Sr. Principal Engineer, Scale-Up (Switch Silicon & Traffic Management)
US - Santa Clara, CA
$263k–$284k a yearOn-siteLead & above
Job details
- US - Santa Clara, CA
- On-site
- $263k–$284k a year (posted by the company)
- Full-time · Lead & above
- AI Network Software
- Posted 1 month ago
Sponsorship history
Past filings by Upscale Ai in public government data. Not a promise of sponsorship.
- Filed 15 H-1B LCAs in FY2025
Full job description
Why join Upscale, Inc.
Upscale, Inc., is building the high-performance infrastructure powering the next generation of artificial intelligence. Backed by over $500M in funding and rapid global adoption, we are scaling systems designed for the world’s most demanding AI workloads.
We focus on first-principles engineering across silicon, systems, and networking—where performance, scale, and execution are critical. Our team is talent-dense and high-performing. We value ownership, technical rigor, and speed, and we offer the opportunity to work on foundational problems with immediate, real-world impact.
If you’re looking to do high-impact work, move fast, and help define the infrastructure behind the future of AI—Upscale is where you can produce meaningful work at the frontier—and operate at a high standard.
Position Overview
Senior individual-contributor validation role. You own how data-path features are qualified on our scale-up AI networking platform, from the SAI/SDK programming path through ASIC forwarding and traffic-management behavior. You set the test strategy, own feature readiness and release sign-off, debug the hardest cross-stack failures, and grow the QA engineers around you. Primary depth is traffic management. Feature scope (cycle-accurate model and silicon)
Key Responsibilities
- Classification and marking: DSCP and 802.1p to traffic class, trust, TC-to-queue, remarking, port versus switch map binding, classification miss, ACL rewrite
- Queueing: unicast and multicast queues, counters, occupancy, watermarks, drops
- Scheduling: strict priority, WRR/DWRR, hierarchy, shaping, rate limiting, fairness under oversubscription
- Buffering: ingress and egress pools, static and dynamic thresholds, priority groups, headroom, reserved versus shared
- PFC and lossless Ethernet: pause, headroom, lossless queues, watchdog, storm and deadlock, DCB
- ECN and WRED: thresholds, drop probability, ECT/CE, interaction with DCQCN under RoCEv2
- Policing: color-blind and color-aware, CIR/PIR/CBS/PBS, trTCM
- Forwarding and counters: L2/L3, ECMP, mirror and sample, telemetry that makes the above observable
- Own the validation strategy, test plans, coverage, and release entry/exit criteria for the features above
- Qualify lossless and congestion behavior under RoCEv2 and GPU collectives (NCCL/RCCL) at line rate, including incast, PFC headroom, ECN accuracy, and no-drop
- Build the frameworks and debug tools the team depends on, including model-based checks of classification, queueing, scheduling, and drop or mark outcomes
