Models of artificial intelligence (AI) that have billions of parameters can achieve high accuracy across a range of tasks1,2, but they exacerbate the poor energy efficiency of conventional general-purpose processors, such as graphics processing units or central processing units. Analog in-memory computing (analog-AI)3-7 can provide better energy efficiency by performing matrix-vector multiplications in parallel on 'memory tiles'. However, analog-AI has yet to demonstrate software-equivalent (SWeq) accuracy on models that require many such tiles and efficient communication of neural-network activations between the tiles. Here we present an analog-AI chip that combines 35 million phase-change memory devices across 34 tiles, massively parallel inter-tile communication and analog, low-power peripheral circuitry that can achieve up to 12.4 tera-operations per second per watt (TOPS/W) chip-sustained performance. We demonstrate fully end-to-end SWeq accuracy for a small keyword-spotting network and near-SWeq accuracy on the much larger MLPerf8 recurrent neural-network transducer (RNNT), with more than 45 million weights mapped onto more than 140 million phase-change memory devices across five chips.

Calcium buffering capacity declines with age in sympathetic nerves of rat tail artery. To test whether smooth endoplasmic reticulum (SER) calcium buffering declines with age, effects of two SER calcium-ATPase inhibitors on norepinephrine release and intracellular calcium were determined. Thapsigargin or cyclopiazonic acid caused a significant increase in stimulation-evoked norepinephrine release from 6 month tail arteries with much less effect in 20 months. In isolated superior cervical ganglion cells, the rate of rise of calcium with K+-depolarization increased only in young cells with either cyclopiazonic acid or thapsigargin, with no effect in the old. In young cells, cyclopiazonic acid significantly influenced time to peak, rate of decline, and time to basal of K+-evoked calcium transients, but had no effect in old cells. Thapsigargin caused a significant increase in rate of decline in young, but not old, cells. These differential effects suggest an age-related decline in function of SER calcium buffering mechanisms in the sympathetic nervous system causing older nerves to become more reliant on mitochondria to buffer calcium.

Stimulation-evoked norepinephrine release from rat tail artery adrenergic nerves increased with advancing age in the Fischer-344 rat when function of norepinephrine uptake mechanisms and prejunctional alpha-2 adrenoceptors were blocked. When calcium channels were bypassed with the ionophore, ionomycin (4 microM), norepinephrine release from aged nerves (20 months) was still elevated as compared to 6-month-old nerves. Norepinephrine release stimulated by high K+ was also higher in 20-month nerves. The intracellular calcium chelator, 1,2 bis(2-aminophenoxy)ethane-N,N,N',N'-tetraacetomethylester (BAPTA/AM), was used to determine whether age-related increases in norepinephrine release could be reversed with the addition of an artificial intracellular calcium buffer. Exposure to BAPTA/AM decreased stimulation-evoked norepinephrine release in both old and young tail arteries; however, the effect was significantly greater in older arteries. When mitochondrial calcium uptake was compromised using the uncoupler of mitochondrial oxidative phosphorylation, dinitrophenol, BAPTA caused a further decrease in stimulation-evoked norepinephrine release in 20-month tail arteries with much less effect in 6-month-old nerves. These results suggest that intracellular calcium buffering is less efficient in older nerves.