All NVIDIA GPUs support general-purpose computation (GPGPU), but not all GPUs offer the same performance or support the same features. The consumer line of GeForce GPUs (GTX Titan X, in particular) may be attractive to those running GPU-accelerated applications. However, it’s wise to keep in mind the differences between the products. There are many features only available on the professional Tesla and Quadro GPUs.
FP64 64-bit (Double Precision) Floating Point Calculations
Many applications require higher-accuracy mathematical calculations. In these applications, data is represented by values that are twice as large (using 64 binary bits instead of 32 bits). These larger values are called double-precision (64-bit). Less accurate values are called single-precision (32-bit).
Although almost all NVIDIA GPU products support both single- and double-precision calculations, the performance for double-precision values is significantly lower on the consumer-level GeForce GPUs. Here is a comparison of the double-precision floating-point calculation performance between GeForce and Tesla/Quadro GPUs:
|NVIDIA GPU Model||Double-precision (64-bit) Floating Point Performance|
|GeForce GTX Titan X Maxwell||up to 0.206 TFLOPS|
|GeForce GTX 1080 Ti||up to 0.355 TFLOPS|
|GeForce Titan Xp||up to 0.380 TFLOPS|
|Tesla K80||1.87+ TFLOPS|
|Tesla P100*||4.7 ~ 5.3 TFLOPS|
|Quadro GP100||5.2 TFLOPS|
|Tesla V100*||7 ~ 7.8 TFLOPS|
Some earlier versions of the GeForce GTX Titan products did support up to 1.3 TFLOPS double-precision calculations. However, those models were discontinued in 2014.
* Exact value depends upon PCI-Express or SXM2 SKU
FP16 16-bit (Half Precision) Floating Point Calculations
Training neural networks for AI/Deep Learning does not typically require higher-accuracy. In fact, the industry has settled on a 16-bit “half-precision” operation when training neural networks. Having GPUs with high FP16 performance is crucial for quickly training new networks.
Support for FP16 operations was introduced in the “Pascal” generation of GPUs. Although all NVIDIA “Pascal” and later GPU generations support FP16, performance is significantly lower on the consumer-level GeForce GPUs. Here is a comparison of the half-precision floating-point calculation performance between GeForce and Tesla/Quadro GPUs:
|NVIDIA GPU Model||Half-precision (16-bit) Floating Point Performance|
|GeForce GTX Titan X Maxwell||N/A|
|GeForce GTX 1080 Ti||less than 0.177 TFLOPS|
|GeForce Titan Xp||less than 0.190 TFLOPS|
|Tesla P100*||18.7 ~ 21.2 TFLOPS|
|Quadro GP100||20.7 TFLOPS|
|Tesla V100*||112 ~ 125 TFLOPS|
* Exact value depends upon PCI-Express or SXM2 SKU
Error Detection and Correction
On a GPU running a computer game, one memory error typically causes no issues (e.g., one pixel color might be incorrect for one frame). The user is very unlikely to even be aware of the issue.
However, technical computing applications rely on the accuracy of the data returned by the GPU. For some applications, a single error can cause the simulation to be grossly and obviously incorrect. For others, a single-bit error may not be so easy to detect (returning incorrect results which appear reasonable).
Titan X GPUs do not include error correction or error detection capabilities. Neither the GPU nor the system can alert the user to errors should they occur. It is up to the user to detect errors (whether they cause application crashes, obviously incorrect data, or subtly incorrect data). Such issues are not uncommon – our technicians regularly encounter memory errors on consumer gaming GPUs.
NVIDIA Tesla GPUs are able to correct single-bit errors and detect & alert on double-bit errors. On the latest Tesla P100 and Quadro GP100 GPUs, ECC support is included in the main HBM2 memory, as well as in register files, shared memories, L1 cache and L2 cache.
NVIDIA’s warranty on GeForce GPU products explicitly states that the GeForce products are not designed for installation in servers. Running GeForce GPUs in a server system will void the warranty. From NVIDIA’s manufacturer warranty website:
Warranted Product is intended for consumer end user purposes only, and is not intended for datacenter use and/or GPU cluster commercial deployments (“Enterprise Use”). Any use of Warranted Product for Enterprise Use shall void this warranty.
GPU Memory Performance
Computationally-intensive applications require high-performance compute units, but fast access to data is also critical. For many HPC applications, an increase in compute performance does not help unless memory performance is also improved. For this reason, the Tesla GPUs provide better real-world performance than the GeForce GPUs:
|NVIDIA GPU Model||GPU Memory Bandwidth|
|GeForce GTX Titan X Maxwell||336 GB/s|
|GeForce GTX 1080 Ti||484 GB/s|
|GeForce Titan Xp||548 GB/s|
|Tesla K80||480 GB/s|
|Tesla P40||346 GB/s|
|Tesla P100 12GB||549 GB/s|
|Tesla P100 16GB||732 GB/s|
|Quadro GP100||717 GB/s|
|Tesla V100 16GB||900 GB/s|
The primary reason for this performance disparity is that GeForce GPUs use GDDR5 memory, while the latest Tesla GPUs use on-die HBM2 memory.
