Nvidia GeForce GTX 980 Quick reference guide

Whitepaper
NVIDIA GeForce GTX 980
Featuring Maxwell, The Most Advanced GPU Ever Made.
V1.1

GeForce GTX 980 Whitepaper
Table of Contents
Introduction .................................................................................................................................................. 3
Extraordinary Gaming Performance for the Latest Displays .................................................................... 3
Incredible Energy Efficiency...................................................................................................................... 4
Dramatic Leap Forward In Lighting with VXGI .......................................................................................... 5
GM204 Hardware Architecture In-Depth ..................................................................................................... 6
Maxwell Streaming Multiprocessor.......................................................................................................... 8
PolyMorph Engine 3.0............................................................................................................................... 9
GM204 Memory Subsystem ...................................................................................................................10
New Display and Video Engines..............................................................................................................11
Maxwell: Enabling The Next Frontier in PC Graphics .................................................................................13
Hardware Acceleration for VXGI –Multi-Projection and Conservative Raster ......................................21
Tiled Resources .......................................................................................................................................23
Raster Ordered View............................................................................................................................... 24
DirectX 12................................................................................................................................................ 25
Advancing the State-Of-The-Art in Image Quality ......................................................................................27
Dynamic Super Resolution......................................................................................................................29
Conclusion...................................................................................................................................................31
LEGAL NOTICE .............................................................................................................................................32

GeForce GTX 980 Whitepaper
INTRODUCTION
3
Introduction
In our 21-year quest to bring the most realistic 3D graphics to gamers, NVIDIA has introduced a number
of innovations. With its hardware transform and lighting engine, NVIDIA’s GeForce 256 ushered in the
era of the GPU in 1999, bringing T&L support to consumer graphics for the first time. In late 2006 with
the G80 GPU and the GeForce 8800 GTX graphics board, CUDA and the formalization of GPGPU
computing as we know it today was first brought to the world. More recently, in 2010 we launched the
GeForce GTX 480—our first GPU built using the Fermi graphics architecture. GeForce GTX 480
incorporated fifteen parallel tessellation units, enabling dramatic speedups in geometry processing
compared to prior GPUs. With great strides in architectural efficiency, Kepler GPUs delivered
significantly improved performance and power efficiency when introduced in 2012.
Earlier this year we launched the GeForce GTX 750 Ti, our first GPU from the Maxwell architectural
family. Maxwell GPUs were designed from the ground up for extreme power efficiency and exceptional
performance per watt consumed. In many DX11 applications, the GTX 750 Ti is capable of matching or
even beating the performance of our once flagship GeForce GTX 480, while consuming only a fourth the
power. Thanks to its remarkable power efficiency, our first generation Maxwell GPUs were ideal for use
in power-limited environments like notebooks and small form factor PCs, in addition to mainstream
desktops.
NVIDIA’s latest GPU, GM204, is the first to use the full realization of our 10th generation GPU
architecture, Maxwell. Our design goals for GM204 were to deliver:
Extraordinary Gaming Performance for the Latest Displays
Incredible Energy Efficiency
Dramatic Leap Forward In Lighting With VXGI
Extraordinary Gaming Performance for the Latest Displays
After years of little or no progress in the evolution of PC displays, the pace of new technology
introductions has dramatically increased. Screen resolutions are growing, with the first 4K displays for
gamers hitting the market a little over a year ago. While 4K was initially expensive, thanks to rapidly
falling prices, 4K adoption is now beginning to take off. In order to drive all these pixels, GM2xx GPUs
were designed to deliver more performance at the highest resolutions, and also support new display
tech for 4K like HDMI 2.0. In addition, with Dynamic Super Resolution, Maxwell introduces the ability to
significantly enhance display quality on sub-4K displays by rendering at higher resolution and then
applying an advanced resizing filter that leverages the high resolution information.
Virtual Reality displays are another area that is likely to drive a dramatic increase in rendering rate
requirements. Two displays for two eyes doubles the rendering workload, and ever higher resolution
and refresh rate capability will be crucial to eliminating rendering artifacts that break the illusion of
reality for the VR user.

