← systems
Datacenter · AI

Open AI server node

The apex system — where the host is settled reuse, and the cost and the joules concentrate 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 as settled reuse. The cost and the differentiation concentrate in two places — the accelerator silicon and the links between accelerators — and both are an energy problem.

6 of 7 settled or cyclical — 1 on the frontier. Each part placed by cost and differentiation — the landscape, not a verdict. The true cost of building without knowing the market and the benefit, surfaced — so you choose by your goal.

Exploded

Decompose the product into its stack. Each layer is colour-coded: settled, cyclical, or frontier.

48 V ORv3 power shelf Settled
CPU tray (2× Xeon) + DDR5 Cyclical
8× OAM accelerators (UBB) Frontier
OCP NIC 3.0 + E1.S NVMe Cyclical

Bill of materials

Settled ×2 Cyclical ×4 Frontier ×1

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

OAM ×8 (H100 / MI300X) OCP form factor Frontier · higher cost and risk — differentiation, SOTA, or science
AI accelerators — 8 modules on a Universal Baseboard

Where the value and the joules concentrate. Frontier options: the compute die itself, and co-packaged optical I/O so bits move between chips in light rather than copper — that interconnect is the photonic tile from the frontier build, in the rack. Buy off-the-shelf accelerators, or design here for energy and bandwidth.

https://www.opencompute.org/documents/ocp-accelerator-module-design-specification-v1p5-final-20220223-docx-1-pdf →
Xeon Sapphire Rapids ×2 Intel Cyclical · refreshes on a clock
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 Cyclical · refreshes on a clock
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 Settled · mature — lowest cost and risk
Management — server BMC running OpenBMC

Settled management SoC. Reuse.

https://www.aspeedtech.com/server_ast2600/ →
OCP NIC 3.0 OCP spec Cyclical · refreshes on a clock
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 Cyclical · refreshes on a clock
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 Settled · mature — lowest cost and risk
Power — 15 kW/shelf with BBU

Standard rack power. Reuse.

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

Where openWafer can help

In an AI node the host — CPU, DRAM, NIC, BMC, power — is settled reuse; OCP standardised all of it. The cost and the differentiation concentrate in two places: the accelerator silicon and the links between accelerators. The frontier options are the compute die itself and co-packaged optical I/O — moving bits in light rather than copper, for energy and bandwidth — which is the photonic tile from the frontier build, in the rack. Buy accelerators off the shelf, or design where the joules are: both are real strategies with different cost, risk, and payoff, and the right one follows from the goal.

Open sources — decomposed and linked, not hosted