An analog-AI chip for energy-efficient speech recognition and transcription.

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.

Electrically conductive bulk composites through a contact-connected aggregate.

This paper introduces a concept that allows the creation of low-resistance composites using a network of compliant conductive aggregate units, connected through contact, embedded within the composite. Due to the straight-forward fabrication method of the aggregate, conductive composites can be created in nearly arbitrary shapes and sizes, with a lower bound near the length scale of the conductive cell used in the aggregate. The described instantiation involves aggregate cells that are approximately spherical copper coils-of-coils within a polymeric matrix, but the concept can be implemented with a wide range of conductor elements, cell geometries, and matrix materials due to its lack of reliance on specific material chemistries. The aggregate cell network provides a conductive pathway that can have orders of magnitude lower resistance than that of the matrix material--from 10(12) ohm-cm (approx.) for pure silicone rubber to as low as 1 ohm-cm for the silicone/copper composite at room temperature for the presented example. After describing the basic concept and key factors involved in its success, three methods of implementing the aggregate into a matrix are then addressed--unjammed packing, jammed packing, and pre-stressed jammed packing--with an analysis of the tradeoffs between increased stiffness and improved resistivity.

Length and substituent-scrambling energies of parent and halogen-substituted conjugated polyynes.

Conjugated polyynes are a class of species of diverse and increasing interest. Length-scrambling and substituent scrambling reaction energies were examined using ab initio quantum chemistry calculations to investigate issues concerning the energetic effects of the molecular ends (substituent communication). Computations were performed for the parent, monohalogenated, and dihalogenated (F, Cl, Br, I) polyynes of up to 60 carbon atoms. A study of resonance effects using natural resonance theory and bond lengths demonstrates lone-pair-donating effects that increase in the series F < Cl < Br < I, but run counter to the halogen inductive effects which decrease in this series and dominate energetic effects.

Functional specialization of calreticulin domains.

Calreticulin is a Ca2+-binding chaperone in the endoplasmic reticulum (ER), and calreticulin gene knockout is embryonic lethal. Here, we used calreticulin-deficient mouse embryonic fibroblasts to examine the function of calreticulin as a regulator of Ca2+ homeostasis. In cells without calreticulin, the ER has a lower capacity for Ca2+ storage, although the free ER luminal Ca2+ concentration is unchanged. Calreticulin-deficient cells show inhibited Ca2+ release in response to bradykinin, yet they release Ca2+ upon direct activation with the inositol 1,4,5-trisphosphate (InsP3). These cells fail to produce a measurable level of InsP3 upon stimulation with bradykinin, likely because the binding of bradykinin to its cell surface receptor is impaired. Bradykinin binding and bradykinin-induced Ca2+ release are both restored by expression of full-length calreticulin and the N + P domain of the protein. Expression of the P + C domain of calreticulin does not affect bradykinin-induced Ca2+ release but restores the ER Ca2+ storage capacity. Our results indicate that calreticulin may play a role in folding of the bradykinin receptor, which affects its ability to initiate InsP3-dependent Ca2+ release in calreticulin-deficient cells. We concluded that the C domain of calreticulin plays a role in Ca2+ storage and that the N domain may participate in its chaperone functions.

Ultrasound guided percutaneous drainage of abdominal abscesses.

OBJECTIVE: To see yield of ultrasound guided percutaneous drainage of abdominal abscesses. SETTING: Department of General Surgery Unit I, Lahore General Hospital and North Surgical Unit, Mayo Hospital, Lahore. DESIGN: A prospective case-controlled study. SUBJECTS: Patients with spontaneous or post operative abdominal abscesses. RESULTS: Overall success rate was 96.2%, major complications 11.5%, minor complications 26.9% and mortality 3.8%. In simple abscesses (20 patients), the success rate was 100% and in complex abscesses (six patients) 83.4%, with respective mortality 0.0% and 16.6%. Mean hospital stay was 10.7 days (range 3-25). CONCLUSION: Ultrasound guided percutaneous drainage of abdominal abscesses is an effective and safe method, suited particularly for simple abscesses.