GPU Memory Quantity
In general, the more memory a system has the faster it will run. For some HPC applications, it’s not even possible to perform a single run unless there is sufficient memory. For others, the quality and fidelity of the results will be degraded unless sufficient memory is available. Tesla GPUs offer as much as twice the memory of GeForce GPUs:
|NVIDIA GPU Model||GPU Memory Quantity|
|GeForce GTX 1080 Ti||11GB|
|GeForce Titan Xp||12GB|
|Tesla P100||12GB or 16GB*|
* note that Tesla/Quadro Pascal Unified Memory allows GPUs to share each other’s memory to load even larger datasets
PCI-E vs NVLink – Device-to-Host and Device-to-Device Throughput
One of the largest potential bottlenecks is in waiting for data to be transferred to the GPU. Additional bottlenecks are present when multiple GPUs operate in parallel. Faster data transfers directly result in faster application performance.
The GeForce GPUs connect via PCI-Express, which has a theoretical peak throughput of 16GB/s. NVIDIA Tesla/Quadro GPUs with NVLink are able to leverage much faster connectivity. The NVLink in NVIDIA’s “Pascal” generation allows each GPU to communicate at up to 80GB/s (160GB/s bidirectional). The NVLink 2.0 in NVIDIA’s “Volta” generation allows each GPU to communicate at up to 150GB/s (300GB/s bidirectional). NVLink connections are supported between GPUs, and also between the CPUs and the GPUs on supported OpenPOWER platforms. Only the Tesla and Quadro line of GPUs supports NVLink.
Application Software Support
While some software programs are able to operate on any GPU which supports CUDA, others are designed and optimized for the professional GPU series. Most professional software packages only officially support the NVIDIA Tesla and Quadro GPUs. Using a GeForce GPU may be possible, but will not be supported by the software vendor. In other cases, the applications will not function at all when launched on a GeForce GPU (for example, the software products from Schrödinger, LLC).
Operating System Support
Although NVIDIA’s GPU drivers are quite flexible, there are no GeForce drivers available for Windows Server operating systems. GeForce GPUs are only supported on Windows 7, Windows 8, and Windows 10. Groups that use Windows Server should look to NVIDIA’s professional Tesla and Quadro GPU products. The Linux drivers, on the other hand, support all NVIDIA GPUs.
Product Life Cycle
Due to the nature of the consumer GPU market, GeForce products have a relatively short lifecycle (commonly no more than a year between product release and end of production). Projects which require a longer product lifetime (such as those which might require replacement parts 3+ years after purchase) should use a professional GPU.
NVIDIA’s professional Tesla and Quadro GPU products have an extended lifecycle and long-term support from the manufacturer (including notices of product End of Life and opportunities for last buys before production is halted). Furthermore, the professional GPUs undergo a more thorough testing and validation process during production.
GeForce GPUs are intended for consumer gaming usage, and are not usually designed for power efficiency. In contrast, the Tesla GPUs are designed for large-scale deployment where power efficiency is important. This makes the Tesla GPUs a better choice for larger installations.
For example, the GeForce GTX Titan X is popular for desktop deep learning workloads. In server deployments, the Tesla P40 GPU provides matching performance and double the memory capacity. However, when put side-by-side the Tesla consumes less power and generates less heat.
The Direct Memory Access (DMA) Engine of a GPU allows for speedy data transfers between the system memory and the GPU memory. Because such transfers are part of any real-world application, the performance is vital to GPU-acceleration. Slow transfers cause the GPU cores to sit idle until the data arrives in GPU memory. Likewise, slow returns cause the CPU to wait until the GPU has finished returning results.
GeForce products feature a single DMA Engine* which is able to transfer data in one direction at a time. If data is being uploaded to the GPU, any results computed by the GPU cannot be returned until the upload is complete. Likewise, results being returned from the GPU will block any new data which needs to be uploaded to the GPU.
* one GeForce GPU model, the GeForce GTX Titan X, features dual DMA engines
The Tesla GPU products feature dual DMA Engines to alleviate this bottleneck. Data may be transferred into the GPU and out of the GPU simultaneously.
GPU Direct RDMA
NVIDIA’s GPU-Direct technology allows for greatly improved data transfer speeds between GPUs. Various capabilities fall under the GPU-Direct umbrella, but the RDMA capability promises the largest performance gain.
Traditionally, sending data between the GPUs of a cluster required 3 memory copies (once to the GPU’s system memory, once to the CPU’s system memory and once to the InfiniBand driver’s memory). GPU Direct RDMA removes the system memory copies, allowing the GPU to send data directly through InfiniBand to a remote system. In practice, this has resulted in up to 67% reductions in latency and 430% increases in bandwidth for small MPI message sizes .
In CUDA version 8.0, NVIDIA has introduced GPU Direct RDMA ASYNC, which allows the GPU to initiate RDMA transfers without any interaction with the CPU.