GeForce GTX 980 Whitepaper
INTRODUCTION
4
Incredible Energy Efficiency
All GeForce GPUs are made to deliver performance leadership in their respective segment. But PC
gamers also expect efficiency—not noisy fans, excessive heat, and over-taxed power supplies. The
desktop PC gaming market is seeing increasing popularity of smaller and thinner form factors, and the
laptop PC gaming market has grown explosively in the past few years.
The Maxwell architecture was designed to provide an extraordinary leap in power efficiency and deliver
unrivaled performance while simultaneously reducing power consumption from the previous
generation. With a combination of advances originally developed for Tegra K1, new architectural
approaches seen first in the GeForce GTX 750 Ti, and further refinement for GM204, Maxwell now
delivers 2x the performance per watt compared to Kepler products.
Figure 1: GM204 delivers 2X higher perf/watt compared to Kepler

GeForce GTX 980 Whitepaper
INTRODUCTION
5
Dramatic Leap Forward In Lighting with VXGI
Lighting is the biggest hurdle in computer graphics that must be overcome in order to achieve
photorealism. Global illumination gets us significantly closer to achieving this goal. Global illumination is
often used in the movie industry to produce CG scenes that look stunning, but rendering truly dynamic
global illumination in real time on the GPU has been impractical until now.
NVIDIA Voxel Global Illumination (VXGI) is the next big leap forward in lighting. VXGI is based on the
concept pioneered by NVIDIA researcher Cyril Crassin in 2011, of using a 3D data structure (“voxels”) to
capture coverage and lighting information at every point in the scene. This data structure can then be
traced during the final rendering stage to accurately determine the effect of light bouncing around in
the scene.
Cyril’s original implementation relied on voxels that were stored in an octree structure. While it was able
to run successfully on a GeForce GTX 680, it had limitations. We’ve spent the last three years developing
an implementation that can be accelerated natively by the GPU, as well as improving the algorithm. The
result of that effort is VXGI.
VXGI provides dramatically improved performance for global illumination. While VXGI’s software
algorithm will run on all GPUs, the performance benefits of VXGI hardware acceleration will only be
available on our latest Maxwell GPUs.
The first graphics cards to ship with our new GM204 GPU are the GeForce GTX 980 and GeForce GTX
970. For simplicity purposes we’ll be focusing solely on the GeForce GTX 980 in this document. With
2048 CUDA Cores and 4GB of GDDR5 memory, GM204 is the fastest GPU in the world, yet with a TDP of
just 165W, it’s also the most efficient.

GeForce GTX 980 Whitepaper
GM204 HARDWARE ARCHITECTURE
IN-DEPTH
6
GM204 Hardware Architecture In-Depth
Figure 2: GM204 Full-chip block diagram
Like Fermi and Kepler, GM204 is composed of an array of Graphics Processing Clusters (GPCs), Streaming
Multiprocessors (SMs), and memory controllers. GM204 consists of four GPCs, 16 Maxwell SMs (SMM),
and four memory controllers. GeForce GTX 980 uses the full complement of these architectural
components (if you are not well versed in these structures, we suggest you first read the Kepler and
Fermi whitepapers).
Another version of the chip, with 13 SMs, will ship concurrently and be called GeForce GTX 970. In the
future we plan to offer additional products based on GM204 that will ship with different combinations
of GPCs, SMs, and memory controllers to address various segments of the graphics market.

