Maxwell adds desktop GPU technologies to mobile devices

NVIDIA Corp. (NVDA) this week pulled the wraps off its second major mobile GPU family release of the year, introducing a pair of mobile graphics chips based on second-generation Maxwell parts.

I. The Pace of Graphics Progress: Leaps and Iterations

Over the last half decade NVIDIA has managed to keep gamers entertained with two die shrinks three major microarchitecture introductions.  The last few years have seen:
  • Oct. 2009 - Move to 40 nm
    • Die shrink occurs on low-end and mid-range GeForce 200 Series parts
    • Fabricated by Taiwan Semiconductor Manufacturing Comp., Ltd. (TPE:2330) (TSMC)
    • Shrunk dies are Tesla microarchitecture
  • March 2010 - Launch of the GeForce 400 Series
    • Introduces the Fermi microarchitecture
    • Built on a 40 nanometer (nm) process from Taiwan Semiconductor Manufacturing Comp., Ltd. (TPE:2330) (TSMC)
    • Part name: "GF1XX"
  • March 2012 - Launch of the GeForce 600 Series
    • Introduces the Kepler microarchitecture
    • Also introduces a die shrink to 28 nm from TSMC
    • Part name: "GK1XX"
  • Feb. 2014 - New GeForce 700 Series models
    • Introduce the Maxwell microarchitecture
    • Parts slots into the preexisting Kepler-based GeForce 700 Series line
    • Built on mature 28 nm TSMC process
    • Part name: "GK1XX"
Maxwell chip
Maxwell is NVIDIA's third major microarchitecture release in the past half-decade.

Looking back, Kepler was perhaps the biggest launch, given that introduced both a new microarchitecture and a die shrink at the same time.  That said, NVIDIA has also introduced several generations which don't bring a wholly new microarchitecture, but still manage to differentiate themselves in other ways.  To do this NVIDIA has used a lot of different strategies including gradually increasing its CUDA core count and clock speeds, as well as specialized rendering and encoding technologies.

Despite its innovation, NVIDIA deserves some admonishment for its idiosyncratic naming schemes, which are perhaps somewhat of a tradition for the graphics veteran.  After deciding to squeeze early Maxwell parts (based on the GM107 part) into mid-range GeForce 700 Series cards for the desktop in February of this year, NVIDIA opted for a different strategy a month later on the mobile front, releasing the GeForce 800M series, whose high end models were Kepler-based (GK104) and whose mid-range models were Maxwell-based (GM107/108).


Despite the naming confusion, the entrance of Maxwell in Q1 2014 did offer some potential performance leaps.  CUDA core count was up in like-named SKUs by between 50 and 66 percent.  The improved second generation "Boost" technology (first introduced with Kepler) was also included.  The thermal design power (TDP) also dipped between 15 to 25 percent, reflecting the gains of the new, more efficient microarchitecture.

II. Second-Gen Maxwell Cometh

After a half year of making Maxwell parts, NVIDIA has some tweaks up its sleeve and that brings us to the GeForce 900/900M Series.  

Sanity has finally returned to the naming scheme, as both desktop and laptop are now united under a single banner for the latest and greats GPUs.  The desktop parts -- based on the GM204 core debutted in the middle of last month.

