Start with less heat.
A lower-power foundation changes what is possible above the logic.
Intelligent
by nature.
Nurtured
by physics.
Intelligence shouldn’t come with a heat penalty.
Meet a new foundation for AI.
8×
Lower power density
64×
Lower energy per byte
Same
Models. Weights. Possibilities.
75%
Target platform advantages. Performance and projected savings depend on workload and configuration.Lower customer GPU capex
01 — BREAK THE CYCLE
The GPU is
a heat trap.
AI’s ambition is limitless.
Its thermal budget isn’t.
THE GPU CYCLEARCHITECTURE EXPLAINED
HEAT
BEGETS HEAT
BEGETS HEAT
01Compute runs hot
02Memory stays off-die
03Data travels farther
04More energy. More heat.
You can’t cool your way
to a new architecture.
GPU compute runs hot. Its thermal load makes stacking memory above the logic difficult. So high-bandwidth memory sits beside the die—and every trip across the interposer consumes energy that becomes more heat.
More cooling treats the symptom. Asycliq changes the starting point: a processor with lower power density, designed to bring memory directly above the logic.
Less heat. A shorter path.
A different foundation for AI.
POWER DENSITY8× lower
ENERGY PER BYTE64× lower
Target architecture comparisons; results depend on operating conditions and workload.02 — INTRODUCING GROVE™
Meet Grove™.
A foundational accelerator for AI
Built to give intelligence room to grow.
Shorten the journey.
Bring memory closer to compute, reducing the energy spent moving each byte.
Change the foundation.
Keep the models and weights. Explore a new path to efficient inference and on-chip learning.
THE FUNDAMENTAL SHIFT
Cooler at the core.
Bigger in possibility.
03 — ONE GROVE™. EVERY SCALE.
From DC
to PC.
Datacenters first. PCs on the horizon.
One architecture, designed to scale in both directions.
Start where demand is greatest.
Our first target is the datacenter: bringing Grove™ to the infrastructure powering demanding AI workloads.
DATACENTER FIRSTScale out. Think hyperscale.
Grow across systems and larger deployments with a common architecture designed for expanding compute needs.
SCALE OUT TO HYPERSCALEBring intelligence closer.
Our direction extends toward PCs: scaling the same architectural foundation into smaller, more personal computing environments.
SCALE DOWN TOWARD PCTHE SAME MODEL. A DIFFERENT PACE.
Same weights.
Two machines.
Explore the comparison.
Choose a prompt. Watch both responses unfold.
Illustrative replay · no live inference
THE PROMPT
In three sentences, why is the sky blue?
Native GPU
RECORDED BASELINEGrove™
PROJECTED PLAYBACKRecreated from the supplied comparison. GPU animation approximates recorded timing; Grove™ playback uses a projection, not measured silicon performance.
About this comparison
The example responses use gemma-3-4b-it. The reference GPU run used fp32 eager execution on an RTX PRO 6000. Projected playback uses 11,600 tokens/second, a system-level estimate for a 70B dense int4 workload applied to these example responses. These are different configurations, not a like-for-like hardware benchmark. Responses are prerecorded; this interactive replay does not run a model.
Watch the original comparison ↗04 — FROM PRINCIPLE TO PROOF
Real hardware.
Revolutionized.
Grove™ brings a new approach to training, inference, and on-chip learning.
Physics, measured.
FPGA demonstrated algorithm and data path.
It learns on the chip.
The first training, inference, and on-line learning silicon on the market.
A full-stack foundation.
18 inventions across five layers, spanning device physics, algorithms, systems, and workloads.