GeForce GTX 980 Whitepaper
GM204 HARDWARE ARCHITECTURE
IN-DEPTH
7
In GeForce GTX 980, each GPC ships with a dedicated raster engine and four SMMs. Each SMM has 128
CUDA cores, a PolyMorph Engine, and eight texture units. With 16 SMMs, the GeForce GTX 980 ships
with a total of 2048 CUDA cores and 128 texture units.
The GeForce GTX 980 features four 64-bit memory controllers (256-bit total). Tied to each memory
controller are 16 ROP units and 512KB of L2 cache. The full chip ships with a total of 64 ROPs and
2048KB of L2 cache (this compared to 32 ROPs and 512K L2 on GK104).
The following table provides a high-level comparison of Maxwell vs. our previous-generation GK104
GPU:
GPU
GeForce GTX 680 (Kepler)
GeForce GTX 980 (Maxwell)
SMs
8
16
CUDA Cores
1536
2048
Base Clock
1006 MHz
1126 MHz
GPU Boost Clock
1058 MHz
1216 MHz
GFLOPs
3090
46121
Texture Units
128
128
Texel fill-rate
128.8 Gigatexels/sec
144.1 Gigatexels/sec
Memory Clock
6000 MHz
7000 MHz
Memory Bandwidth
192 GB/sec
224 GB/sec
ROPs
32
64
L2 Cache Size
512KB
2048KB
TDP
195 Watts
165 Watts
Transistors
3.54 billion
5.2 billion
Die Size
294 mm²
398 mm²
Manufacturing Process
28-nm
28-nm
The GeForce GTX 980 has double the SMs compared to the GK104 GPU used in the GeForce GTX 680
released two years ago. Because of the changes implemented in GTX 980’s new Maxwell SM, we were
able to integrate 2x more SMs without doubling the die size. With each SM also containing its own
dedicated PolyMorph Engine, GeForce GTX 980 also has twice the number of geometry units as its direct
predecessor. We’ll be discussing more details on the new SM design in the next section of the
whitepaper.
Based on efficiency and workload analysis, and math vs. texture processing requirements of modern
games, NVIDIA engineers determined that eight texture units per SMM is the best architectural balance
for Maxwell; therefore, the total number of texture units is the same as Kepler, 128. However, thanks to
GeForce GTX 980’s higher clocks, texture fill rate improves by 12% from one generation to the next. To
improve performance in high AA/high resolution gaming scenarios, we doubled the number of ROPs
1
The GFLOPS and texel fill rates in this chart are based on GPU Base Clock

GeForce GTX 980 Whitepaper
GM204 HARDWARE ARCHITECTURE
IN-DEPTH
8
from 32 to 64. Again, thanks to the added benefit of higher clocks, pixel fill-rate is actually more than
double that of GTX 680: 72 Gpixels/sec for GTX 980 versus 32.2 Gpixels/sec for GTX 680.
The memory subsystem has also been significantly revamped. GTX 980’s memory clock is over 15%
higher than GTX 680, and GM204’s cache is larger and more efficient than Kepler’s design, reducing the
number of memory requests that have to be made to DRAM. Improvements in our implementation of
memory compression provide a further benefit in reducing DRAM traffic—effectively amplifying the raw
DRAM bandwidth in the system.
Maxwell Streaming Multiprocessor
The SM is the heart of our GPUs. Almost
every operation flows through the SM at
some point in the rendering pipeline.
Maxwell GPUs feature a new SM that’s
been designed to provide dramatically
improved performance per watt than prior
GeForce GPUs.
Compared to GPUs based on our Kepler
architecture, Maxwell’s new SMM design
has been reconfigured to improve
efficiency. Each SMM contains four warp
schedulers, and each warp scheduler is
capable of dispatching two instructions per
warp every clock. Compared to Kepler’s
scheduling logic, we’ve integrated a
number of improvements in the scheduler
to further reduce redundant re-
computation of scheduling decisions,
improving energy efficiency. We’ve also
integrated a completely new datapath
organization. Whereas Kepler’s SM shipped
with 192 CUDA Cores—a non-power-of-two
organization—the Maxwell SMM is
partitioned into four distinct 32-CUDA core
processing blocks (128 CUDA cores total
per SM), each with its own dedicated
resources for scheduling and instruction
buffering. This new configuration in
Maxwell aligns with warp size, making it
easier to utilize efficiently and saving area
Figure 3: GM204 SMM Diagram (GM204 also features 4 DP units per
SMM, which are not depicted on this diagram)

