First generation Maxwell architecture brings major perf/watt improvements, but hits mainstream rather than enthusiast segments on début.
First generation Maxwell architecture brings major perf/watt improvements, but hits mainstream rather than enthusiast segments on début.
Today NVIDIA are releasing two new GPUs into the wild – the GeForce GTX 750 and 750 Ti. Our review of the GTX 750Ti is of course available, but this article is intended to provide some context for the release of this GPU, and more widely NVIDIA’s new Maxwell architecture. If all you need are raw performance numbers and solid conclusions from our graphics expert we urge that you check out Rich’s take, but otherwise read on.
The Way It’s Meant To Be Played
You’ll no doubt have seen the phrase littering marketing literature over the last several years, but it’s taking new prominance as NVIDIA’s platform moniker. TWIMTBP is the umbrella term for all of NVIDIA’s gaming IP and encompasses Gameworks developer relations programme, Gamestream, G-SYNC and GeForce Experience. This slight change in emphasis coincides with the release of the first Maxwell GPUs, but will be a major part of platform development moving ahead.
NVIDIA’s GeForce GTX 750-class GPUs have been dubbed the first generation of Maxwell, and for good reason. Unlike previous evolutions of their graphics architecture Maxwell is being introduced at the mature 28nm process, the same as the now veteran Kepler architecture. That’s an exceptional step – generally speaking manufacturers have relied on a die shrink to bring about meaningful performance per watt improvements, giving them greater overhead for performance at the top-end. Later generations of Maxwell are planned for the 20nm-era once the process becomes viable, and will encompass all performance strata. For the record, the specific core used in GTX 750-class GPUs is the GM107.
Both GTX 750 and 750Ti SKUs are pitched at the mainstream, slotting in between the GTX 650 and GTX 660 to replace the GTX 650Ti & 650Ti Boost. Once again this is an uncommon step for NVIDIA who have tended to bring new architectures to the market at the bleeding edge of performance. This step shouldn’t be seen as a pull-back from the enthusiast market, already well served by GeForce designs, but rather a pre-emptive beefing up of the mainstream before new third-party products come online in the next few months. More on that later.
The GeForce GTX 750 Ti – Justified By Efficiency
According to the Steam Hardware Survey the most widely installed gaming GPU is currently the Fermi-vintage GTX 550Ti. Although a great card it’s really had its time, and the GTX 750Ti is designed as a direct replacement both for OEMs and consumers.
GTX 750Ti is the first generation of Maxwell, and due to critical technical improvements offers double the performance per watt of Kepler and 135% better performance per core. That’s an unheard of jump for an architecture on the same process node, and indicates a major redesign of the core to enable better utilisation of available resources.
NVIDIA have been reticent in describing exactly how this perf/watt improvement has been achieved but it is at least partially down to the redesign of the SMX to SMM and shifting some part of the thread scheduling off the GPU onto the CPU. They stress however that the performance impact on the CPU of doing it in software is negligible (fractions of a percent), whilst the efficiency gains are significant. Even in CPU-bound scenarios you would struggle to tell the difference.
Power efficiency isn’t really a selling point for an upgrade however, which is why the new GPU targets the 550Ti. Compared to the older GPU the new cards will offer 220% better performance within a 60W envelope, compared with 116W. This means that in-place upgrades can be made without commensurate upgrades to the GPU.
As a mainstream part, it’s important that the 750Ti can keep up with standard gaming resolutions of 1080p. Once again we refer you to the review for performance numbers, but NVIDIA believe that a good experience is assured at appropriate image quality settings found through GeForce Experience.
For OEMs and system integrators however the improved power efficiency is extremely important. For one it allows them to run more affordable PSUs, an important consideration in keeping down overall system cost. Furthermore the cooling solution required can be far less elaborate, reducing noise generated overall. As a consequence the GTX 750Ti stack up very well in a small form factor designs for the livingroom space.
Reference GTX 750Ti designs come with no 6-pin PCI-E connector, instead drawing all it requires from the bus. Partner designs may well include the connector to unlock additional overclocking potential, either as a value-added factory feature or allowing end-users greater flexibility. Whilst reference clocks are pegged at a little over 1GHz, some tests have indicated a possible overclock of 1.2 – 1.3GHz within reasonable thermal boundaries.
