inches long. Even the lower cost XFX R7 270X is seven inches long.

The GTX 750 Ti is tiny, even compared to the relatively compact XFX card.

Just for fun, let’s also take a look at how EVGA’s card looks next to Nvidia’s reference design.

The cards are actually the same length, but the EVGA card looks a little bulkier due to the shroud around the fan.

Power and Size

Power and physical card size is becoming an increasingly important component in modern PC systems. Small form factor systems are increasingly popular, with a new mini-ITX motherboard or case popping up on the scene every few weeks, or so it seems. Nvidia makes much of this in the reviewer’s guide, suggesting the GTX 750 Ti is ideal for very compact systems where gaming performance may still be important.

Nvidia rates power usage at 60 watts; EVGA also rates their superclocked card at 60W. How does total system power stack up against the competition? I collected system power consumption at idle, while running 3DMark 2013 at very high resolutions and with Metro Last Light with all the eye candy maxed out.

The EVGA card is more power efficient than the rest, though the margins aren’t really big, except when compared to the higher end GTX 760.

As you can see, the GTX 750 Ti is pretty power efficient. But how does it perform?


Spoiler alert: the EVGA GTX 750 Ti performs pretty darned well at 1080p. Let’s look at benchmarks.

The GTX 750 Ti posts substantially better scores in 3DMark 2013’s Fire Strike test than other budget-minded cards. Sure, the GTX 760 is faster, but it’s also about $100 more.

Maxwell is substantially faster in the Unreal-engine based Bioshock Infinite than AMD’s R7 260X. Even the GTX 650 outpaces AMD’s budget card in this benchmark.

EVGA’s iteration on Maxwell is the first budget card we’ve seen that exceeds 30fps in Metro Last Light, at the high setting. If you’re willing to dial back anti-aliasing, you can get over 45fps.

When the new Tomb Raider arrived, much was made of the game’s TressFX hair rendering, and AMD had a big advantage over Nvidia at the time Tomb Raider shipped. The GTX 750 Ti pretty much crushes the R7 260X; I suspect even the R7 265 wouldn’t match the EVGA card.

Racing games tend not to stress the GPU too much, and GRiD2 is no exception. At the high setting, the game looks pretty much CPU bound, with the resulting performance roughly in the same ballpark for all the cards. Cranking it up to Ultra, with Soft Ambient Occlusion, separates the men from the boys, however, with Maxwell dancing past the other budget choices.

I’ve started using this test recently; it’s a series of flythroughs of scenes of differing complexity, and can be pretty hard on a GPU, though not to the extent of Metro Last Light. You won’t run Sleeping Dogs in Ultra mode with any of these cards, but at high settings, EVGA’s GTX 750 Ti SC generates playable frame rates.

ShadowPlay Recording

The latest version of ShadowPlay generates gameplay videos while playing with no noticeable frame rate hit.

Nvidia’s ShadowPlay records gameplay videos as you’re playing. Early versions of ShadowPlay hit frame rates pretty hard, even on high end cards. However, the latest version of ShadowPlay generates gameplay videos while playing with no noticeable frame rate hit.

Nvidia’s significantly tweaked their H.264 hardware encode engine to improve performance and efficiency. It’s impressive that you can record up to 20 minutes of video while playing a game, and not worry about hits on your frame rate.

Maximizing Budget Efficiency

At first, it seemed surprising that the first iteration of Maxwell would be a budget-priced card. Based on what I’ve seen, though, it was a smart move on Nvidia’s part. Cards at the $150 price point sell a lot more units than the high end gear. Given that the GTX 750 Ti can run most games at 1080p while maintaining relatively high graphic fidelity makes it likely that a lot of these cards will end up in gamer’s systems. Toss in great power efficiency and compact size, and EVGA’s superclocked version of the GTX 750 ti is a real winner.