Maxwell’s Feature Set: Kepler Refined

To start our look at the Maxwell architecture, we’ll start with a look at the feature set, as this will be the shorter and easier subject to begin with.

In short, Maxwell only offers a handful of new features compared to Kepler. Kepler itself was a natural evolution of Fermi, further building on NVIDIA’s SM design and Direct3D 11 functionality. Maxwell in turn is a smaller evolution yet.

From a graphics/gaming perspective there will not be any changes. Maxwell remains a Direct3D 11.0 compliant design, supporting the base 11.0 functionality along with many (but not all) of the features required for Direct3D 11.1 and 11.2. NVIDIA as a whole has not professed much of an interest in being 11.1/11.2 compliant – they weren’t in a rush on 10.1 either – so this didn’t come as a great surprise to us. Nevertheless it is unfortunate, as NVIDIA carries enough market share that their support (or lack thereof) for a feature is often the deciding factor whether it’s used. Developers can still use cap bits to access the individual features of D3D 11.1/11.2 that Maxwell does support, but we will not be seeing 11.1 or 11.2 becoming a baseline for PC gaming hardware this year.

On the other hand this means that for the purposes of the GeForce family the GTX 750 series will fit in nicely into the current stack, despite the architectural differences. As a consumer perspective is still analogous to a graphics perspective, Maxwell does not have any features that will explicitly set it apart from Kepler. All 700 series parts will support the same features, even NVIDIA ecosystem features such as GameWorks, NVENC, and G-Sync, so Maxwell is fully aligned with Kepler in that respect.

At a lower level the feature set has only changed to a slightly greater degree. I/O functionality is identical to Kepler, with 4 display controllers backing NVIDIA’s capabilities. HDMI 1.4 and DisplayPort 1.2 functionality join the usual DVI support, with Maxwell being a bit early to support any next generation display connectivity standards.

Video Encode & Decode

Meanwhile turning our gaze towards video encoding and decoding, we find one of the few areas that has received a feature upgrade on Maxwell. NVENC, NVIDIA’s video encoder, has received an explicit performance boost. NVIDIA tells us that Maxwell’s NVENC should be 1.5x-2x faster than Kepler’s NVENC, or in absolute terms capable of encoding speeds 6x-8x faster than real time.

For the purposes of the GTX 750 series, the impact of this upgrade will heavy depend on how NVENC is being leveraged. For real time applications such as ShadowPlay and GameStream, which by the very definition can’t operate faster than real time, the benefit will primarily be a reduction in encoding latency by upwards of several milliseconds. For offline video transcoding using utilities such as Cyberlink’s MediaEspresso, the greater throughput should directly translate into faster transcoding.

The bigger impact of this will be felt in mobile and server applications, when GM107 makes its introduction in those product lines. In the case of mobile usage the greater performance of Maxwell’s NVENC block directly corresponds with lower power usage, which will reduce the energy costs of using it when operating off of a battery. Meanwhile in server applications the greater performance will allow a sliding scale of latency reductions and an increase in the number of client sessions being streamed off of a single GPU, which for NVIDIA’s purposes means they will get to increase the client density of their GRID products.

Speaking of video, decoding is also receiving a bit of a lift. Maxwell’s VP video decode block won’t feature full H.265 (HEVC) support, but NVIDIA is telling us that they will offer partial hardware acceleration, relying on a mix of software and hardware to decode H.265. We had been hoping for full hardware support on Maxwell, but it looks like it’s a bit premature for that in a discrete GPU. The downside to this is that the long upgrade cycle for video cards – many users are averaging 4 years these days – means there’s a good chance that GTX 750 owners will still be on their GTX 750 cards when H.265 content starts arriving in force, so it will be interesting to see just how much of the process NVIDIA can offload onto their hardware as it stands.

H.265 aside, video decoding overall is getting faster and lower power. NVIDIA tells us that decoding is getting a 8x-10x performance boost due to the implementation of a local decoder cache and an increase in memory efficiency for video decoding. As for power consumption, combined with the aforementioned performance gains, NVIDIA has implemented a new power state called “GC5” specifically for low usage tasks such as video playback. Unfortunately NVIDIA isn’t telling us much about how GC5 works, but as we’ll see in our benchmarks there is a small but distinct improvement in power consumption in the video decode process.

Introducing Maxwell Maxwell: Designed For Energy Efficiency
POST A COMMENT

181 Comments

View All Comments

  • Harry Lloyd - Tuesday, February 18, 2014 - link

    20 nm Maxwell will be epic. Gimme. Reply
  • TheinsanegamerN - Tuesday, February 18, 2014 - link

    Imagine. OCed Geforce 690 level performance, out of a single chip, with 8 GB of RAM on a 512 bit bus, pulling the same amount of power as a geforce 770. One can dream.... Reply
  • ddriver - Tuesday, February 18, 2014 - link

    LOL, epic? Crippling FP64 performance further from 1/24 to 1/32 - looks like yet another nvidia architecture I'll be skipping due to abysmal compute performance per $ ratio... Reply
  • JDG1980 - Tuesday, February 18, 2014 - link

    This card is designed for gaming and HTPC. Only a tiny fraction of users need FP64. Reply
  • nathanddrews - Tuesday, February 18, 2014 - link

    So I guess we'll have to wait for the 750TIB before we can see SLI benchmarks. Two of these would be within reach of 770 while using considerably less power. Hypothetically, that is. Reply
  • ddriver - Tuesday, February 18, 2014 - link

    You do realize the high end GPUs on the same architecture will have the same limitation? Reply
  • Morawka - Tuesday, February 18, 2014 - link

    I thought the higher end Maxwell cards will have Denver/aRM cores on the PCB as well. Reply
  • Mr Perfect - Wednesday, February 19, 2014 - link

    It might be a software/firmware limitation though. From what the compute enthusiasts have said, the only difference between the Titan's full compute and 780Ti's cut down compute is firmware based. They've got the same chip underneath, and some people hack their 780s for full compute. They're probably doing the same thing with the Maxwell stack. Reply
  • chrnochime - Wednesday, February 19, 2014 - link

    Got link for the hack? Sounds interesting. Reply
  • Mr Perfect - Thursday, February 20, 2014 - link

    I don't myself, but if you're interested look up IvanIvanovich over at bit-tech.net. He was talking about vbios mods and resistor replacement tweaks that can do that. Reply

Log in

Don't have an account? Sign up now