Power is now the bottleneck in AI data centers – not compute.
The industry has spent years optimising silicon, scaling GPU clusters all the while pushing model architectures to their limits. But there’s a constraint that doesn’t get the same attention: the power delivery systems feeding all that compute.
At weeteq, we’ve been working on a breakthrough that fundamentally changes that equation.
The Hidden Tax in Today’s Power Architectures
Modern GPUs and AI accelerators demand sub-microsecond load transient response, aggressive dynamic voltage and frequency scaling (DVFS), and tighter regulation margins to control runaway power consumption.
Conventional power architectures can’t respond fast enough to these dynamic loads. The typical path – 48v to 12v to 0.8v – introduces conversion stages that simply cannot keep pace with what modern AI silicon demands.
The workaround? Engineers add voltage margin. They run processors at higher voltages than theoretically necessary, just to ensure stability during transient events. This regulation margin is a hidden tax baked into every AI data center today – increasing steady-state power consumption and limiting system-level efficiency.
Introducing DIVA
DIVA – Dynamic Current (I) and Voltage control for ASICs – is a new power delivery architecture designed for high-performance AI and GPU/xPU workloads.
Rather than forcing a trade-off between efficiency and dynamic response, DIVA decouples them entirely through a parallel power delivery architecture combining two complementary paths:
High-Efficiency DC-DC Converter delivers average load power with optimal efficiency, supporting single-step conversion directly to the processor power domain.
Ultra-Fast DIVA Regulation Path supplies instantaneous load transients with response greater than 10 A/ns and enables dynamic voltage scaling faster than 10 mV/ns.
These two paths operate seamlessly in parallel. Average power is delivered efficiently, while fast transient events are handled locally and instantly – without forcing voltage over-margining.
The Numbers
DIVA enables operation at lower average power domain voltage while maintaining identical compute performance:
- Voltage reduction: 0.8V → 0.7V at the power domain
- Power domain consumption: ~30% reduction
- Compute tray power: 15 kW → 10 kW
- Rack power: 120 kW → 80 kW
- Compute output: Equivalent token throughput
The result: 40 kW savings per rack, reduced cooling burden, higher deployment density, and materially lower total cost of ownership.
Why This Matters Now
AI infrastructure is scaling rapidly. Data centres are pushing toward megawatt-scale racks, and the industry is transitioning from 48V/54V to 800V DC architectures to manage the sheer current demands. But voltage conversion is only part of the problem.
The real inefficiency sits at the final stage – where power meets silicon. That’s where voltage margins compound into kilowatts of waste, multiplied across thousands of GPUs.
DIVA addresses this directly. It’s scalable from 48V to 800V DC environments, and it delivers benefits at every level – from silicon to rack.
A Step Change, Not an Incremental Tweak
Conventional architectures trade efficiency for stability. Ultra Edge® DIVA delivers both, at scale.
This isn’t about squeezing another percentage point of efficiency from existing approaches. It’s a step change in how power is delivered to AI compute.
