Frontier Inference Clusters
San Jose
Joined April 2024
- Tweets50
- Following3
- Followers61.1K
- Likes340
If you're excited to tackle the challenges required to run the world's inference like:
- Building a Gigawatt-scale factory
- Orchestrating quadrillion-token clusters
- Making self-improving kernel agents
- Designing new product lines from scratch
You should join us:
etched.com/join
We're grateful for our customers, suppliers, and investors for enabling us to ramp production.
Under 1% of the world has access to frontier models. Scaling intelligence requires a new kind of inference hardware.
Our mission has never been more urgent.
etched.com/progress/from-zer…
Chip design is what connects our inference system.
Our entire silicon flow is custom for inference. We develop critical IP in-house, including our fabrics, interconnects, math blocks, and memories.
Closing blocks at low voltages is a very hard PD problem. Etched heavily invests in owning our own physical design, creating novel solutions to solve challenging problems like IR drop. We build custom PnR and sign-off flows to push PPA to its limit.
Chip designers own a block through this entire process: uArch, RTL, DV, PD, signoff, and (most importantly) bringup and production!
Firmware at Etched builds the software that controls our inference clusters.
That means custom algorithms and schemas to achieve maximum performance: doorbell-free submission paths, high-speed swapping to live spares, and nanosecond-accurate profiling tools.
Firmware co-designs directly with our inference and ASIC teams to keep the lowest levels of the stack optimized for production workloads. Firmware owns the code closest to the metal and every cycle you save pushes the Pareto frontier of inference.
Electrical engineering at Etched is uniquely challenging.
Enabling low-voltage inference and cluster-scale memory requires breakthroughs across our platform.
Our PCBs handle tons of current and high-speed SerDes while being manufacturable at scale. This means asymmetric PCB stackups, vertical power delivery, high-fidelity FEA sims, and well-designed test points.
EEs own our platform work from stack-up to bring-up, working closely with our firmware, thermal, and inference teams to get production workloads running as quickly as possible.
Inference at Etched is much more than writing kernels. We own the full stack, from KV cache management to sharding algorithms to workload-aware SerDes tuning.
We push the Pareto curve daily, co-design models with customers, collaborate with chip designers, and are always improving our model harness for recursive kernel development.
If you're excited by things like KV tree masking, diffusion speculators, latent compression, and large scale-up domains, Etched is the right place for you.
We have an unlimited token budget.
We're hiring across inference, firmware, platform, chip design, and more.
We believe in end-to-end ownership that spans beyond most engineering roles.
We look for people that are deeply curious to understand the entire inference system, often beyond what's comfortable.
We're grateful for the support from our partners, suppliers, and team. We have a lot of work to do as we ramp production.
etched.com/progress
We’ve raised $300M in Series C funding at a $10.3B valuation from Sequoia, Andreessen Horowitz, Jane Street, Argo, and SK Hynix.
Our mission is to run the world's inference. This round accelerates production of our inference clusters.
We've opened an 80,000-sqft, 10-MW facility 15 minutes from our office to expedite production and prototyping.
Vertically integrating the full inference hardware stack is extremely hard.
Hear @UbertiGavin and @robertwachen share three years of battle stories leading into today's launch
