← systems
Datacenter · AI

Open AI server node

The apex system — where the host is all reuse, and the value and the joules live in two places.

baseline: OCP Grand Teton (Meta) — Open Compute contribution
host 2× Xeon Sapphire Rapidsmemory ≤ 2 TB DDR5accel 8× OAMpower 48 V ORv3

Grand Teton is an open AI platform: dual-socket host, eight accelerators on a baseboard, 48 V power. OCP standardised the entire host. What is left to design is exactly the value — the accelerator silicon and the wires between accelerators — and both are an energy problem.

6 of 7 parts are reuse — 1 is the slice worth designing. That ratio is the NRE bill. The descent below is how you find it before you pay it.

Exploded

Decompose the product into its stack. Each layer is colour-coded by what it is to the next build.

48 V ORv3 power shelf Common · stable
CPU tray (2× Xeon) + DDR5 Common · cyclic
8× OAM accelerators (UBB) Opportunity
OCP NIC 3.0 + E1.S NVMe Common · cyclic

Bill of materials

stable ×2 cyclic ×4 Opportunity ×1

Every part is a real, sourced component from the open baseline — no invented part numbers.

OAM ×8 (H100 / MI300X) OCP form factor Opportunity · the slice openWafer designs
AI accelerators — 8 modules on a Universal Baseboard

The value and the joules. openWafer’s slice is both the compute die and co-packaged optical I/O — so bits move between chips in light, not copper, where a moved bit costs the fewest joules. That optical interconnect is the photonic tile from the frontier build, dropped into the rack.

https://www.opencompute.org/documents/ocp-accelerator-module-design-specification-v1p5-final-20220223-docx-1-pdf →
Xeon Sapphire Rapids ×2 Intel Common · cyclic · refreshes each generation — reuse the current one
Host CPU — up to 56 cores/CPU, 350 W TDP each (AMD EPYC tray also specced)

Standard server CPU — refreshes on a node clock. Reuse the current generation.

https://www.opencompute.org/documents/grand-teton-intel-based-cpu-tray-specification-v1-0-pdf →
DDR5 ≤ 2 TB Common · cyclic · refreshes each generation — reuse the current one
System memory — 32 DIMMs, 8 channels/CPU, 2 DPC

Standard DIMMs. Reuse the current generation.

https://www.opencompute.org/documents/grand-teton-intel-based-cpu-tray-specification-v1-0-pdf →
AST2600 ASPEED Common · stable · reuse as-is — zero NRE
Management — server BMC running OpenBMC

Settled management SoC. Reuse.

https://www.aspeedtech.com/server_ast2600/ →
OCP NIC 3.0 OCP spec Common · cyclic · refreshes each generation — reuse the current one
Network — PCIe 5.0 x16, hot-insertable

Standard NIC form factor. Reuse.

https://www.opencompute.org/documents/grand-teton-intel-based-cpu-tray-specification-v1-0-pdf →
E1.S NVMe Common · cyclic · refreshes each generation — reuse the current one
Storage — 8× front-accessible, + M.2 boot

Standard NVMe. Reuse.

https://www.opencompute.org/documents/grand-teton-intel-based-cpu-tray-specification-v1-0-pdf →
ORv3 48 V shelf + busbar OCP Common · stable · reuse as-is — zero NRE
Power — 15 kW/shelf with BBU

Standard rack power. Reuse.

https://engineering.fb.com/2022/10/18/open-source/ocp-summit-2022-grand-teton/ →

The slice openWafer designs

In an AI node the host is all reuse — CPU, DRAM, NIC, BMC, power; OCP standardised every bit of it. The value and the energy concentrate in two places: the accelerator silicon and the wires between accelerators. openWafer’s slice is both — the compute die, and co-packaged optical I/O so bits move between chips in light instead of copper, where a moved bit costs the fewest joules. That optical interconnect is the photonic tile from the frontier build, dropped into the rack. The datacenter is where the wrist and the frontier meet the same law: design where the joules are.

Open sources — decomposed and linked, not hosted