GeForce GPUs do not support GPU-Direct RDMA. Although the MPI calls will still return successfully, the transfers will be performed through the standard memory-copy paths. The only form of GPU-Direct which is supported on the GeForce cards is GPU Direct Peer-to-Peer (P2P). This allows for fast transfers within a single computer, but does nothing for applications which run across multiple servers/compute nodes.
Tesla GPUs have full support for GPU Direct RDMA and the various other GPU Direct capabilities. They are the primary target for these capabilities and thus have the most testing and use in the field.
Hyper-Q Proxy for MPI and CUDA Streams allows multiple CPU threads or processes to launch work on a single GPU. This is particularly important for existing parallel applications written with MPI, as these codes have been designed to take advantage of multiple CPU cores. Allowing the GPU to accept work from each of the MPI threads running on a system can offer a potentially significant performance boost. It can also reduce the amount of source code re-architecting required to add GPU acceleration to an existing application.
However, the only form of Hyper-Q which is supported on the GeForce GPUs is Hyper-Q for CUDA Streams. This allows the GeForce to efficiently accept and run parallel calculations from separate CPU cores, but applications running across multiple computers will be unable to efficiently launch work on the GPU.
GPU Health Monitoring and Management Capabilities
Many health monitoring and GPU management capabilities (which are vital for maintaining multiple GPU systems) are only supported on the professional Tesla GPUs. Health features which are not supported on the GeForce GPUs include:
- NVML/nvidia-smi for monitoring and managing the state and capabilities of each GPU. This enables GPU support from a number of 3rd party applications and tools such as Ganglia. Perl and Python bindings are also available.
- OOB (out of band monitoring via IPMI) allows the system to monitor GPU health, adjust fan speeds to appropriately cool the devices and send alerts when an issue is seen
- InfoROM (persistent configuration and state data) provides the system with additional data about each GPU
- NVHealthmon utility provides cluster administrators with a ready-to-use GPU health status tool
- TCC allows GPUs to be specifically set to display-only or compute-only modes
- ECC (memory error detection & correction)
Cluster tools rely upon the capabilities provided by NVIDIA NVML. Roughly 60% of the capabilities are not available on GeForce – this table offers a more detailed comparison of the NVML features supported in Tesla and GeForce GPUs:
|Show GPU Count||yes||yes|
|PCI-Express Generation (e.g., 2.0 vs 3.0)||yes||–|
|PCI-Express Link Width (e.g., x4, x8, x16)||yes||–|
|Current Fan Speed||yes||yes|
|Current Performance State||yes||–|
|Clock Throttle Status||yes||–|
|Current GPU Usage (percentage)||yes||–|
|Current Memory Usage (percentage)||yes||yes|
|GPU Boost Capability||yes||yes^|
|ECC Error Detection/Correction Support||yes||–|
|List Retired Pages||yes||–|
|Current Power Draw||yes||–|
|Set Power Draw Limit||yes||–|
|Current GPU Clock Speed||yes||–|
|Current Memory Clock Speed||yes||–|
|Show Available Clock Speeds||yes||–|
|Show Available Memory Speeds||yes||–|
|Set GPU Boost Speed (core clock and memory clock)||yes||–|
|Show Current Compute Processes||yes||–|
|Card Serial Number||yes||–|
|InfoROM image and objects||yes||–|
|Accounting Capability (resource usage per process)||yes||–|
|NVIDIA Driver Version||yes||yes|
|NVIDIA VBIOS Version||yes||yes|
* Temperature reading is not available to the system platform, which means fan speeds cannot be adjusted.
^ GPU Boost is disabled during double precision calculations. Additionally, GeForce clock speeds will be automatically reduced in certain scenarios.
All of the latest NVIDIA GPU products support GPU Boost, but their implementations vary depending upon the intended usage scenario. GeForce cards are built for interactive desktop usage and gaming. Tesla GPUs are built for intensive, constant number crunching with stability and reliability placed at a premium. Given the differences between these two use cases, GPU Boost functions differently on Tesla than on GeForce.
In Geforce’s case, the graphics card automatically determines clock speed and voltage based on the temperature of the GPU. Temperature is the appropriate independent variable as heat generation affects fan speed. For less graphically-intense games or for general desktop usage, the end user can enjoy a quieter computing experience. When playing games that require serious GPU compute, however, GPU Boost automatically cranks up the voltage and clock speeds (in addition to generating more noise).
Tesla’s GPU boost level, on the other hand, may be specified by the system administrator or computational user – the desired clock speed is set to a specific frequency. Rather than floating the clock speed at various levels, the desired clock speed may be statically maintained unless the power consumption threshold (TDP) is reached. This is an important consideration because accelerators in an HPC environment often need to be in sync with one other. The deterministic aspect of Tesla’s GPU boost allows system administrators to determine optimal clock speeds and lock them in across all GPUs.
For applications that require additional performance, the most recent Tesla GPUs include Auto Boost within synchronous boost groups. With Auto Boost enabled, each group of GPUs will increase clock speeds when headroom allows. The group will keep clocks in sync with each other to ensure matching performance across the group.