- Euclyd CRAFTWERK SiP claims bandwidth ranges far past Nvidia’s designs
- The rack-scale system is claimed to succeed in over one exaflop efficiency
- Energy effectivity claims recommend huge good points, although unbiased testing is lacking
European startup Euclyd has introduced new {hardware} geared toward large-scale AI inference.
The system, referred to as CRAFTWERK, was launched on the KISACO Infrastructure Summit 2025 in Santa Clara.
The firm describes it as designed particularly for agentic AI workloads, with specs that set it aside from present accelerators.
Contained in the CRAFTWERK structure
On the core of the discharge is the CRAFTWERK SiP, a system-in-package that matches into the palm of a hand.
It integrates 16,384 customized SIMD processors alongside 1TB of customized ultra-bandwidth reminiscence, or UBM.
Euclyd claims this reminiscence can ship 8,000 terabytes per second of bandwidth.
The compute efficiency is listed at as much as 8 petaflops in FP16 and 32 petaflops in FP4 precision.
These figures place the module above what established corporations reminiscent of Nvidia at present promote.
“Our Crafted Compute philosophy reimagines inference from the bottom up, customized processors, customized reminiscence, and superior packaging,” mentioned Bernardo Kastrup, CEO of Euclyd.
“We’ve engineered each gate for max effectivity and minimal energy draw, by far the bottom within the trade.”
The corporate additionally revealed the CRAFTWERK STATION CWS 32, a rack-scale platform constructed from 32 SiPs.
On this configuration, Euclyd states the system reaches 1.024 exaflops of FP4 compute, backed by 32TB of UBM.
It’s mentioned to generate 7.68 million tokens per second in multi-user mode, and its energy consumption is reported at 125 kilowatts.
In response to the corporate, this can be a hundred-fold achieve in each vitality use and value effectivity in comparison with current options.
The benchmark used to determine these enhancements was modeled efficiency with Llama 4 Maverick.
Headquartered in Eindhoven, the Netherlands, the corporate additionally maintains workplaces in San Jose, California.
It promotes environmentally accountable engineering and effectivity good points in information heart infrastructure.
“I consider AI inference will dominate datacenter silicon. CRAFTWERK’s economics will speed up agentic AI adoption and usher in an period of plentiful inference,” mentioned investor Peter Wennink, former CEO of ASML.
Whereas the specs of CRAFTWERK are bold, the claims stay untested exterior the corporate’s personal framework.
Startups within the semiconductor house usually face challenges in manufacturing at scale, constructing dependable software program assist, and guaranteeing integration with present information heart infrastructure.
Euclyd’s announcement suggests a design that would, on paper, outperform main accelerators, however whether or not it will probably ship in apply will rely upon outcomes seen in actual deployments.
Till these outcomes emerge, the {hardware} stays a set of spectacular numbers with an unsure path to widespread adoption.
