Neural networks are both computationally intensive and memory intensive, making them difficult to deploy on embedded systems with limited hardware resources. To address this limitation, we introduce "deep compression", a three stage pipeline: pruning, trained quantization and Huffman coding, that work together to reduce the storage requirement of neural networks by 35× to 49× without affecting their accuracy. Our method first prunes the network by learning only the important connections. Next, we quantize the weights to enforce weight sharing, finally, we apply Huffman coding. After the first two steps we retrain the network to fine tune the remaining connections and the quantized centroids. Pruning, reduces the number of connections by 9× to 13×; Quantization then reduces the number of bits that represent each connection from 32 to 5. On the ImageNet dataset, our method reduced the storage required by AlexNet by 35×, from 240MB to 6.9MB, without loss of accuracy. Our method reduced the size of VGG-16 by 49× from 552MB to 11.3MB, again with no loss of accuracy. This allows fitting the model into on-chip SRAM cache rather than off-chip DRAM memory. Our compression method also facilitates the use of complex neural networks in mobile applications where application size and download bandwidth are constrained. Benchmarked on CPU, GPU and mobile GPU, compressed network has 3× to 4× layerwise speedup and 3× to 7× better energy efficiency.

State-of-the-art deep neural networks (DNNs) have hundreds of millions of connections and are both computationally and memory intensive, making them difficult to deploy on embedded systems with limited hardware resources and power budgets. While custom hardware helps the computation, fetching weights from DRAM is two orders of magnitude more expensive than ALU operations, and dominates the required power.Previously proposed 'Deep Compression' makes it possible to fit large DNNs (AlexNet and VGGNet) fully in on-chip SRAM. This compression is achieved by pruning the redundant connections and having multiple connections share the same weight. We propose an energy efficient inference engine (EIE) that performs inference on this compressed network model and accelerates the resulting sparse matrix-vector multiplication with weight sharing. Going from DRAM to SRAM gives EIE 120× energy saving; Exploiting sparsity saves 10×; Weight sharing gives 8×; Skipping zero activations from ReLU saves another 3×. Evaluated on nine DNN benchmarks, EIE is 189× and 13× faster when compared to CPU and GPU implementations of the same DNN without compression. EIE has a processing power of 102 GOPS/s working directly on a compressed network, corresponding to 3 TOPS/s on an uncompressed network, and processes FC layers of AlexNet at 1.88×10 4 frames/sec with a power dissipation of only 600mW. It is 24,000× and 3,400× more energy efficient than a CPU and GPU respectively. Compared with DaDianNao, EIE has 2.9×, 19× and 3× better throughput, energy efficiency and area efficiency.

The bandwidth and latency of a memory system are strongly dependent on the manner in which accesses interact with the "3-D" structure of banks, rows, and columns characteristic of contemporary DRAM chips. There is nearly an order of magnitude difference in bandwidth between successive references to different columns within a row and different rows within a bank. This paper introduces memory access scheduling, a technique that improves the performance of a memory system by reordering memory references to exploit locality within the 3-D memory structure. Conservative reordering, in which the first ready reference in a sequence is performed, improves bandwidth by 40% for traces from five media benchmarks. Aggressive reordering, in which operations are scheduled to optimize memory bandwidth, improves bandwidth by 93% for the same set of applications. Memory access scheduling is particularly important for media processors where it enables the processor to make the most efficient use of scarce memory bandwidth.

No abstract