Speakers, voice panels and voice-enabled appliances spend most of their time in standby, not recognizing speech. In standby the firmware has two jobs: keep the wake-word model somewhere it can be read immediately, and keep the PSRAM holding that model in a low-power mode. The two pull in opposite directions — holding the model whole takes density, while whole-device standby takes low current. APS12804O-SQRHX-SN meets both: 128Mb of density holds the model, and Halfsleep mode drops the memory's standby current to 19.5μA.
With the model in NOR, every recognition starts with a Flash read. Flash is far slower than memory, latency jitters, and the more false wake-up attempts occur, the more often the Flash chip is read and the more standby power climbs with it. Move the model into 128Mb PSRAM and it is read only once at power-up: the whole model is loaded into the 16MB PSRAM and kept there, and every inference afterwards reads from memory without touching Flash. A wake-word model is only about 300KB and its addresses are broadly contiguous, so a burst read from PSRAM takes about 4ms, the decision rhythm stays steady, and the same phrase no longer comes back fast one time and slow the next.
The 16MB is not all model space. On a voice panel with a display, a realistic split looks like this: 6MB for the wake-word and command models, 1.5MB for prompts and response audio, 512KB each for the echo-cancellation and noise-suppression double buffers, 1MB for UI sounds and language packs, leaving just over 6MB for OTA staging and run logs.
The ordinary standby current of APS12804O-SQRHX-SN is tens of μA at 25°C and up to 420μA maximum at 85°C. Halfsleep is the feature aimed at battery-type products: typically 19.5μA at 25°C, model data retained, immediate exit when the wake word arrives, no reload needed. The chip does not enter Halfsleep by itself; firmware must issue the command during idle. Skip that command and the part stays at the ordinary standby tier, and the 19.5μA never materializes. Across the whole device the gap between the two states is large: over one day of standby, Halfsleep consumes roughly 0.5mAh, while ordinary standby at the 85°C maximum costs about 10mAh — a twenty-fold difference. On dry-cell or coin-cell products, that difference decides whether the battery swap cycle is measured in months or years.
During Halfsleep the PSRAM still refreshes its array internally, so power must stay on. If you cut the 1.8V rail while idle to save power, the model has to reload after repowering and wake-up latency goes back to hundreds of milliseconds. If power-down is truly required, pair it with a NOR model image and a fast-load routine — switching the power off alone is not a recovery plan.
During audio playback the MCU fetches a batch of data for the codec at short intervals, and the batch size is set by the codec — commonly 32 or 64 bytes. APS12804O-SQRHX-SN lets you configure the fetch length as 16/32/64/2048. Match it to the codec's batch size and the audio plays continuously; mismatch it and a faint, evenly spaced noise creeps into the output, easy to misdiagnose as a codec or power problem.
The second gain is transfer speed. This chip reads 2KB in one continuous access, while common 1024-byte-page parts manage only 1KB. For large moves such as power-on model loading or whole-clip audio reads, interrupts are halved and the job finishes sooner. Response audio lives in the PSRAM and streams straight to the MCU during playback, with no second copy in MCU memory — the space saved goes to other system modules.
After plug-in or a battery swap, the user may speak right away. At 144MHz over four lines (about 72MB/s theoretical), reading a 6MB model takes roughly 83ms; with cache initialization and codec power-up, the whole sequence fits inside 200ms. The risk is not speed but order: sending commands before initialization completes returns invalid data, and the symptom is a wake word that never triggers on the first try. A gate in the power-up flow that requires the model CRC check to pass helps more than endless tuning of algorithm parameters. Battery products repeat this sequence at every boot, and users feel the difference between 200ms and 500ms directly. On the interface, both SPI and QPI modes are configurable: pin-constrained modules can start on SPI and switch to four-bit QPI when load speed matters, without changing the memory device.
| Item | Value |
|---|---|
| Ordering part number | APS12804O-SQRHX-SN |
| Density / organization | 128Mb, 16M×8bit |
| Interface frequency | 144MHz Wrapped Burst / 84MHz linear burst |
| Supply voltage | 1.62~1.98V |
| Operating current | 11mA @144MHz |
| Halfsleep current | typ. 19.5μA @25°C |
| Package | SOP-8 150mil |
With MCU SRAM alone the model does not fit; you are down to a few fixed command words and the product shrinks outright. Keep it in NOR and every inference pays a model read first — latency and current both suffer, and the buffer pressure of a four-mic array (four channels at 16kHz, 16-bit, 128KB per second) stacks on top. Holding the model whole plus low-current standby both come down to one 128Mb PSRAM.
Q: Can the SQRHX-SN part number go straight into the BOM?A: The ordering table lists only SQRH, SQRHX (bare die) and SQRH-SN (SOP-8, 85°C) as separate entries. The combination of 105°C plus SOP-8 is not listed separately; confirm the temperature definition and volume availability with the manufacturer before committing it to a production BOM.
Q: How big is the gap between Halfsleep and standby?A: Standby reaches 420μA maximum at 85°C; Halfsleep is typically 19.5μA at 25°C. Data is retained in both states; the difference is whether firmware explicitly issues the command.
Q: Is cutting the 1.8V rail while idle an acceptable way to save power?A: Power-down loses the model, wake-up needs a reload, and latency goes back to hundreds of milliseconds. If you must cut power, pair it with a NOR image and a fast-load routine.
Q: How much buffer does a four-mic array need?A: Four channels at 16kHz, 16-bit landing simultaneously produce 128KB per second, nearly four times a single mic. Multiply the density selection by the number of microphones.
Q: Is there room for future model iterations?A: Models grow ten to twenty percent per generation. After the model, audio and double buffers, 16MB still has 6MB left — enough for two or three generations plus OTA.
Need samples or the full datasheet for APS12804O-SQRHX-SN? Shenzhen Manduic Technology Co., Ltd. is an authorized distributor of AP Memory and can also arrange samples and documentation for other 105°C-grade parts in the same series.
For samples, product datasheets, detailed adaptation solutions or technical support, please contact us:
Tel: +86-150-1290-5940
Email: sales@manduic.com
Website: www.manduic.com
Telephone: +86-150-1290-5940
Mobile phone: +86-150-1290-5940
Mailbox: sales@manduic.com
Address: Room 1115, R&D Building, DAFU Technology, Fenghuang Sub‑district, Guangming District, Shenzhen City