Today Nvidia is pulling the wraps off its all-new Maxwell architecture, which has one defining feature -- it's twice as efficient as Kepler. Instead of launching with a high-end $500 GPU like it's done in the past though, this time it's going the opposite direction by launching at the entry level with the promise that it will eventually release 200w+ TDP cards based on Maxwell. For now this is the new low-power king at the $150 price point with its rock-bottom TDP of just 60w and the ability to pull all the juice it needs from a PCIe connector, no six-pin power connector required. Let's take a look at what Maxwell is all about, and how the card fares against its rival from the red team.
Right off the bat let's clear up two things about Maxwell . First, it is using the same 28nm process that Nvidia used for Kepler instead of 22nm like everyone thought. Second, though there has been a lot of discussion online (and even from us) about how Maxwell would be the first GPU to have an embedded ARM CPU, that is not happening on the desktop cards as far as we can tell, so we were wrong about that, along with everyone else. It now appears that all that talk about Maxwell having an integrated ARM core was simply wrong, or maybe it was just the next-gen Tegra part named K1 , which has a Kepler core and an ARM CPU.
Now that we know what Maxwell isn't, let's talk about what it is, which is a radically redesigned architecture that makes Kepler look like a fire breathing dragon. The chip inside the GTX 750 Ti is named GM107, and according to Nvidia it was designed, "for use in power-limited environments like notebooks and small form factor PCs." Nvidia is targeting very small Steam Machines with the GTX 750 Ti as well. Nvidia notes that the first-generation Maxwell products will be focused on low power operation, with "higher performing second generation Maxwell GPUs addressing the enthusiast graphics segments at a later date."
Nvidia claims as it transitioned Kepler into a mobile part it had to make a lot of changes to the architecture to save power, and that all that knowledge went into Maxwell. It has completely revamped the structure of the chip with specific attention paid to the arrangement and management of the CUDA cores. The SMX design from Kepler is gone, replaced by a new structure named SM for Streaming Multiprocessor. According to Nvidia, the improvements are mainly in "control logic partitioning, workload balancing, clock-gating granularity, compiler-based scheduling, and number of instructions issued per clock cycle" as well as other enhancements. These changes allow Nvidia to use more SM units in general but each of them have fewer CUDA cores (more on this in a bit). The result is that a chip like GM107 ships with five SMs compared to two in the GK107 chip used in the GTX 650 it replaces.
The GM107 compared to Kepler's GK107 - more power, with less power.
One other big change in Maxwell has a fat 2MB of L2 cache compared to just 256KB in GK107. Nvidia says the extra cache reduces the amount of requests sent to the graphics card memory, which conserves energy. Overall, the GM107 contains a single Graphics Processing Cluster (GPC), which includes five SM units for a total of 640 CUDA cores, and two 64-bit memory controllers for a 128-bit bus.
The biggest change in Maxwell is of course the redesigned Streaming Multiprocessor units. Intead of having 192 CUDA cores per unit with one piece of control logic for all of them, the new design has each SM partitioned into four distinct processing blocks with each one having its own control logic. Each partitioned block hosts 32 CUDA cores, so with four blocks per unit the total CUDA cores is 128 per block, and Nvidia claims this approach allows for 35 percent more performance per CUDA core on workloads that are limited by shader performance. Pairs of these blocks share texture filtering units and cache, thereby saving space and power. Overall, the smaller blocks deliver 90 percent of a Kepler SMX unit while taking up much less space, allowing Nvidia to put more of them on the GPU. Since each block's control logic has to only deal with 32 CUDA cores instead of 192 in Kepler, it makes their job simpler and more efficient. As we noted previously, it has placed five of them in the GM107 compared with just two in the previous chip, allowing it to utilize 1.7 times more CUDA cores and thus over 200 percent more shader performance. It basically boils down to less cores being used equals less power consumed, but since the design is more efficient performance is also improved.
Maxwell's SM units feature four "processing blocks" per unit, with four partitions of 32 CUDA cores each.
Continue reading for a closer look at the card, benchmarks, and closing thoughts.
To recap briefly, the GPU at the heart of this card goes by the call-sign GM107 and it is a full implementation with 640 CUDA cores, dual 64-bit memory controllers for a 128-bit wide bus, and 1GB or 2GB of GDDR5 memory clocked at 5.4GHz. The 750 Ti includes the same GPU Boost 2.0 technology found in the GTX 700 series of cards, so it'll automatically overclock itself under load despite having surprisingly high stock clocks of 1,020MHz base clock and a 1,085MHz boost clock. Nvidia says the card can easily do 1,250MHz as well. Nvidia's add-in-board partners will also be offering the usual overclocked versions of the card too. The card slides in right above the GTX 650 in Nvidia's product stack, replacing the GTX 650 Ti, and is one notch below the GTX 660.
Despite its entry-level nature, the GTX 750 Ti supports ShadowPlay and G-Sync.