GeForce GTX 980 Whitepaper
GM204 HARDWARE ARCHITECTURE
IN-DEPTH
9
and power that had to be spent to manage data transfer in the more complex datapath organization
used by Kepler.
Compared to Kepler, the SMM’s memory hierarchy has also changed. Rather than implementing a
combined shared memory/L1 cache block as in Kepler SMX, Maxwell SMM units feature a 96KB
dedicated shared memory, while the L1 caching function has been moved to be shared with the texture
caching function.
As a result of these changes, each Maxwell CUDA core is able to deliver roughly 1.4x more performance
per core compared to a Kepler CUDA core, and 2x the performance per watt. At the SM level, with 33%
fewer total cores per SM, but 1.4x performance per core, each Maxwell SMM can deliver total per-SM
performance similar to Kepler’s SMX, and the area savings from this more efficient architecture enabled
us to then double up the total SM count, compared to GK104.
PolyMorph Engine 3.0
Tessellation was one of DirectX 11’s key features and will play a bigger role in the future as the next
generation of games are designed to use more tessellation. With the addition of more SMs in GM204,
GTX 980 also benefits from 2x the Polymorph Engines, compared to GTX 680. As a result, performance
on geometry heavy workloads is roughly doubled, and due to architectural improvements within the PE,
can achieve up to 3x performance improvement with high tessellation expansion factors.
0.0
0.5
1.0
1.5
2.0
2.5
3.0
3.5
4.0
1357911 13 15 17 19 21 23 25 27 29 31
Performance
Expansion Factor
Microsoft SubD11 SDK Test
GTX 980
GTX 680

GeForce GTX 980 Whitepaper
GM204 HARDWARE ARCHITECTURE
IN-DEPTH
10
GM204 Memory Subsystem
In GM204 one ROP partition contains 16 ROP units (compared to eight ROP units per partition in Kepler);
each ROP can process a single color sample. With four ROP partitions, a full GM204 has 64 ROPs, twice
that of its predecessor, GK104, dramatically improving ROP throughput.
GM204 has a 256-bit memory interface with 7Gbps GDDR5 memory, the fastest in the industry. GM204
also features a unified 2048KB L2 cache that is shared across the GPU. In addition, GM204 has made
significant enhancements to our memory compression implementation.
To reduce DRAM bandwidth demands, NVIDIA GPUs make use of lossless compression techniques as
data is written out to memory. The bandwidth savings from this compression is realized a second time
when clients such as the Texture Unit later read the data. As illustrated in the preceding figure, our
compression engine has multiple layers of compression algorithms. Any block going out to memory will
first be examined to see if 4x2 pixel regions within the block are constant, in which case the data will be
compressed 8:1 (i.e., from 256B to 32B of data, for 32b color). If that fails, but 2x2 pixel regions are
constant, we will compress the data 4:1.
These modes are very effective for AA surfaces, but less so for 1xAA rendering. Therefore, starting in
Fermi we also implemented support for a “delta color compression” mode. In this mode, we calculate
the difference between each pixel in the block and its neighbor, and then try to pack these different
values together using the minimum number of bits. For example if pixel A’s red value is 253 (8 bits) and
pixel B’s red value is 250 (also 8 bits), the difference is 3, which can be represented in only 2 bits.

GeForce GTX 980 Whitepaper
GM204 HARDWARE ARCHITECTURE
IN-DEPTH
11
Finally, if the block cannot be compressed in any of these modes, then the GPU will write out data
uncompressed, preserving the lossless rendering requirement.
However, the effectiveness of delta color compression depends on the specifics of which pixel ordering
is chosen for the delta color calculation. Maxwell contains our third generation of delta color
compression, which improves effectiveness by offering more choices of delta calculation to the
compressor.
Thanks to the improvements in caching and compression in Maxwell, the GPU is able to significantly
reduce the number of bytes that have to be fetched from memory per frame. In tests with a variety of
games, Maxwell uses roughly 25% fewer bytes per frame compared to Kepler. This means that from the
perspective of the GPU core, a Kepler-style memory system running at 9.3Gbps would provide effective
bandwidth similar to the bandwidth that Maxwell’s enhanced memory system provides.
New Display and Video Engines
As the rapid adoption rate of 4K displays shows, consumer demand for high resolution devices has
intensified this year. To meet this demand, all Maxwell GPUs ship with a display engine capable of
supporting resolutions up to 5K. In addition, up to four 4K MST displays can be driven by the GPU,
compared to a maximum of two 4K MST displays in Kepler.
GeForce GTX 980 is also the world’s first GPU to support HDMI 2.0. Whereas the previous generation,
HDMI 1.4, can support 4K display at 30Hz for “444” RGB pixels, and 60Hz for “420” YUV pixels, with
HDMI 2.0, the GPU can now drive full-resolution “444” RGB pixels at 4K resolution at 60Hz.

