Mixtile Aims at AI Performance with the Core 3588E, a Drop-In Alternative to NVIDIA's Jetson TX2 NX

Embedded and cluster computing specialist Mixtile has announced the launch of a new system-on-module, designed as a drop-in replacement for carrier boards compatible with the NVIDIA Jetson TX2 NX — and offering a claimed six tera-operations per second (TOPS) of compute for on-device machine learning and artificial intelligence (ML and AI) workloads.

"Powered by RK3588's cutting-edge quad-core CPU architecture, Mixtile Core 3588E packs a punch in a compact system-on-module (SOM)," the company writes of its latest hardware launch. "With blazing fast 6 TOPS AI performance enabled by the Cortex-A76 and Cortex-A55 cores, Core 3588E readily handles demanding workloads like high-res video encoding/decoding, graphics processing, and running AI apps."

The Rockchip RK3588 SoC features eight high-performance Arm Cortex-A76 cores running at up to 2.4GHz and four lower-power Cortex-A55 cores running at up to 1.8GHz, an Arm Mali-G610 MC4 graphics processor and 2D acceleration module, and a triple-core neural processing unit (NPU) coprocessor which delivers a claimed 6 TOPS of compute performance at INT4 precision with support for INT4, INT8, INT16, FP16, BF16, and TF32 precision modes.

To this, Mixtile has added a choice of 4GB, 8GB, 16GB, or an impressive 32GB of LPDDR4 memory and 32GB, 64GB, 128GB, or 256GB of on-board eMMC 5.12 flash storage. There's hardware codecs for 8k60 video decoding in H.264, H.265, and VP9 formats and 8k30 encoding in H.264 or H.265, an 8k60-capable HDMI output and a DisplayPort/Embedded DisplayPort combination interface supporting 4k60.

For peripherals, the SOM offers three four-lane or five two-lane MIPI Camera Serial Interfaces (CSIs) connecting to a 48 megapixel image signal processor (ISP) with high dynamic range (HDR) support, one USB 3.0 Gen. 1 and three USB 2.0 lanes, a PCI Express Gen. 3 x4 and a PCI Express Gen 2.1 x1 lane, and a single gigabit Ethernet port.

The edge connector also brings out 15 general-purpose input/output (GPIO) pins, three pulse-width modulation (PWM) channels, two SPI, four I2C, four I2S, and one CAN bus, one debug UART and two UARTs with flow control, SD 4.0, SD Host 4.0, and SDIO 3.0 connectivity.

Given the performance of the CPU, GPU, and NPU, there's no surprise to see Mixtile is pushing the Core 3588E for on-device edge-AI applications — including running large language models (LLMs) on-device. The company claims a single Core 3588E can handle Bloom with a 1.4 billion parameter size at a performance up to six tokens per second, or ChatGLM with six billion parameters at one token per second.

The more demanding LLaMA, recently released by Facebook parent Meta's AI division, can run with a 65 billion parameter size at one token per second, too — if you're willing to network 65 Core 3588E boards in parallel.

Mixtile is taking orders for the SOM on its website at $109 for the 4GB/32GB variant, $159 for the 16GB/128GB version, and $269 for the 32GB/256GB model, with an optional active heatsink available for an additional $9 — all representing an 18 per cent discount for pre-order customers. The company has not yet confirmed pricing for the promised 8GB/64GB version.