NewIntroducing XPU Grasshopper. Pre-orders are open.

FPGA prototype devkit

XPU Grasshopper

The Gen 1 architecture on your desk, before silicon. End to end, programmable down to the instruction.

XPU Grasshopper (KU040) is a PCIe devkit carrying the Grasshopper architecture. It is for building on the new architecture today, and it is not meant for production serving.

$9,599$11,999

pre-order price

  • LIMITED EDITION, N of 100
  • 20% off for pre-orders
  • Early cluster access, batch #0
  • Open-source runtime, driver, and libxpu
Pre-order for $9,599

Ships February 2027*. $960 refundable deposit.

Read the announcementGen 1 enterprise and cluster orders

Architecture

GPUs were made for graphics. TPUs for matmuls. XPUs for frontier AI.

Agentic workloads

LLMs in agent harnesses run long, stateful trajectories that are irregular and sparse, a poor fit for chips built around a steady stream of dense matmuls.

Reconfigurable, simple

The dataflow reconfigures for each model instead of freezing one paradigm into silicon, and it is simple enough for AI agents to program and tune across the whole cluster.

Perf per dollar, per watt

Moving a byte costs far more than computing on it, so the data stays close to the math and the design optimizes for performance per dollar and per watt.

FOSS software stack

Own your intelligence

A fully open-source stack.

In the UNIX tradition, the hardware is a file. The whole stack above it ships open source, and you control the device completely from any language.

Layer 1

Write the kernel

You, a compiler, or an AI writes it in any language, ahead of time or just in time. The result is one binary with every MTU's program and where its inputs and outputs live.

gpt2.py
weights = Buffer(WEIGHTS).at(g0.dram[0])
tokens = Buffer(256, base=0).at(g0.spm[0])
next = Buffer(4, base=256).at(g0.spm[0])
@kernel
def gpt2(x):
# one hardware loop
for _ in range(12):
heads = (
attention.at(0, 0)
| attention.at(1, 0)
| attention.at(0, 1)
| attention.at(1, 1)
)
x = (heads >> mlp.at(0, 0))(
x, weights
)
return argmax(x, next)
graph = gpt2.graph(embed(tokens))
kernel = graph.assemble(g0)

Specs

XPU Grasshopper (KU040)

Two AMD KU040 FPGAs linked by SerDes, running the Grasshopper architecture. Use it as an XPU, or as a regular FPGA board.

The devkit

ArchitectureGrasshopper 2x2, 4 MTUs
Host interfacePCIe Gen3 x8
InterconnectQSFP-DD, degree 4
SoftwareOpen-source runtime, driver, and libxpu
EditionLimited edition, Gen0, N/100
ShipsFebruary 2027

The FPGAs

Devices2x AMD Kintex UltraScale XCKU040
LinkSerDes
Process20 nm
Logic cells1,060,500
DSP slices3,840
Block RAM42.2 Mb

Timeline

Where it is going

It runs on FPGA today. Next comes a cluster you can use, then boards for local prototyping, then silicon, then one machine.

  1. Grasshopper (VU47P) E2E

    Inference, in-house

  2. XPU (FPGA) Cluster

    Serving API

  3. XPU Grasshopper (KU040)

    Pre-orders open, ships February 2027

  4. Gen 1 silicon

    Tapeout

  5. Monolith

    Frontier AI computer

Pre-order XPU Grasshopper (KU040)

The first run is 100 numbered boards. Pre-orders get 20% off and batch #0 of early API and SSH access to the cluster.

Limited edition

XPU Grasshopper (KU040)

Limited edition, boards numbered 1 to 100.

The KU040 board runs the same 2x2 mesh as the benchmarked VU47P version.

$9,599$11,999

20% off when you pre-order

Pre-order

$960 deposit today, fully refundable

  • LIMITED EDITION, N of 100
  • 20% off for pre-orders
  • Early cluster access, batch #0
  • Open-source runtime, driver, and libxpu
  1. 1The board ships February 1, 2027. Component supply or manufacturing delays may move that date, and if they do, you will hear from us first.
  2. 2The deposit is fully refundable until the board ships. The balance is due when the board is ready to ship.
  3. 3Early API and SSH access to the cluster opens in batches, starting with first-run pre-orders in batch #0. Dates come with the next launch.
  4. 4Prices in USD, before tax and shipping. Ships within the United States only.
Pre-order terms

Enterprise

Gen 1 for clusters and enterprise

Enterprise and cluster inquiries are for Gen 1 silicon, not the devkit. Tell us your power budget and what you want to run, and we will get back to you.

Monolith - soon

The complete frontier AI computer

XPU Grasshopper (KU040) serves one model on one XPU. Monolith connects many XPUs into one machine that behaves as one chip, and runs the whole loop on it, with agents, experience generation, and training on the same silicon. Soon.

Explore Monolith

Join the Zettascale team

Zettascale Computing is building the highest-performance, most energy-efficient AI chips for artificial superintelligence to accelerate discovery and further humanity's understanding of the universe. Join us.

View open roles