Many architects believe that major improvements in cost-energyperformance must now come from domain-specific hardware. This paper evaluates a custom ASIC-called a Tensor Processing Unit (TPU)-deployed in datacenters since 2015 that accelerates the inference phase of neural networks (NN). The heart of the TPU is a 65,536 8-bit MAC matrix multiply unit that offers a peak throughput of 92 TeraOps/second (TOPS) and a large (28 MiB) software-managed on-chip memory. The TPU's deterministic execution model is a better match to the 99th-percentile responsetime requirement of our NN applications than are the time-varying optimizations of CPUs and GPUs that help average throughput more than guaranteed latency. The lack of such features helps explain why, despite having myriad MACs and a big memory, the TPU is relatively small and low power. We compare the TPU to a server-class Intel Haswell CPU and an Nvidia K80 GPU, which are contemporaries deployed in the same datacenters. Our workload, written in the high-level TensorFlow framework, uses production NN applications (MLPs, CNNs, and LSTMs) that represent 95% of our datacenters' NN inference demand. Despite low utilization for some applications, the TPU is on average about 15X -30X faster than its contemporary GPU or CPU, with TOPS/Watt about 30X -80X higher. Moreover, using the GPU's GDDR5 memory in the TPU would triple achieved TOPS and raise TOPS/Watt to nearly 70X the GPU and 200X the CPU.

Many architects believe that major improvements in cost-energyperformance must now come from domain-specific hardware. This paper evaluates a custom ASIC-called a Tensor Processing Unit (TPU)-deployed in datacenters since 2015 that accelerates the inference phase of neural networks (NN). The heart of the TPU is a 65,536 8-bit MAC matrix multiply unit that offers a peak throughput of 92 TeraOps/second (TOPS) and a large (28 MiB) software-managed on-chip memory. The TPU's deterministic execution model is a better match to the 99th-percentile responsetime requirement of our NN applications than are the time-varying optimizations of CPUs and GPUs that help average throughput more than guaranteed latency. The lack of such features helps explain why, despite having myriad MACs and a big memory, the TPU is relatively small and low power. We compare the TPU to a server-class Intel Haswell CPU and an Nvidia K80 GPU, which are contemporaries deployed in the same datacenters. Our workload, written in the high-level TensorFlow framework, uses production NN applications (MLPs, CNNs, and LSTMs) that represent 95% of our datacenters' NN inference demand. Despite low utilization for some applications, the TPU is on average about 15X-30X faster than its contemporary GPU or CPU, with TOPS/Watt about 30X-80X higher. Moreover, using the GPU's GDDR5 memory in the TPU would triple achieved TOPS and raise TOPS/Watt to nearly 70X the GPU and 200X the CPU.

No abstract

Mycobacterium tuberculosis and filarial coinfection is highly prevalent, and the presence of a tissue-invasive helminth may modulate the predominant type 1 T helper (Th1; interferon [IFN]–γ–mediated) response needed to control M. tuberculosis infection. By analyzing the cellular responses to mycobacterial antigens in patients who had latent tuberculosis with or without filarial infection, we were able to demonstrate that filarial infection coincident with M. tuberculosis infection significantly diminishes M. tuberculosis–specific Th1 (interleukin [IL]–12 and IFN-γ) and type 17 T helper (Th17; IL-23 and IL-17) responses related to increased expression of cytotoxic T lymphocyte antigen (CTLA)–4 and programmed death (PD)–1. Blockade of CTLA-4 restored production of both IFN-γ and IL-17, whereas PD-1 blockade restored IFN-γ production only. Thus, coincident filarial infection exerted a profound inhibitory effect on protective mycobacteria-specific Th1 and Th17 responses in latent tuberculosis, suggesting a mechanism by which concomitant filarial (and other systemic helminth) infections predispose to the development of active tuberculosis in humans.

Reactions to pathogens are usually tuned to effect immunity and limit tissue damage. Several host counterinflammatory mechanisms inhibit tissue damage but these may also act to constrain the effectiveness of immunity to acute infections, as we demonstrate in mice acutely infected with influenza A virus (IAV). We show that compared with wild type (WT), galectin-9 knockout (G9KO) mice mounted a more robust acute phase virus-specific CD8 T-cell response as well as higher and more rapid virus-specific serum IgM, IgG, and IgA responses and also cleared virus more rapidly than did WT mice. Blocking galectin-9 signals to Tim-3-expressing cells using a Tim-3 fusion protein resulted in improved immune responses in WT mice. When IAV immune mice were challenged with a heterologous IAV, the secondary IAV-specific CD8 T-cell responses were fourto fivefold higher in G9KO compared with WT mice. Our results indicate that manipulating galectin signals may represent a convenient approach to improve immune responses to some vaccines.T he host immune response to pathogens needs precise regulation to minimize tissue damage while still achieving defense (1, 2). Some bystander tissue damage usually happens because several host defenses can destroy cells or orchestrate inflammatory reactions. With chronic infections, for example, immune-mediated tissue damage would be more severe were it not for several cellular and chemical host components that inhibit inflammatory reactions (1). However, the activity of some of these counterinflammatory mechanisms could act to constrain the efficiency of protective immune components (3). For instance, regulatory T cells (Tregs) can inhibit inflammatory reactions associated with chronic virus infections (4), but the same Treg response can also limit the magnitude of protective immunity to a virus or induced by a vaccine (5, 6). Other host components may also function to limit and help resolve inflammatory reactions. These include some cytokines (7), groups of molecules derived from omega-3 polyunsaturated fatty acids (8), as well as some of the carbohydrate binding proteins of the galectin family (9). Galectin-9 (Gal-9), for example, upon binding to Tim-3 on T cells acts to limit the extent of immunopathological lesions in autoimmunity (10) as well as in some chronic infections (11-13). In the present study, we investigated whether the inhibitory effects of Gal-9 on Tim-3-expressing cells could influence the outcome of acute infection with influenza A virus (IAV). We demonstrate that animals lacking the regulatory effects of Gal-9/Tim-3 triggering mounted superior CD8 T-cell and humoral immune responses and they were more refractory to IAV. Moreover, IAV immune G9KO mice challenged with a heterologous IAV strain generated better virus-specific memory CD8 T-cell responses than WT animals. Our results indicate that manipulating galectin signaling may represent a convenient approach to improve responses to some vaccines. Results Virus-Specific CD8 T cells Up-Regulate Tim-3 Expression after IAVInfection. ...