GeForce GTX 970
NVIDIA GeForce GTX 970

GeForce GTX 980
NVIDIA GeForce GTX 980

You'll see NVIDIA referring to GM2XX chips (including those in the titular GeForce 900M release) as "second generation Maxwell".  The marketing lingo is a bit misleading, as GM204 is likely nearly identical to GM107 on an architectural basis.  That said, the tuned up Maxwell parts do add a number of new rendering technologies, including:
  • Dynamic Super Resolution [PDF]
    • Downsampling from 4K to 1080p
  • Third Generation Delta Color Compression [PDF]
    • Allows greater memory efficiency via lossless compression of outgoing pixel color data on the bus
  • Nvidia VXGI (Real-Time-Voxel-Global Illumination)
    • Voxel = opacity geometry
    • More precisely voxels are cube (six-sided) volume elements which make up space filling models (think LEGO models) of an object.  They carry an opacity value,
    • Basically, voxel lighting is a major advance over direct lighting in gaming realism.  It uses voxels to calculate reflective lighting, by turning illuminated objects into weak lights and then using those with the voxels to determined secondary illumination.  The results are then combined with the direct lighting for a finished scene.
    • This is a poor man's alternative to ray tracing, which remains out of reach for real time gaming
    • Still too slow, without dedicated hardware acceleration (which Maxwell adds)
  • VR Direct
    • Tunes latency to minimize motion sickness in near-eye displays found inside the new wave of VR devices like Oculus VR
  • Multi-Projection Acceleration [PDF]
    • Works hand in hand with other techniques like VXGI by allowing parallel projection of geometry to multiple surfaces
  • Multi-Frame Sampled Anti-Aliasing (MFAA) and Multi-Pixel Programming Sampling
    • Anti-aliasing technique performed by rendering several locations within a pixel and combining the results
    • Uses sampling patterns stored in read only memory
    • Patterns are spatially and temporally alternated to maximize quality
    • Delivers aliasing at half the computational cost compared to the older multisample anti-aliasing (MSAA).
    • Note that this technology also eliminates the older coverage sampling anti-aliasing (CSAA), which provided similar computation savings in earlier GeForce models
    • Some have complained that the decision to drop antialiasing modes using the older CSAA and MSAA techniques makes Maxwell chips technically not fully compatible with the DirectX 11 spec, which mandates "pure" MSAA 8x
And the GM2XX Maxwell parts also add support for some new hardware/connectivity options including:
  • 4K hardware accelerated encoding via the NVIDIA Encoder (NVENC) engine
    • Encodes at H.264 1440p60 and 4Kp60
    • First generation Maxwell only supported H.264 1080p60 encoding
  • HDMI 2.0 support
The GeForce GTX 980 was an interesting card, as it cut the CUDA core count by 11 percent from the Kepler-based GeForce GTX 780 (reducing the core count to 2048), but compensates by shaving $100 off the price (taking it to $549 USD), cutting the TDP by a third, and by increasing the shader clock by nearly a third.

NVIDIA GeForce GTX 980

Its companion, the GeForce GTX 970 uses a differents strategy over its predecessor, increasing the core count by roughly 8 percent (to 1664), increasing the boost clock by a similar amount, and by trimming $70 USD off the price (taking it to $329 USD).

III. The Laptop Launch of GM204 and What's Next

With last month's desktop release it was clear that the second generation Maxwell brings a lot of new features to the table, along with some more minor obvious performance bumps.

Remember, the GTX 770 vs. GTX 980 comparison is more stark as you're comparing Kepler to second-generation Maxwell parts (remember, first generation Maxwell was only the midrange of the desktop GeForce 700 Series).

The same is true with this week's release of the GeForce GTX 970M and GeForce GTX 980M.  

GeForce GTX 970M and 980M

While the CUDA core count is virtually unchanged (and power even increases in the 980M), nonetheless you can expect some mild performance gains from going from Kepler to Maxwell.  At the same time you get access to the special rendering technologies like VXGI lighting, which NVIDIA has rolled into the second generation Maxwell parts.

NVIDIA is also claiming that the TDP drop with the second-generation Maxwell parts is sufficient to game for two hours on a battery charge for an average gaming laptop.

GeForce GTX 980M

You'll find the new mobile GPU duo onboard select models from Micro-Star Int'l Comp., Ltd. (TPE:2377) (the MSI GT72, GS70 and GS60 models), the ASUSTek Computer, Inc. (TPE:2357) G751, and a pair of models from Gigabyte Tech. Comp., Ltd. (TPE:2376) (the Gigabyte Aurus X7 and P35).  Adding in models from boutique vendors AVADirect. MainGear, and OriginPC, there's "over a dozen SKUs... now available" with the chips inside, according to NVIDIA.

If you're looking for a more serious advance you'll have to wait for next year's third generation iteration of Maxwell, which is expected to move to TSMC's 20 nm process next year (which is currently somewhat tied up fulfilling orders of Apple, Inc.'s (AAPL) A8 system-on-a-chip (SoC)).  That should bring some nice power efficiency gains.

And if you're waiting for a wholly new architecture, be prepared to wait nearly a year as NVIDIA doesn't look to be breaking from its 2-year cycle for GPU architectures.  It's already announced that Maxwell's successor, Pascal, will drop in 2016.

GeForce GTX 980
Next year's 20 nm, third-generation Maxwell and 2016's Pascal will have to wait -- for now we have second-generation 28 nm Maxwell.

A final note is that Maxwell marks the first time NVIDIA accomplished a nearly simultaneous release of mobile and desktop GPUs based on the same part (the GM204).  
GeForce 900M -- closing the gap
NVIDIA has finally closed the gap between its PC and mobile parts.

That's good news for laptop gamers, and perhaps yet another sign of the waning role of the desktop client in the consumer market -- even amongst the gamer crowd.

Source: NVIDIA [blog]

"The Space Elevator will be built about 50 years after everyone stops laughing" -- Sir Arthur C. Clarke
Related Articles

Copyright 2017 DailyTech LLC. - RSS Feed | Advertise | About Us | Ethics | FAQ | Terms, Conditions & Privacy Information | Kristopher Kubicki