NVIDIA GeForce GTX TITAN Z, of the GeForce 700-line of graphics-cards, is the final major iteration featuring the Kepler microarchitecture (GK110-350-B1) and the final NVIDIA graphics card supporting dual-GPU configuration on a single PCB.
Die shot of a GK110 A1 GPU, found inside GeForce GTX Titan cards
The goal of Nvidia's previous architecture was design focused on increasing performance on compute and tessellation. With the Kepler architecture, Nvidia targeted their focus on efficiency, programmability, and performance.[2][3] The efficiency aim was achieved through the use of a unified GPU clock, simplified static scheduling of instruction and higher emphasis on performance per watt.[4] By abandoning the shader clock found in their previous GPU designs, efficiency is increased, even though it requires additional cores to achieve higher levels of performance. This is not only because the cores are more power-friendly (two Kepler cores using 90% power of one Fermi core, according to Nvidia's numbers), but also the change to a unified GPU clock scheme delivers a 50% reduction in power consumption in that area.[5]
Programmability aim was achieved with Kepler's Hyper-Q, Dynamic Parallelism and multiple new Compute Capabilities 3.x functionality. With it, higher GPU utilization and simplified code management was achievable with GK GPUs thus enabling more flexibility in programming for Kepler GPUs.[6]
Finally with the performance aim, additional execution resources (more CUDA cores, registers and cache) and with Kepler's ability to achieve a memory clock speed of 7 GHz, increases Kepler's performance when compared to previous Nvidia GPUs.[5][7]
Features
The GK Series GPU contains features from both the older Fermi and newer Kepler generations. Kepler based members add the following standard features:
Hyper-Q (Hyper-Q's MPI functionality reserve for Tesla only)
Grid Management Unit
NVIDIA GPUDirect (GPU Direct's RDMA functionality reserve for Tesla only)
Next Generation Streaming Multiprocessor (SMX)
GTX 780 PCB and die - A later revision of Kepler with more similarities to the GK110 than the initial 680.
Kepler employs a new streaming multiprocessor architecture called SMX. CUDA execution core counts were increased from 32 per each of 16 SMs to 192 per each of 8 SMX; the register file was only doubled per SMX to 65,536 x 32-bit for an overall lower ratio; between this and other compromises, despite the 3x overall increase in CUDA cores and clock increase (on the 680 vs. the Fermi 580), the actual performance gains in most operations were well under 3x. Dedicated FP64 CUDA cores are used rather than treating two FP32 cores as a single unit as was done previously, and very few were included on the consumer models resulting in 1/24th speed FP64 calculation compared to FP32. [9]
On the HPC models, the GK110/210, the SMX count was raised to 13-15 depending on the product, and more FP64 cores were included to bring the compute ratio up to 1/3rd FP32. On the GK110, per-thread register limit was quadrupled over fermi to 255, but this still only allows a thread using half of the registers to parallelize to 1/4 of each SMX. The GK210 (released at the same time) increased the register limit to 512 to improve performance in high register pressure situations like this. Texture cache, which programmers had already been using for compute as a read-only buffer in previous generations, was increased in size and the data path optimized for faster throughput when using this method. All levels of memory including the register file are single-bit ECC as well.
Another notable feature is that while Fermi GPUs could only be accessed by one CPU thread at a time, the HPC Kepler GPUs added multithreading support so high core count processors could open 32 connections and more easily saturate the compute capability. [10]
Simplified Instruction Scheduler
Additional die space reduction and power saving was achieved by removing a complex hardware block that handled the prevention of data hazards. [3][5][11][12]
GPU Boost
GPU Boost is a new feature which is roughly analogous to turbo boosting of a CPU. The GPU is always guaranteed to run at a minimum clock speed, referred to as the "base clock". This clock speed is set to the level which will ensure that the GPU stays within TDP specifications, even at maximum loads.[3] When loads are lower, however, there is room for the clock speed to be increased without exceeding the TDP. In these scenarios, GPU Boost will gradually increase the clock speed in steps, until the GPU reaches a predefined power target of 170W by default (on the 680 card).[5] By taking this approach, the GPU will ramp its clock up or down dynamically, so that it is providing the maximum amount of speed possible while remaining within TDP specifications.
The power target, as well as the size of the clock increase steps that the GPU will take, are both adjustable via third-party utilities and provide a means of overclocking Kepler-based cards.[3]
Microsoft Direct3D Support
Nvidia Fermi and Kepler GPUs of the GeForce 600 series support the Direct3D 11.0 specification. Nvidia originally stated that the Kepler architecture has full DirectX 11.1 support, which includes the Direct3D 11.1 path.[13] The following "Modern UI" Direct3D 11.1 features, however, are not supported:[14][15]
UAV (Unordered Access View) in non-pixel-shader stages.