GeForce GTX 980 Whitepaper
GM204 HARDWARE ARCHITECTURE
IN-DEPTH
12
The GeForce GTX 980 reference board design ships with three DisplayPort 1.2 connectors, one HDMI 2.0
connector, and one dual-link DVI connector. Up to four display heads can be driven simultaneously from
one card.
Kepler
Maxwell
Number of Heads
4
4
Number of Links
6
7
Max Pixel Clock
540 MHz
1045 MHz
Max Resolution
3840 x 2160 @
60Hz
5120 x 3200 @ 60Hz
Digital Protocols
LVDS,
TMDS/HDMI 1.4,
DP 1.2
LVDS, TMDS/HDMI 2.0,
DP 1.2, eDP 1.4
When combined with a G-SYNC display, the GeForce GTX 980 delivers a gaming experience that’s free of
the distracting screen tearing that currently plagues gaming when Vsync is disabled. G-SYNC also
eliminates display-subsystem-generated stutter and reduces input lag that gamers put up with today.
Utilizing DisplayPort, the GeForce GTX 980 can drive up to three G-SYNC displays in Surround.
GM2xx Maxwell also ships with an enhanced NVENC encoder that adds support for H.265 (also known
has HEVC) encoding. H.265 compression offers significant bandwidth savings versus H.264 at the same
quality level, which will help enable the next generation of streamed gaming experiences within the
home or powered by the cloud.
In addition, Maxwell’s video encoder improves H.264 video encode throughput by 2.5x over Kepler,
enabling it to encode 4K video at 60 fps. We’re supporting this feature today with ShadowPlay.

GeForce GTX 980 Whitepaper
MAXWELL: ENABLING THE NEXT
FRONTIER IN PC GRAPHICS
13
Maxwell: Enabling The Next Frontier in PC Graphics
Correct modelling of lighting is the most computationally difficult problem in graphics, and with
Maxwell, our objective was to make an enormous leap forward in the capability of the GPU to perform
near-photo-real lighting calculations in real time.
In the real world, all objects are lit by a combination of direct light (photons that travel directly from a
light source to illuminate an object) and indirect light (photons that travel from the light source, hit one
object and bounce off of it and then hit a second object, thus indirectly illuminating that object). “Global
illumination” (GI) is a term for lighting systems that model this effect. Without indirect lighting, scenes
can look harsh and artificial. However, while light received directly is fairly simple to compute, indirect
lighting computations are highly complex and computationally heavy.

GeForce GTX 980 Whitepaper
MAXWELL: ENABLING THE NEXT
FRONTIER IN PC GRAPHICS
14
Figure 4: Scene rendered with direct lighting only
Figure 5: Same scene with GI enabled, note the indirect light and specular reflections on the floor
Because it’s a computationally expensive lighting technique (particularly in highly detailed scenes), GI
has been primarily used to render complex CG scenes in movies using offline GPU rendering farms.
While some forms of GI have been used in many of today’s most popular games, their implementations
have relied on pre-computed lighting. These “prebaked” techniques are used for performance reasons;
however, they require additional artwork, as the desired lighting effects must be computed beforehand.

GeForce GTX 980 Whitepaper
MAXWELL: ENABLING THE NEXT
FRONTIER IN PC GRAPHICS
15
Because prebaked lighting is not dynamic, it’s often difficult or impossible to update the indirect light
sources when in-game changes occur; say for instance an additional light source is added or something
in the scene moves or is destroyed. Prebaked indirect lighting models the static objects of the scene, but
doesn’t properly apply to the animated characters or moving objects.
In 2011, NVIDIA engineers developed and demonstrated an innovative new approach to computing a
fast, approximate form of global illumination dynamically in real time on the GPU. This new GI
technology uses a voxel grid to store scene and lighting information, and a novel voxel cone tracing
process to gather indirect lighting from the voxel grid. NVIDIA’s Cyril Crassin describes the technique in
his paper on the topic and a video from GTC 2012 is available here. Epic’s ‘Elemental’ Unreal Engine 4
tech demo from 2012 used a similar technique.
Figure 6: Epic's UE4 'Elemental' tech demo used voxel cone tracing for its jaw-dropping GI
Since that time, NVIDIA has been working on the next generation of this technology—VXGI—that
combines new software algorithms and special hardware acceleration in the Maxwell architecture.
To understand how voxel global illumination works, it is helpful to first understand voxels. The term
“voxel” is related to “pixel.”Whereas a pixel represents a 2D point in space, a voxel represents a small
cube (a volume) of 3D space. To perform global illumination, we need to understand the light emitting
from all of the objects in the scene, not just the direct lights. To accomplish this, we dice the entire 3D
space of the scene in all three dimensions, into small cubes called voxels. “Voxelization” is the process of
determining the content of the scene at every voxel, analogous to “rasterization” which is the process of
determining the value of a scene at a given 2D coordinate.

