Pascal (microarchitecture)

Lore & Background

Pascal was announced as the successor to Maxwell in March 2014, announced on May 6, 2016, and released on May 27 of the same year. The architecture encompasses multiple chip variants: GP100 (used in Tesla P100 and Quadro GP100), GP102 (Titan Xp, Titan X Pascal, GeForce GTX 1080 Ti, Quadro P6000, Tesla P40), GP104 (GeForce GTX 1070, GTX 1070 Ti, GTX 1080, some GTX 1060 6 GB, Quadro P5000/P4000/P3200, Tesla P4), GP106 (GeForce GTX 1060, Quadro P2000), GP107 (GeForce GTX 1050/1050 Ti, Quadro P1000/P600/P620/P400), and GP108 (GeForce GT 1010, GT 1030). Architectural improvements in GP100 include High Bandwidth Memory 2 (16 GiB HBM2 in four stacks, 4096-bit bus, 720 GB/s bandwidth), unified memory with Page Migration Engine, NVLink (80–200 GB/s), FP16 operations at twice the rate of FP32, FP64 at half the rate of FP32, doubled registers per CUDA core, more shared memory, dynamic load balancing scheduling, and instruction-level and thread-level preemption. GP104 improvements include GDDR5X memory (10 Gbit/s data rates), Simultaneous Multi-Projection, DisplayPort 1.4, HDMI 2.0b, fourth-generation Delta Color Compression, enhanced SLI, PureVideo Feature Set H (HEVC Main10/Main12, VP9 hardware decoding), HDCP 2.2, NVENC HEVC Main10 10-bit encoding, GPU Boost 3.0, and instruction-level preemption.

In Pascal, a streaming multiprocessor (SM) consists of 64–128 CUDA cores depending on the chip: GP100 SMs have 64 single-precision cores (plus 32 double-precision), while GP104 SMs have 128 single-precision ALUs. The GP100 SM is partitioned into two processing blocks, each with 32 single-precision CUDA cores, an instruction buffer, a warp scheduler, 2 texture mapping units, and 2 dispatch units. The Polymorph Engine 4.0 handles tessellation and was moved from the shader module to the TPC to feed multiple SMs.

Reader's Guide

Pascal represents a significant generational step in Nvidia's GPU architecture, introducing several technologies that shaped subsequent designs. The GP100 variant brought HBM2 memory and NVLink, targeting high-performance computing and deep learning with enhanced FP16 and FP64 throughput. The consumer GP104 and its derivatives brought GDDR5X memory, Simultaneous Multi-Projection, and improved video encoding/decoding capabilities.

Pascal's unified memory architecture with Page Migration Engine allowed CPU and GPU to share system and graphics memory, while dynamic load balancing enabled asynchronous compute in the driver. The architecture's support for DirectX Raytracing via software implementation (enabled in April 2019 on GTX 1060 6 GB and 16 series cards) extended its relevance beyond its original feature set. Pascal was succeeded by Volta in 2017 for HPC, cloud, and self-driving car markets, and by Turing in 2018 for consumer and business markets.

Did You Know?

• The GP100 SM contains 64 single-precision CUDA cores and 32 double-precision units, giving a 2:1 ratio of single- to double-precision throughput.
• The theoretical double-precision performance of GP102, GP104, GP106, GP107, and GP108 is 1/32 of their single-precision performance.