Maxwell is NVIDIA's next-generation architecture for CUDA compute applications. Maxwell retains and extends the same CUDA programming model as in previous NVIDIA architectures such as Fermi and Kepler, and applications that follow the best practices for those architectures should typically see speedups on the Maxwell architecture without any code changes. This guide summarizes the ways that an application can be fine-tuned to gain additional speedups by leveraging Maxwell architectural features.1
Maxwell introduces an all-new design for the Streaming Multiprocessor (SM) that dramatically improves energy efficiency. Although the Kepler SMX design was extremely efficient for its generation, through its development, NVIDIA's GPU architects saw an opportunity for another big leap forward in architectural efficiency; the Maxwell SM is the realization of that vision. Improvements to control logic partitioning, workload balancing, clock-gating granularity, compiler-based scheduling, number of instructions issued per clock cycle, and many other enhancements allow the Maxwell SM (also called SMM) to far exceed Kepler SMX efficiency.
The first Maxwell-based GPU is codenamed GM107 and is designed for use in power-limited environments like notebooks and small form factor (SFF) PCs. GM107 is described in a whitepaper entitled NVIDIA GeForce GTX 750 Ti: Featuring First-Generation Maxwell GPU Technology, Designed for Extreme Performance per Watt.2
The first GPU using the second-generation Maxwell architecture is codenamed GM204. Second-generation Maxwell GPUs retain the power efficiency of the earlier generation while delivering significantly higher performance. GM204 is described in a whitepaper entitled NVIDIA GeForce GTX 980: Featuring Maxwell, The Most Advanced GPU Ever Made.
Compute programming features of GM204 are similar to those of GM107, except where explicitly noted in this guide. For details on the programming features discussed in this guide, please refer to the CUDA C Programming Guide.
The performance guidelines and best practices described in the CUDA C Programming Guide and the CUDA C Best Practices Guide apply to all CUDA-capable GPU architectures. Programmers must primarily focus on following those recommendations to achieve the best performance.
- Find ways to parallelize sequential code,
- Minimize data transfers between the host and the device,
- Adjust kernel launch configuration to maximize device utilization,
- Ensure global memory accesses are coalesced,
- Minimize redundant accesses to global memory whenever possible,
- Avoid long sequences of diverged execution by threads within the same warp.
Before addressing specific performance tuning issues covered in this guide, refer to the Maxwell Compatibility Guide for CUDA Applications to ensure that your application is compiled in a way that is compatible with Maxwell.
The Maxwell Streaming Multiprocessor, SMM, is similar in many respects to the Kepler architecture's SMX. The key enhancements of SMM over SMX are geared toward improving efficiency without requiring significant increases in available parallelism per SM from the application.
- The register file size (64k 32-bit registers) is the same as that of SMX.
- The maximum registers per thread, 255, matches that of Kepler GK110. As with Kepler, experimentation should be used to determine the optimum balance of register spilling vs. occupancy, however.
- The maximum number of thread blocks per SM has been increased from 16 to 32. This should result in an automatic occupancy improvement for kernels with small thread blocks of 64 or fewer threads (shared memory and register file resource requirements permitting). Such kernels would have tended to under-utilize SMX, but less so SMM.
- Shared memory capacity is increased (see Shared Memory Capacity).
As such, developers can expect similar or improved occupancy on SMM without changes to their application. At the same time, warp occupancy requirements (i.e., available parallelism) for maximum device utilization are similar to or less than those of SMX (see Instruction Latencies).
The number of CUDA Cores per SM has been reduced to a power of two, however with Maxwell's improved execution efficiency, performance per SM is usually within 10% of Kepler performance, and the improved area efficiency of SMM means CUDA Cores per GPU will be substantially higher vs. comparable Fermi or Kepler chips. SMM retains the same number of instruction issue slots per clock and reduces arithmetic latencies compared to the Kepler design.
As with SMX, each SMM has four warp schedulers. Unlike SMX, however, all SMM core functional units are assigned to a particular scheduler, with no shared units. Along with the selection of a power-of-two number of CUDA Cores per SM, which simplifies scheduling and reduces stall cycles, this partitioning of SM computational resources in SMM is a major component of the streamlined efficiency of SMM.