GeForce GTX 980 Whitepaper
MAXWELL: ENABLING THE NEXT
FRONTIER IN PC GRAPHICS
16
The following picture illustrates this approach applied to a single triangle. In the picture, each cube that
contains a portion of the purple triangle is highlighted in blue.
Figure 7: Voxels touched by a triangle
In VXGI, we store two pieces of key information in each voxel: (a) the fraction of the voxel that contains
an actual object, and (b) for any voxel that contains an object, the properties of light coming from that
object (i.e., bouncing off of it from primary light sources), including direction and intensity.
The first step, as illustrated in the following figure, is the coverage calculation step. In this step, each
triangle needs to be checked from the perspective of each face of the cube, to assess what fraction of
the voxel is covered. The picture on the left shows a traditional rasterized image of a simple scene. The
picture on the right is a visualization of the voxelized result. In this picture, empty voxels are not drawn,
voxels that are fully covered are drawn in red, and voxels that are partially covered are represented by a
shade between blue (minimally covered) and red (fully covered). Note that the interior of each box is
solid red, as expected, while the edges are blue, due to the edge of the triangle intersecting but not fully
covering a voxel.

GeForce GTX 980 Whitepaper
MAXWELL: ENABLING THE NEXT
FRONTIER IN PC GRAPHICS
17
Figure 8: Voxelization results
Once the voxel coverage stage is complete, we store information into each voxel describing how the
physical geometry will respond to light. This includes encoding the material’s opacity and
emissive/reflective properties. We then evaluate direct lighting at each non-empty voxel, and render the
scene multiple times from the point of view of different light sources, capturing the amount of light that
hits each voxel.
In the figure below, the direct light source indicated by the yellow dot causes light to strike the white
walls and some of the surfaces of the red and green boxes. Each will then emit reflected light based on
the color of the object (i.e., the white wall will emit a small amount of reflected white light, while the
green box will emit reflected green light).

GeForce GTX 980 Whitepaper
MAXWELL: ENABLING THE NEXT
FRONTIER IN PC GRAPHICS
18
Figure 9: Results of light injection
After these steps, we now have a complete voxel data structure representing the lighting information in
the scene.
The next and final step is to rasterize the scene. This step is largely the same as it would be for a scene
rendered with other lighting approaches; the main difference is that the final rasterization and lighting
now has a new and more powerful data structure (the voxel data structure) that it can use in its lighting
calculations, along with other structures such as shadow maps.
The approach of calculating indirect lighting during the final rendering pass of VXGI is called cone
tracing. Cone tracing is an approximation of the effect of secondary rays that are used in ray tracing
methods. Using cones results in very realistic approximations of global illumination at a much lower
computational cost than standard ray tracing.

GeForce GTX 980 Whitepaper
MAXWELL: ENABLING THE NEXT
FRONTIER IN PC GRAPHICS
19
Rendering a real-time reflection from a glossy curved surface has been difficult in the past. In order to
render glossy reflections, you would traditionally need to launch hundreds or thousands of scattered
secondary rays for each ray that bounces from the original reflector. It’s incredibly challenging to reflect
these lights realistically, especially when you also factor in the material properties of the various light
reflectors. Using our approach, we’ve replaced the thousands of secondary rays with just a handful of
voxel cones that are traced through the voxel grid.

GeForce GTX 980 Whitepaper
MAXWELL: ENABLING THE NEXT
FRONTIER IN PC GRAPHICS
20
We also use the same approach to quickly compute diffuse or specular lighting with only a few scattered
cones. Ultimately as a result, we’re able to compute approximate GI at high frame rates in real time,
allowing us to realistically render glossy and metallic surfaces.
Figure 10: In the example above voxel cones are used to produce various forms of diffuse and specular light