According to the definition by Microsoft, Direct3D feature level 11_1 must be complete, otherwise the Direct3D 11.1 path can not be executed.[16]
The integrated Direct3D features of the Kepler architecture are the same as those of the GeForce 400 series Fermi architecture.[15]
Next Microsoft Direct3D Support
NVIDIA Kepler GPUs of the GeForce 600/700 series support Direct3D 12 feature level 11_0.[17]
TXAA Support
Exclusive to Kepler GPUs, TXAA is a new anti-aliasing method from Nvidia that is designed for direct implementation into game engines. TXAA is based on the MSAA technique and custom resolve filters. It is designed to address a key problem in games known as shimmering or temporal aliasing. TXAA resolves that by smoothing out the scene in motion, making sure that any in-game scene is being cleared of any aliasing and shimmering.[3]
Shuffle Instructions
The GK110 had a small number of instructions added to further improve performance. New shuffle instructions allow for threads within a warp to share data amongst themselves with an instruction that completes the normal store and load operations that previously required two accesses to local memory within one instruction, making the process around 6% faster than using local data storage. Atomic operations were also improved, with 9x increases in speed for some instructions and the addition of more atomic 64-bit operations, namely min, max, and, or, and xor.[11]
Hyper-Q
Hyper-Q expands GK110 hardware work queues from 1 to 32. The significance of this being that having a single work queue meant that Fermi could be under occupied at times as there wasn't enough work in that queue to fill every SM. By having 32 work queues, GK110 can in many scenarios, achieve higher utilization by being able to put different task streams on what would otherwise be an idle SMX. The simple nature of Hyper-Q is further reinforced by the fact that it's easily mapped to MPI, a common message passing interface frequently used in HPC. As legacy MPI-based algorithms that were originally designed for multi-CPU systems that became bottlenecked by false dependencies now have a solution. By increasing the number of MPI jobs, it's possible to utilize Hyper-Q on these algorithms to improve the efficiency all without changing the code itself.[11]
Dynamic Parallelism
Dynamic Parallelism ability is for kernels to be able to dispatch other kernels. With Fermi, only the CPU could dispatch a kernel, which incurs a certain amount of overhead by having to communicate back to the CPU. By giving kernels the ability to dispatch their own child kernels, GK110 can both save time by not having to go back to the CPU, and in the process free up the CPU to work on other tasks.[11]
Grid Management Unit
Enabling Dynamic Parallelism requires a new grid management and dispatch control system. The new Grid Management Unit (GMU) manages and prioritizes grids to be executed. The GMU can pause the dispatch of new grids and queue pending and suspended grids until they are ready to execute, providing the flexibility to enable powerful runtimes, such as Dynamic Parallelism. The CUDA Work Distributor in Kepler holds grids that are ready to dispatch, and is able to dispatch 32 active grids, which is double the capacity of the Fermi CWD. The Kepler CWD communicates with the GMU via a bidirectional link that allows the GMU to pause the dispatch of new grids and to hold pending and suspended grids until needed. The GMU also has a direct connection to the Kepler SMX units to permit grids that launch additional work on the GPU via Dynamic Parallelism to send the new work back to GMU to be prioritized and dispatched. If the kernel that dispatched the additional workload pauses, the GMU will hold it inactive until the dependent work has completed.[12]
NVIDIA GPUDirect
NVIDIA GPUDirect is a capability that enables GPUs within a single computer, or GPUs in different servers located across a network, to directly exchange data without needing to go to CPU/system memory. The RDMA feature in GPUDirect allows third party devices such as SSDs, NICs, and IB adapters to directly access memory on multiple GPUs within the same system, significantly decreasing the latency of MPI send and receive messages to/from GPU memory.[18] It also reduces demands on system memory bandwidth and frees the GPU DMA engines for use by other CUDA tasks. Kepler GK110 also supports other GPUDirect features including Peer‐to‐Peer and GPUDirect for Video.
NVENC is Nvidia's power efficient fixed-function encode that is able to take codecs, decode, preprocess, and encode H.264-based content. NVENC specification input formats are limited to H.264 output. But still, NVENC, through its limited format, can support up to 4096x4096 encode.[19]
Like Intel's Quick Sync, NVENC is currently exposed through a proprietary API, though Nvidia does have plans to provide NVENC usage through CUDA.[19]
Performance
The theoretical single-precision processing power of a Kepler GPU in GFLOPS is computed as 2 (operations per FMA instruction per CUDA core per cycle) × number of CUDA cores × core clock speed (in GHz). Note that like the previous generation Fermi, Kepler is not able to benefit from increased processing power by dual-issuing MAD+MUL like Tesla was capable of.
The theoretical double-precision processing power of a Kepler GK110/210 GPU is 1/3 of its single precision performance. This double-precision processing power is however only available on professional Quadro, Tesla, and high-end TITAN-branded GeForce cards, while drivers for consumer GeForce cards limit the performance to 1/24 of the single precision performance.[20] The lower performance GK10x chips are similarly capped to 1/24 of the single precision performance.[21]
^Angelini, Chris (7 November 2013). "Nvidia GeForce GTX 780 Ti Review: GK110, Fully Unlocked". Tom's Hardware. p. 1. Retrieved 6 December 2015. The card's driver deliberately operates GK110's FP64 units at 1/8 of the GPU's clock rate. When you multiply that by the 3:1 ratio of single- to double-precision CUDA cores, you get a 1/24 rate
.jpg)