The power-of-two number of CUDA Cores per partition simplifies scheduling, as each of SMM's warp schedulers issue to a dedicated set of CUDA Cores equal to the warp width. Each warp scheduler still has the flexibility to dual-issue (such as issuing a math operation to a CUDA Core in the same cycle as a memory operation to a load/store unit), but single-issue is now sufficient to fully utilize all CUDA Cores.
Another major improvement of SMM is that dependent math latencies have been significantly reduced; a consequence of this is a further reduction of stall cycles, as the available warp-level parallelism (i.e., occupancy) on SMM should be equal to or greater than that of SMX (see Occupancy), while at the same time each math operation takes less time to complete, improving utilization and throughput.
- The change in number of CUDA Cores per SM brings with it a corresponding change in peak single-precision floating point operations per clock per SM. However, since the number of SMs is typically increased, the result is an increase in aggregate peak throughput; furthermore, the scheduling and latency improvements also discussed above make this peak easier to approach.
- The throughput of many integer operations including multiply, logical operations and shift is improved. In addition, there are now specialized integer instructions that can accelerate pointer arithmetic. These instructions are most efficient when data structures are a power of two in size.
Maxwell combines the functionality of the L1 and texture caches into a single unit.
As with Kepler, global loads in Maxwell are cached in L2 only, unless using the LDG read-only data cache mechanism introduced in Kepler.
In a manner similar to Kepler GK110B, GM204 retains this behavior by default but also allows applications to opt-in to caching of global loads in its unified L1/Texture cache. The opt-in mechanism is the same as with GK110B: pass the -Xptxas -dlcm=ca flag to nvcc at compile time.
Local loads also are cached in L2 only, which could increase the cost of register spilling if L1 local load hit rates were high with Kepler. The balance of occupancy versus spilling should therefore be reevaluated to ensure best performance. Especially given the improvements to arithmetic latencies, code built for Maxwell may benefit from somewhat lower occupancy (due to increased registers per thread) in exchange for lower spilling.
The unified L1/texture cache acts as a coalescing buffer for memory accesses, gathering up the data requested by the threads of a warp prior to delivery of that data to the warp. This function previously was served by the separate L1 cache in Fermi and Kepler.
Two new device attributes were added in CUDA Toolkit 6.0: globalL1CacheSupported and localL1CacheSupported. Developers who wish to have separately-tuned paths for various architecture generations can use these fields to simplify the path selection process.
GK110 introduced a new architectural feature called Dynamic Parallelism, which allows the GPU to create additional work for itself. A programming model enhancement leveraging this feature was introduced in CUDA 5.0 to enable kernels running on GK110 to launch additional kernels onto the same GPU.
SMM brings Dynamic Parallelism into the mainstream by supporting it across the product line, even in lower-power chips such as GM107. This will benefit developers, as it means that applications will no longer need special-case algorithm implementations for high-end GPUs that differ from those usable in more power-constrained environments.
- Initial Public Release
- Updated for second-generation Maxwell (compute capability 5.2).
ALL NVIDIA DESIGN SPECIFICATIONS, REFERENCE BOARDS, FILES, DRAWINGS, DIAGNOSTICS, LISTS, AND OTHER DOCUMENTS (TOGETHER AND SEPARATELY, "MATERIALS") ARE BEING PROVIDED "AS IS." NVIDIA MAKES NO WARRANTIES, EXPRESSED, IMPLIED, STATUTORY, OR OTHERWISE WITH RESPECT TO THE MATERIALS, AND EXPRESSLY DISCLAIMS ALL IMPLIED WARRANTIES OF NONINFRINGEMENT, MERCHANTABILITY, AND FITNESS FOR A PARTICULAR PURPOSE.
Information furnished is believed to be accurate and reliable. However, NVIDIA Corporation assumes no responsibility for the consequences of use of such information or for any infringement of patents or other rights of third parties that may result from its use. No license is granted by implication of otherwise under any patent rights of NVIDIA Corporation. Specifications mentioned in this publication are subject to change without notice. This publication supersedes and replaces all other information previously supplied. NVIDIA Corporation products are not authorized as critical components in life support devices or systems without express written approval of NVIDIA Corporation.
NVIDIA and the NVIDIA logo are trademarks or registered trademarks of NVIDIA Corporation in the U.S. and other countries. Other company and product names may be trademarks of the respective companies with which they are associated.