The card replaces NVIDIA’s GTX 650Ti and 650Ti Boost in the product stack, but remains bookended by the GTX 650 and 660. They acknowledge that this may cause some confusion, but anticipate that it would be less than introducing the 800-series with a mid-range part; furthermore, they feel that pricing and performance reviews will serve to reduce any confusion. That’s certainly a shot across the bows of AMD’s R7 & R9 range.
Perhaps the only feature the GTX 750-class lacks compared to the card it replaces is SLI, a technology which has been sacrificed on the altar of power efficiency. Admittedly that’s not too much of a hardship at this price point and enough of the infrastructure has been retained for the cards to support G-SYNC and GPU Boost 2, so for the vast majority of potential users the impact will be small.
You should expect to pay around £114.99 for the GTX 750Ti, putting it squarely in the budget of many gamers outside of the hardcore crowd looking for an upgrade it a minimal cost.
NVIDIA Performance Comparison Charts, GTX 750Ti vs R7 260X
AMDs chief competitor at the price point is the Radeon R7 260X (the R7 265 not quite being available yet). Whilst performance and price may be comparable within a margin of error, the Radeon GPU is a 115W part. For many end-users the difference will be imperceptible, but for OEMs and system integrators generating designs to ship in the thousands and hundreds of thousands the difference is pronounced.
The battleground for the new cards will be upcoming Steam Machines and similar small form factor designs. NVIDIA may well have the edge in terms of power consumption and noise, but that will also be down to the respective cooling solutions AIB partners ship. G-SYNC may not be a factor just yet, especially in the living room, and in the UK at least SHIELD isn’t a relevant technology. AMD have Mantle and TrueAudio for selected titles, but it seems likely that NVIDIA will increasingly leverage Gameworks for vendor-specific performance tweaks.
In terms of the raw specs it’s difficult to see how AMD can compete except by reducing price and cracking on with an architectural update of their own. In the meantime NVIDIA no doubt want to make hay whilst the sun shines, and press their dominance if an when Steam Machines hit the mainstream.
GeForce GTX 750 – The <£100 Sweetspot
A small notch below the GTX 750Ti comes the vanilla GTX 750, a more affordable card pitched at or around the £90 price point. Once again superseding the GTX 650 Ti, for now it still coexists in the same product stack as the entry-level GTX 650.
In much the same was that the GTX 750Ti is intended to replace legacy parts so is the GTX 750. In this case it will serve as a direct update to systems running the GTS 450, which not coincidentally is the GPU with 2nd-highest share on the Steam Hardware Survey. Offering three times the performance within a similar 55W TDP, it’s an extremely important part for budget system builders.
NVIDIA are boasting of a card capable of 58fps in Skyrim on High settings within an extremely small package, making it excellent for casual gaming from a wide selection of Steam Indie titles or older AAA games. In terms of performance and features the card aims to surpass the AMD Radeon R7 260, which is effectively the Cape Verdi-based HD 7770. That’s despite having only 80% the number of CUDA cores, and speaks to the scalability of Maxwell.
There are no reference designs of the GTX 750 card, whilst will only be sold through NVIDIA’s partners. Once again, although a 6-pin power connector isn’t necessary, some designs may use one for additional clock speed flexibility.
In Summary – A Vital Addition To The Stack
Although Maxwell GPUs probably won’t make many headlines in the enthusiast section of the market their impact should be pronounced. These are the cards which are typically shipped in mass quantities and are essential for major system integrators keen to push good mainstream performance on a budget; in this respect their dreadfully important to the industry as a whole. It they claim a level of performance they had best deliver, as hundreds of thousands of users are likely to have their hands on one in a relatively short space of time.
Two other major releases will also tend to push these cards into consumer hands. The first is Intel’s Haswell CPU refresh, due May (or April according to some reports). New CPU releases always provide a small bump, and so something new rather than rehashed from NVIDIA would push consumers minds towards Maxwell GPUs. AMD cannot offer this, instead relying on the now ageing first generation GCN cards.
The second the Valves Steam Machine initiative. Low power and just as importantly low noise gaming is essential for the various vendor designs to take off, and the GTX 750-class is poised to take advantage of this need in a way that the R7 260X & 265 don’t appear to be able to match. Mobile variants used in integrated Steam Box’s from ZOTAC, GIGABYTE and Scan Computers will also be critical for the very small form factor designs (< mini-ITX). NVIDIA are poised to take advantage of a surge in demand, if it appears.
From our perspective, we’re looking forward to more Maxwell designs. Boosting performance per watt in the way they have done, without moving to 20nm, is close to astonishing and bodes well for enthusiast designs in the future.
