All News | Boards | Chips | Devices | Software | LinuxDevices.com Archive | About | Contact | Subscribe
Follow LinuxGizmos:
Twitter Facebook Pinterest RSS feed
*   get email updates   *

Edge AI server packs in a 16-core Cortex-A72 CPU plus up to 32 i.MX8M SoCs and 128 NPUs

Jan 15, 2020 — by Eric Brown — 1126 views

SolidRun’s “Janux GS31 AI Inference Server” runs Linux on its CEx7 LX2160A Type 7 module equipped with NXP’s 16-core Cortex-A72 LX2160A. The system also supplies up to 32 i.MX8M SoCs for video and up to 128 Grylfalcon Lightspeeur 2803 NPUs via multiple “Snowball” modules.

When people talk about edge AI servers, they might be referring to some of the high-end embedded systems we regularly cover here at LinuxGizmos or perhaps something more server-like such as SolidRun’s rackmount form factor Janux GS31 AI Inference Server. The system would generally exceed the upper limits of our product coverage, but it’s a particularly intriguing beastie. The Janux GS31 is based on a SolidRun CEx7 LX2160A COM Express Type 7 module, which also powers the SolidRun HoneyComb LX2K networking board that we covered in June.



Janux GS31 AI Inference Server
(click image to enlarge)

In addition to this CPU module, the Janux GS31 integrates up to 4x “Snowball” modules, each with 8x NXP i.MX8M SoCs, a PolarFire MPF200T FPGA, and up to 32 Gyrfalcon Lightspeeur SPR2803 NPUs. These combine to create a vision processing subsystem with up to 32 i.MX8M SoCs and 128 NPUs that supports “ultra-low latency transcoding and video analytics” on up to 128 channels of 1080p60 video. Applications include monitoring smart cities, infrastructure, intelligent enterprise/industrial video surveillance, object detection, recognition & classification, and smart visual analysis.

SolidRun had no announcement, pricing, or availability for the “coming soon” Janux GS31 AI Server — which, let’s face it, you probably couldn’t afford anyway. However, CNXSoft spotted a product page with substantial details.


HoneyComb LX2K
with CEx7 LX2160A

The CEx7 LX2160A, which is now available separately, starts at $550. The Type 7 module is built around NXP’s headless, up to 2.0GHz, 16-core Cortex-A72 Layerscape LX2160A SoC. There is also 64GB of dual-channel DDR4. It’s unclear if the Janux GS31 offers the optional up to 64GB eMMC.

SolidRun has yet to publish specs for the Snowball module, but the Janux GS31 spec list and block diagram help fill in the gaps. The Snowball incorporates NXP’s widely deployed, mid-range i.MX8M, a quad-core, Cortex-A53 SoC at up to 1.5GHz with a Cortex-M4 MCU and 3D-ready Vivante GC7000 Lite GPU with 4K support.

Each of the up to 4x Snowball modules offers 4x i.MX8M SoCs, which is the largest clustering of an i.MX SoC that we’ve ever seen. SolidRun also deploys the i.MX8M on its i.MX8 SOM and HummingBoard Pulse SBC.



Janux GS31 block diagram
(click image to enlarge)

Each Snowball module also provides 32 Lightspeeur SPR2803 Neural Accelerators (or Lightspeeur 2803) chips from Gyrfalcon Technology, Inc (GTI). SolidRun uses the NPU on its i.MX 8M Mini SOM, which is also available on the HummingBoard Pulse.

On the Janux GS31, the Lightspeeur 2803 is listed as offering 2150 tera operations per second @ 300MHz. According to the GTI product page, the 2803 operates at 24 TOPS/W per chip, with optimal peak performance at 16.8 TOPs @ 300MHz and a low-power mode of 16.8 TOPs @ 700mW. The Snowball module supports AI frameworks including TensorFlow, Caffe, and PyTorch and AI networks including VGG, ResNet, and MobileNet.

Since the block diagram describes each module as having two blocks of 16 NPUs, it’s possible SolidRun is using GTI’s Gainboard 2803, a server-oriented PCIe accelerator card with 16 Lightspeeur 2803 NPUs. GTI also offers a lower-power Lightspeeur 2801S NPU, which is used on Shenzhen Xunlong’s Orange Pi 4B SBC and Orange Pi AI Stick Lite USB stick.

Finally, the Snowball modules provide Microchip’s Microsemi PolarFire MPF200T FPGA, a mid-range PolarFire with 192K logic elements. Dual PCIe switches are also on the module.

The Janux GS31 AI Inference Server is further equipped with a single GbE, 2x 10GbE SFP+, 2x USB host, and a micro-USB port. The block diagram also shows a PCIe x16 edge socket, which may be used to connect the Snowball boards and therefore may not be available. There’s also an unspecified amount of NVMe storage via an M.2 2280/22110 socket, as well as an OLED screen.

The 1U rackmount system measures 500 x 445 x 43.5mm and consumes a maximum of 900W via a 100-240V power supply. Even with 8x PWM-controlled fans, there’s a narrow 10 to 35°C operating range and 20% to 80% relative, non-condensing humidity resistance.

 
Further information

No pricing or availability information was provided for the Janux GS31 AI Inference Server. More information may be found on SolidRun’s product page.

 

(advertise here)


Print Friendly, PDF & Email
PLEASE COMMENT BELOW

Please comment here...