HalcyonEmbargoed / 01 — 08
Halcyon N1

Inference at the edge of a watt

A 5 nm accelerator that runs 70B-parameter models in a single PCIe card, at a third of the power.
Process5 nm
Memory96 GB HBM3e
Power150 W
ShipsQ2 2028
HalcyonEmbargoed / 02 — 08
Why we built it
Data centres now spend more on power than on the chips that draw it.
Halcyon analysis of public operator filings, 2026–2027
HalcyonEmbargoed / 03 — 08
First silicon
Back from the fab, working
184 dies on the first wafer, every one passing functional test on the first run.
Halcyon / tokens per joule against today's leading cardEmbargoed / 04 — 08
Efficiency
3.2×
tokens per joule against today's leading card
Measured on a 70B model at batch 32, card power only.
Halcyon lab measurements, October 2027
HalcyonEmbargoed / 05 — 08
The card
N1 at a glance
Compute1.6 PFLOPS FP8dense, no sparsity
Memory96 GB HBM3e4.8 TB/s
InterconnectPCIe 6.0 x16plus 4 N-Link ports
Power150 Wpassive, single slot
PrecisionFP8 · INT4 · BF16per layer
Form factorFHHL cardfits a 1U server
HalcyonEmbargoed / 06 — 08
Software
Your model, unchanged
Compile from ONNX or PyTorch; the runtime speaks the OpenAI-style API your apps already call.
halcyon-sdk
$ pip install halcyon
$ halcyon compile llama-70b.onnx --target n1 --fp8
partitions 1 card · 80 layers
memory 71.4 GB of 96 GB
✓ compiled in 3 min 12 s
 
$ halcyon serve ./llama-70b.n1 --port 8080
serving http://0.0.0.0:8080/v1
first token 182 ms · 41 tok/s per user
HalcyonEmbargoed / 07 — 08
Scale out
From one card to a rack
01Hostany x86 or Arm server
02N1 cardone model per card
03N-Linkeight cards, one fabric
04Rack64 cards in 9.6 kW
HalcyonEmbargoed / 08 — 08
Sampling now
Run it on less
halcyon.example/n1early@halcyon.example
Dev kitJanuary 2028
VolumeQ2 2028
BoothHall C · 2214
Presspress@halcyon.example