Musaib Mohiuddin Einheit
ShopAccount
Concept render of the SATYRN X1 processor package

SATYRN · First processor · In design

SATYRN X1

The group’s first in-house AI processor, and the compute infrastructure built to run it. SATYRN is the program; X1 is its first chip.

Concept visualization

The product

SATYRN X1

Artificial intelligence runs on processors, and most companies rent them. SATYRN is the group’s program to design its own and to run them in its own data halls. X1 is the first: one design for two uses, the BAPH robots and the SATYRN halls.

Concept render of the SATYRN X1 package: brushed black heat spreader marked SATYRN X1, four memory stacks and rows of capacitors on a dark substrate
  1. Compute dieBeneath the integrated heat spreader marked SATYRN X1.
  2. Memory stacksFour stacks of high-bandwidth memory beside the die.
  3. Power deliveryRows of decoupling capacitors steady the supply.
  4. SubstrateFlip-chip package on a multilayer organic substrate.
Sheet SX-001. SATYRN X1 package. Design study. No chip has been manufactured.

Architecture

Inside X1

A die floor plan and the data-hall module, drawn to the current architecture study. Layouts are illustrative and will change as the design matures.

Concept render of the SATYRN X1 die: a grid of compute tiles, memory banks, interface strips and a ring of bond pads
  1. Tensor compute arraySixteen compute tiles in the floor-plan study.
  2. SRAM banksOn-chip memory placed beside each group of tiles.
  3. Network on chipThe spine that moves data between tiles and memory.
  4. Memory interfacePhysical links to the four memory stacks.
  5. Chip-to-chip linksFor joining many X1 chips into one machine.
  6. Safety islandA separate core that enforces the stop command; compute cannot override it.
  7. I/O ringBond pads around the edge of the die.
Sheet SX-002. Die floor plan, illustrative. Block sizes not to scale.

Design targets

Preliminary targets from the architecture study. Figures marked under evaluation will be set before the design is committed to silicon.

Stage
Architecture definition. Not taped out.
Class
AI accelerator for training and inference
Compute
Sixteen tensor compute tiles on a network on chip (floor-plan study)
Number formats (target)
FP8, BF16 and INT8, with FP32 accumulation
Memory
Four stacks of high-bandwidth memory on the package
On-chip memory
SRAM banks beside each group of tiles
Host link
PCI Express class
Scale-out
Chip-to-chip links for multi-accelerator machines
Safety
Independent safety island; a hardware stop path that compute cannot override
Package
Flip-chip with integrated heat spreader; memory integrated beside the die
Cooling
Direct liquid cold plate in the data hall
Form factors (plan)
Accelerator module for SATYRN halls; embedded module for BAPH units
Process node
Under evaluation
Power envelope
Under evaluation
Manufacturing
Contract foundry. No foundry selected.
Concept render of the SATYRN X1 accelerator module with its copper liquid cold plate lifted above the board
  1. Cold plateCopper plate for direct liquid cooling in the data hall.
  2. Coolant fittingsQuick-disconnect couplings for the coolant loop.
  3. X1 packageThe processor at the centre of the module.
  4. Memory stacksHigh-bandwidth memory on the package.
  5. Voltage regulatorsPower stages convert board power for the chip.
  6. ConnectorsHigh-density connectors to the host board.
Sheet SX-003. Data-hall module with liquid cold plate.
Concept render of a 300 millimetre wafer of processor dies held in a cleanroom under amber light
  1. WaferEach square on a 300 mm wafer is one die.
  2. CleanroomChips are made in a contract foundry, not in-house.
Sheet SX-004. Illustration of a production wafer.

Development path

From specification to silicon

The standard path for a new processor. Each step is completed and reviewed before the next begins.

  1. 01ArchitectureInstruction set, tile design, memory system and safety island specified.Now
  2. 02RTL and verificationThe chip written in hardware description language and tested in simulation.Next
  3. 03FPGA prototypeA scaled-down X1 running on programmable hardware.Next
  4. 04Test chipA small test die on a shared multi-project wafer run.Next
  5. 05X1 tape-outThe full design released to a contract foundry.Next
  6. 06Bring-upFirst silicon tested; modules for BAPH and the data halls.Next

Where it runs

Data hall locations

X1 is designed to run in data halls owned by the group. Locations are ranked in order of preference. Exact sites are not disclosed, and no land has been acquired or leased.

  1. First preference

    Hyderabad

    India

    The group’s base in India: one of the country’s principal data center markets, with deep engineering and chip-design talent nearby.

  2. Second preference

    Karachi

    Pakistan

    Pakistan’s largest commercial city and the landing point of its undersea data cables.

  3. Third preference

    Cold highlands

    Afghanistan

    A high, naturally cold region. Cold outside air can cool the halls for much of the year, cutting the water and electricity spent on cooling.

Concept render of a low, dark data hall with walls of louvred air intakes in a snow-covered mountain valley at dawn
  1. Fresh-air intakesLouvred walls draw in cold outside air.
  2. Cold climateAltitude keeps the outside air cold for much of the year.
  3. Power feedA dedicated high-voltage line to the site.
Sheet ST-001. Third-preference concept: a hall cooled by cold outside air. Not a real site.

Purpose

Strategic rationale

  1. Own the processor

    The group’s programs, from BAPH to AEROHOP, will depend on artificial intelligence. Designing the processor keeps that work on silicon the group controls.

  2. Own the compute

    X1 is intended to run in halls the group owns, under its own governance, instead of rented capacity.

  3. Built hall by hall

    Each hall is a repeatable module of racks, cooling, power and network. One is built and proven before the next begins.

  4. The group first

    The first user is the group itself. Capacity may be offered to third parties at a later stage, where the law permits.

Timeline

Status and roadmap

  1. Nov 2026Program initiated and named SATYRN.Done
  2. Nov 2026SATYRN X1 named as the first processor; package, die and module sheets published.Done
  3. Nov 2026Data hall locations ranked: Hyderabad, Karachi, then the cold highlands of Afghanistan.Done
  4. NextX1 architecture specification.Next
  5. NextStudy of the first-preference site: land, power, cooling, network and licences.Next
  6. NextLegal review for each country, including export-control and sanctions rules.Next

Legal

Notice

SATYRN is a program in development. SATYRN X1 is a design study: no chip has been designed to completion, taped out or manufactured, and images of the processor, module, wafer and data hall are concept renders, not photographs of built hardware. No data center has been built, leased, bought or operated, and no site has been chosen. Any chip would be made by a contract foundry under licence terms not yet agreed. Every site depends on funding, land, power, local law and licences, and on United States export-control and sanctions rules, any of which may delay, change or prevent it. Plans and timelines are current intentions, not promises, and may change or be abandoned. See the forward-looking notice.