The big "talking point" of this card is its lack of six-pin PCIe power connector, though the back of the board has solder points for such a connector in case an add-in-board partner wants to put one there, but the reference design is completely bereft of a connector. This translates to the ability to stick it into any PC on the market with a PCI Express slot. Nvidia points out in the press materials this card is meant to be wedged into almost any PC out there, including small form factor boxes, Steam Machines, and even older desktops that are using integrated graphics. The card's TDP is a mere 60w, which is easily the lowest of any GTX card we've ever seen. Nvidia's minimum power supply requirement for this card is just 300w, and it notes that the PCIe slot delivers 75w, so this card isn't even using all of the power the slot can provide. In general, the GTX 750 Ti's mission in life is to replace integrated graphics on older PCs, so it'll slide right into the millions of Dell, HP, and Lenovo desktops that ship with integrated graphics and provide very good 1080p gaming at a decent price.
It's also an entry-level GTX card from Nvidia, so you get access to everything in the ecosystem including Gamestream for your Shield, GeForce Experience, Shadowplay, G-sync, and the GPU's ability to overclock care of GPU Boost 2.0. It does not support SLI, however. Its connectors include two dual-link DVI connectors and one HDMI connector, and like most GPUs it is a dual-slot card. Here are the full specs straight from the green horse's mouth:
Full specs of the GTX 750 Ti
Before we show you the benchmarks, let us be clear about one thing, which is we only had time to test it against its direct competitor, which is the AMD R7 260X. In Nvidia's briefing documents this is/was the card they are/were targeting, and we use the past tense because last Thursday AMD announced the R7 265 in an offensive maneuver to fend off the attack from this very GPU. While doing so, AMD also dropped the price on the R7 260X down to just $120. Nvidia is also launching a non-Ti version of the GTX 750 with the same name for the exact same price of $120. It ships with the same clock speeds as the GTX 750 Ti but has less shader performance due to less CUDA cores via one less SM unit, and a slightly lower memory clock (5GHz vs. 5.4GHz on the Ti). We do not have the GTX 750 nor do we have the R7 265 for review, so we compared the GTX 750 Ti to its closest competitor. We are rounding up all these cards however for a mega-shootout in an upcoming issue, so stay tuned. With that disclaimer out of the way, here are the numbers:
Best scores are bolded. Our test bed is a 3.3GHz Intel Core i7-3960X Extreme Edition in an Asus Rampage Extreme IV motherboard with 16GB of DDR3/1600 memory and a Thermaltake ToughPower 1,050w PSU. The OS is Windows 8 64-bit Enterprise. All games run at 1920x1080 with maximum settings and 4X AA except for 3DMark.
Looking at the benchmark chart there's one piece of information that is missing which puts these numbers in a whole new light. The AMD R7 260X has a TDP of 115w, and the Maxwell card is just 60w. We know that "energy efficiency" isn't the most titillating spec or feature here at Maximum PC, but we give this card big props for being able to hold its own against a card with double the power rating. That is impressive. However, since AMD has launched a new GPU specifically to attack this card, and we don't have that card to benchmark yet (the R7 265), this battle is far from over. Also, the AMD card has a TDP of 150 watts, so it's more than double that of the Maxwell-based GTX 750 Ti.
Given their differences in power consumption it's almost fair to say these cards aren't even playing in the same league as Nvidia is targeting PCs without any PCIe power connectors, and AMD is targeting value-conscious gamers that are looking to upgrade from a much older 1080p GPU.
Now, as a reader of Maximum PC you might not be too stoked on a card designed for "low power environments," and we feel you. This is how Nvidia is handling it right now though, as the $100-$150 is the lion's share of the GPU market globally. Plus, it already has a handle on the high-end of the market with the GTX 780 Ti , and the all-new Titan Black . Also, AMD's pricing and supply issues are only strengthening its hand as well in both the high-end and the mid-range too. That leaves just one area for it to really sock it to AMD, and it's chosen the low-power route. It's tantalizing to consider what a flagship GM110 GPU might be capable of though, but sadly we won't see it until probably Q4 based on an estimate we have pulled from our nether regions.
All in all, the Maxwell-based GTX 750 Ti certainly delivers on the promise of being more efficient than Kepler, which is good news for gamers who like quiet machines and bad news for AMD and its power-hungry GCN architecture. Whether or not that will factor into peoples' purchasing decisions remains to be seen but Nvidia clearly has the performance-per-watt lead with Maxwell, and we can't wait to see it introduce more powerful GPUs using the same architecture. Also, it's worth noting that despite its power-sipping status we were able to overclock the card all the way up to 1,272MHz on stock voltage, and under load it ran at 67 C.
The GTX 750 Ti 2GB version will sell for $150, with the GTX 750 (we assume also with 2GB) selling for $120. Retailers will also be selling 1GB versions of the GTX 750 Ti for $140 later this month.