NVIDIA has officially launched the Jetson AGX Orin, its latest-generation embedded AI platform — and since its unveiling last year has found a way to boost performance from 200 TOPS at INT8 to a whopping 275 TOPS.
Powered by NVIDIA's latest-generation graphics hardware, the Jetson AGX Orin promises "server-class AI performance at the edge" — and comes with the specifications to deliver. The company promises up to 275 TOPS of INT8 compute, up from 32 TOPS in the last-generation Jetson AGX Xavier — but not all of that increase is a generational gain.
Ampere, the GPU technology powering the AGX Orin, supports a new feature missing from its predecessors: Sparse Tensor Cores, an upgraded version of the AI-focused Tensor Cores found in previous generation GPUs. When running specially-prepared sparse networks — which, NVIDIA claims, are rapidly becoming the de facto standard thanks to the performance gains on offer — the AGX Orin hits its headline-grabbing 275 TOPS at INT8; using a traditional dense network, however, limits peak performance to 138 TOPS.
There's a certain amount of brute force to the increase in compute performance, too: Where the AGX Xavier could be configured to draw from 10W up to 30W, the AGX Orin goes from 15W to 60W — twice the AGX Xavier's maximum draw, pulling its successor's performance-per-watt gains down considerably.
The heart of the AGX Orin is a 2,048-core Ampere GPU with 64 Tensor Cores, two NVIDIA Deep-Learning Accelerators (NVDLAs), and a single PVA v2 vision accelerator — replacing two PVAs on the AGX Xavier. The CPU is a 12-core Cortex-A78AE chip with 3MB of L2 and 6MB of L3 cache, connected to 32GB of LPDDR5 memory offering up to 204.8GB/s throughput — and configurable up to 64GB, if you're ordering the AGX Orin module rather than the developer's kit.
There's 64GB of on-board eMMC storage,up from 32GB on the AGX Xavier, and hardware encode and decode for up to two 4k60 and one 8k30 video streams respectively - half that, oddly, achievable on the AGX Xavier. There are 16 lanes of MIPI CSI-2 with 16 virtual channels, again half that of the AGX Xavier, and 22 lanes of PCI Express Gen. 4, eight of which are brought out to a full-size PCIe slot with another four each found on M.2 Key M and Key E slots. An on-board Ethernet port provides wired connectivity, with four 10-gigabit XFIs on the module itself, with a bundled M.2 Wi-Fi card offering wireless support.
In short, the AGX Orin is a beefy upgrade — but it's also a costly one. Previously, NVIDIA has subsidized the cost of its developer kits as a means of boosting interest in the Jetson ecosystem; with a claimed million-plus developers, 6,000 commercial customers, and over 130 ecosystem partners, it would appear those days are now over. The AGX Orin Developer's Kit is priced at $1,999 — a big leap over the AGX Xavier's launch price of $1,299 for those registered in the NVIDIA Developer Program, though cheaper than its $2,499 recommended retail price at launch.
For your money, though, NVIDIA is promising impressive performance with the hope of more to come: The company claims that, based on incremental improvements seen during the lifespan of the AGX Xavier through driver and software optimisation, the overall performance of the AGX Orin could be boosted by around 1.5x before its inevitable replacement by the next generation.
The Jetson AGX Orin Developer's Kit is available to order now at $1,999, though NVIDIA warns that demand is likely to exceed supply for some time after launch.
In Q4 the company will launch the module-only variants, the AGX Orin 32GB with an eight-core CPU and 200 TOPS peak performance at $899 and the AGX Orin 64GB with a 12-core CPU and 275 TOPS peak performance at $1,599. These join two non-AGX models, the Orin NX 8GB with a six-core CPU and 70 TOPS at $399 and the Orin NX 16GB with eight-core CPU and 100 TOPS at $599.
More information is available on the NVIDIA website.