NVIDIA introduced the A2 Tensor Core GPU at GTC 2018, offering 20% more performance than CPUs for inference workloads. NVIDIA is positioning itself to be the top choice in AI training and inferencing with this new breakthrough technology.
NVIDIA has introduced a new GPU with the name A2. The Tensor Core GPU is 20 percent more powerful than CPUs and offers up to twice the performance of its predecessor, the Tesla P100.
NVIDIA image
NVIDIA today announced the A2 Tensor Core GPU, an entry-level GPU that, despite its low power and tiny footprint, can give 20% greater inference performance than CPUs. Inference is a step in the Deep Learning process that involves applying and processing learned models produced during the training (i.e., fact collection or learning) phase. For this reason, NVIDIA’s A2 Tensor Core GPU delivers up to 1.3x greater performance in a variety of intelligent edge use scenarios.
NVIDIA says:
The adaptability, small size, and low power consumption of the A2 surpass the expectations for large-scale edge deployments, rapidly upgrading entry-level CPU servers to perform inference. At an entry-level price point, servers accelerated with A2 GPUs promise stronger inference performance vs CPUs and more efficient intelligent video analytics (IVA) installations than earlier GPU generations. […]
AI inference is used to make consumers’ life easier by providing real-time experiences and allowing them to gather data from billions of end-point sensors and cameras. When compared to CPU-only servers, NVIDIA A2 Tensor Core GPU-based servers provide up to 20X greater inference performance, immediately upgrading any server to handle current AI.
Specifications for the A2 Tensor Core
Peak FP32 | 4.5 TF |
Tensor Core TF32 | 18 TF | 9 TF |
Tensor Core BFLOAT16 | 36 TF | 18 TF |
Peak Tensor Core FP16 | 36 TF | 18 TF |
Peak Tensor Core INT8 | 72 TOPS | 36 TOPS | 36 TOPS | 36 TOPS | 36 TOPS | 36 |
Peak Tensor Core INT4 | 144 TOPS | 72 TOPS |
RT Cores | 10 |
Engines of the media | 1 encoder for video 2 decoders for video (includes AV1 decode) |
Memory on the GPU | GDDR6 16GB |
Memory bandwidth on the GPU | 200GB/s |
Interconnect | x8 PCIe Gen4 |
Factor of appearance | PCIe with one slot and a modest profile |
Thermal design maximum power (TDP) | 40–60 watts (configurable) |
Support for virtual GPU (vGPU) software | NVIDIA AI Enterprise, NVIDIA Virtual Compute Server, NVIDIA Virtual PC (vPC), NVIDIA Virtual Applications (vApps), NVIDIA RTX Virtual Workstation (vWS), NVIDIA RTX Virtual Workstation (vWS), NVIDIA RTX Virtual Workstation (vWS), NVIDIA RTX Virtual Workstation (vWS), NVIDIA RTX Virtual Workstation (vWS (vCS) |
NVIDIA is the source of this information.
Recent Developments
Samsung Announces the First LPDDR5X DRAM with Up to 8.5 Gbps Speeds in the Industry
9 November 2021 9 November 2021
Audio2Face by NVIDIA is a software tool for synchronizing audio with video animations.
9 November 2021 9 November 2021
NVIDIA is said to be capping frame rates in certain games on GeForce NOW Priority Membership, with Cyberpunk 2077 running at just 45 frames per second.
8 November 2021 8 November 2021
The PS5 is one step closer to being jailbroken, according to PlayStation hackers.
8 November 2021 8 November 2021
AMD Introduces the Instinct MI200 GPU Series and Teases “Zen 4” EPYC Processors
8 November 2021 8 November 2021
Metal Gear Solid Games are being temporarily removed from Konami’s website due to historical footage.
8 November 2021 8 November 2021
NVIDIA has introduced a new GPU called the A2 Tensor Core. It is designed to offer 20 percent more inference performance than CPUs. Reference: nvidia tensor cores.
Related Tags
- gpu for inference
- rtx tensor cores
- tensor core vs cuda core
- amd tensor cores
- gpu-based deep learning inference a performance and power analysis