The conclusion first. If the host controller carries an OPI or Xccela interface, the design needs around 32MB of buffer memory, and the project does not want to pay the routing and power cost of DDR, the APS256XXN-OB9-BG is worth trying first. In edge AI and screen-based HMI designs, memory is the link most often underestimated — internal SRAM runs out quickly once a lightweight model and a mid-resolution display are both in play, while external DDR means another layer of PCB routing, another power domain and timing closure. PSRAM exists precisely for this gap: an interface as simple as SRAM with capacity in the tens to hundreds of megabits. Below is the specification profile of this part and the points that matter when using it.
Capacity: 256Mb, or 32MB
Supply: single rail 1.62V to 1.98V, nominal 1.8V; VDD and VDDQ share the same rail, so the IO level lines up directly with the host controller without level shifting
Interface: Octal Peripheral Interface (OPI) and Hexadecimal Peripheral Interface (HPI), operating in Xccela mode; the mode is set by bit MR8[6] of the mode register and defaults to X8
Speed: clock up to 250MHz; X8 mode transfers two bytes per clock for 500MB/s of read/write throughput, X16 mode transfers two words per clock for 1000MB/s. That bandwidth is generous at the MCU tier — refreshing an 800×480 panel or caching weights and intermediate results for a lightweight neural network will not become the bottleneck
Package: BGA24B, 6mm × 8mm × 1.2mm
Temperature range: the APS256XXN-OB9-BG is rated -40°C to +85°C; the APS256XXN-OBx9-BG extends to -40°C to +105°C
Supply limits: absolute maximum for VDD and VDDQ is -0.4V to +2.45V — exceeding 2.45V causes permanent damage; signal overshoot is allowed to -1.0V on the negative side and VDD +1.0V on the positive side, but not for more than 20ns, so routing impedance control matters
Other: storage temperature -55°C to +150°C; package pin capacitance typically 5pF input and 6pF output, with system load capacitance kept within 15pF
| Parameter | X8 mode (default) | X16 mode |
|---|---|---|
| Total capacity | 256Mb | 256Mb |
| Organization | 32M × 8bit | 16M × 16bit |
| Page size | 2048 bytes | 1024 words |
| Column address | AY0 – AY10 | AY0 – AY9 |
| Row address | AX0 – AX13 | AX0 – AX13 |
Capacity is identical in both modes; what changes is the data width and the page size. X8 has a larger page and more address bits, suiting mostly sequential data buffering. X16 moves twice the data per clock, suiting bandwidth-sensitive graphics or matrix workloads.
One package-related note: it carries every signal needed for both X8 and X16 modes; if the design only runs X8, some pins are unused and are explicitly marked DNU in the datasheet, so they can be left unrouted.
Reset. Use the RESET# pin or a global reset command — the datasheet gives complete sequences for both.
Set the width. At init, write the matching value to bit MR8[6] of the mode register to switch between X8 and X16; the default is X8.
Enter a low-power mode. Halfsleep is entered by writing 8'hF0 to MR6 and holding CE# high for at least the minimum tHS; Deep Power Down is entered by writing 8'hC0 to MR6.
Mode switching is done once at initialization, so one part serves either need per project.
| Mode | How to enter | Data retained | Average current | Typical use |
|---|---|---|---|---|
| Halfsleep | Write 8'hF0 to MR6, hold CE# high for ≥ tHS | Yes | Below 100μA | Screen-off, voice standby |
| Deep Power Down | Write 8'hC0 to MR6 | No — registers reset to defaults after exit | Lowest | Overnight standby, before extended power-off |
Halfsleep exits on a low pulse of CE#, after which the part is ready once tXHS has elapsed; data is fully retained and wake-up is fast, which suits screen-off or voice-standby states that must recover instantly. Deep Power Down pushes power to the minimum, but neither the register values nor the memory contents are retained, so the firmware must re-initialize after exit — it therefore suits deep sleep only.
Although average Halfsleep current is small, the internal periodic refresh produces a peak current of up to 25mA lasting a few tens of microseconds. If the system regulator cannot supply that transient peak, the rail sags and data errors follow. The datasheet's recommendation is a 4.7μF to 10μF capacitor dedicated to covering the refresh current, plus a low-ESR capacitor of no more than 1μF placed close to the device to absorb fast transients. These two capacitors are not optional — leave footprints for them on the board.
The first category is edge AI inference. When an MCU or a lightweight NPU runs keyword spotting, face detection or simple image classification, internal SRAM cannot hold the weights and intermediate tensors, and attaching a 256Mb PSRAM as unified memory is the simplest fix — Xccela bandwidth is sufficient and interface complexity is far below DDR.
The second is screen-based HMI. Appliance panels, industrial HMIs and EV-charger displays often sit on a host with only a QSPI interface; the frame buffer, font and image assets all live in PSRAM, and 32MB is ample for double buffering plus animation assets.
The third is communication and gateway equipment that needs a mid-size data buffer without paying the routing and power cost of DDR.
If the host only has a QSPI interface, a QSPI part such as the APS6404L is the better fit — OPI needs eight data lines and will not connect without them, while the QSPI version still delivers half of that 500MB/s at 250MHz. On voltage, this is a 1.8V device suiting low-power, low-voltage hosts; the difference between 1.8V and 3.3V versions is mainly IO level and power, and the 3.3V version connects directly to conventional MCU IO levels. On capacity, 32MB usually holds the weights plus a few layers of intermediates for lightweight-model edge devices; for high-resolution video buffering or larger models, move up to 512Mb or beyond.
Why does Halfsleep still need dedicated capacitors? Because refresh current is pulsed — average current is small but peaks can reach 25mA, faster than the regulator can respond, so a nearby energy-storage capacitor has to carry those few tens of microseconds.
How is it sourced in China? AP Memory PSRAM is supplied in China through distributors. Shenzhen Manduic Technology Co., Ltd. is within that distribution network, and samples, datasheets and technical support can be requested through it.
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