ZEVOXION // FIRST SILICON
Meet HF-1.
Our first AI accelerator, designed from the die up. One card runs models with hundreds of billions of parameters, and answers start almost instantly.
02 — ANATOMY
Six layers. Zero wasted space.
- 01 Lid
- 02 Compute die
- 03 HBM3E stacks
- 04 Silicon interposer
- 05 Package substrate
- 06 Ball grid array
03 — THE DIE
Every block has a job.
Hover or tap any part of the die to see what it does.
HF-1 · A0 · 812 mm²
04 — SPEED
Answers faster than you can read.
The same prompt on a typical GPU server and on HF-1. Press run.
BASELINE · TYPICAL GPU0 tok/s · — first token
ZEVOXION HF-10 tok/s · — first token
SIMULATED DEMO · ILLUSTRATIVE FIGURES
05 — THE CARD
Drop it into the servers you already have.
HF-1 ships on a standard full-height, full-length PCIe card. Hover or tap the markers.
06 — SPECIFICATIONS
The datasheet.
0TFLOPS
FP8 DENSE COMPUTE
0GB
HBM3E MEMORY
0TB/s
MEMORY BANDWIDTH
0W
BOARD POWER
- Process
- 3 nm-class
- Compute
- 48 tensor clusters · 1,000 TFLOPS FP8 · 2,000 TFLOPS FP4
- Precisions
- FP32 · BF16 · FP8 · FP4 · INT8
- On-chip memory
- 256 MB SRAM
- Memory
- 96 GB HBM3E · 4 stacks · 4.8 TB/s
- Host interface
- PCIe 5.0 ×16
- Scale-out
- HF-Link · 800 GB/s card-to-card · up to 8 cards
- Form factor
- Full-height, full-length, dual-slot PCIe card
- Cooling
- Passive, using server airflow
- Board power
- 350 W max · 12V-2×6 connector
PRELIMINARY TARGET SPECIFICATIONS. SUBJECT TO CHANGE.