Architect Labs · San Francisco
Hardware & ElectricalHardwareFullTime
Architect is a frontier AI lab for custom silicon. We partner with frontier labs, clouds / neoclouds, physical AI companies, and advanced fabs to tape-out custom chips co-designed for next-generation AI workloads. Our goal is to compress end-to-end software to silicon timelines, and maximize intelligence per watt and per dollar for the world. We are a small exceptional team across silicon, systems, software and frontier AI. Our team have led research teams at nearly every frontier AI lab, and at some of the most complex SoCs in the world.
As a Founding Member of the Technical Staff on the RTL Design team at Architect, you’ll own the AI-driven microarchitecture and RTL design of the memory subsystem going into production silicon. You will define, drive, and revise the block-level micro-architecture specification for memory controllers, memory hierarchy management, and memory-side accelerators — ensuring maximum bandwidth utilization, minimal latency, and efficient power delivery for compute-intensive ML workloads.
Own the memory subsystem RTL end-to-end: from DDR/HBM controller design through code generation, lint, CDC, synthesis, and timing closure using our AI-driven design flow.
Design and implement memory controllers: including DDR5/LPDDR5X PHY-side controller logic, HBM3/HBM3E pseudo-channel controllers, command scheduling (open-page/close-page policies, bank-level parallelism), refresh management, and ECC/RAS engines.
Architect the memory hierarchy: including multi-level cache controllers, scratchpad memory managers, coherency protocol engines (where applicable), prefetch engines, and bandwidth partitioning/QoS mechanisms to serve diverse traffic profiles from ML accelerator datapaths.
Design memory-side accelerators: near-memory compute logic, scatter-gather DMA engines, address translation/remapping units, compression/decompression engines co-located with memory interfaces, and intelligent prefetchers tuned for ML access patterns.
Work directly with the principal architect to refine microarchitectural specs, resolve implementation trade-offs (bandwidth vs. latency vs. area vs. power), and feed area/timing/power realities back into the architecture and internal AI systems.
Define and maintain interface specifications: DDR PHY interfaces (DFI), HBM PHY interfaces, on-chip SRAM interfaces, AXI/ACE/CHI for memory-facing fabric ports, and custom interfaces for near-memory accelerator datapaths.
Build and maintain RTL infrastructure for our in-house AI-driven flow: design automation scripts, regression flows, lint/CDC waivers, and integration collateral for the memory subsystem.
Close collaboration with DV: Support verification bring-up with memory timing models, protocol-compliant BFMs, SVA assertions for JEDEC protocol compliance, coverage plans targeting worst-case scheduling scenarios, and architectural documentation for verification closure.
Close collaboration with SW and ML: Support and guide our SW and ML experts to revise and improve our in-house AI flow based on your memory subsystem domain expertise — particularly around workload-driven memory access pattern optimization.
Support FPGA prototyping on Xilinx for early functional validation of memory controllers, including bring-up with DDR MIG IPs and HBM validation platforms.
Degree: Bachelor’s, Master’s, or PhD in Electrical Engineering, Computer Engineering, or a closely related field.
Experience: 5+ years (10+ preferred) in RTL design with at least one advanced-node tapeout experience involving memory subsystems (DDR/LPDDR/HBM controllers, cache hierarchies, or memory-intensive SoC subsystems).
Memory Interface Expertise: Deep familiarity with JEDEC memory standards — DDR5/LPDDR5X command/address protocols, timing parameters, training sequences, and/or HBM2E/HBM3 pseudo-channel architecture, stack addressing, and interleaving schemes.
Memory Controller Design: Hands-on experience designing or owning memory controller blocks including command schedulers, bank state machines, refresh engines (per-bank, fine-granularity), read/write turnaround optimization, and PHY interface timing (DFI or proprietary).
Memory Hierarchy Architecture: Experience with multi-level cache design (tag/data arrays, replacement policies, coherence protocols), scratchpad controllers, or unified memory architectures with partitioning and QoS.
SystemVerilog: Clear, synthesizable, lint-clean RTL with strong design habits — parameterization for multi-standard support (DDR5/HBM3), modularity for channel/pseudo-channel instantiation, and configurability for different capacity/bandwidth targets.
Block-Level Depth: Hands-on experience with SRAM controllers and arbiters, bank conflict resolution, address hashing/interleaving, ECC encode/decode engines, and high-bandwidth data movement between on-chip and off-chip memory.
SoC Methodology: Solid grasp of synthesis, timing constraints, clock domain crossings (PHY-to-controller domain, multi-frequency memory interfaces), reset strategies, AMBA protocols (AXI, ACE, CHI), and power management for memory subsystems.
Python: Strong skills for design automation, performance modeling, regression infrastructure, and tooling.
PPA Ownership: Experience taking a memory controller or cache subsystem from RTL through synthesis and working with PD teams on timing/area/power closure — particularly for high-frequency controller logic and wide data buses.
Experience with HBM integration: interposer-level considerations, PHY calibration, thermal management impacts on refresh.
Familiarity with CXL memory pooling, Type 3 device controllers, or disaggregated memory architectures.
Near-memory or processing-in-memory (PIM) design experience.
Low-power design techniques: DVFS-aware memory scheduling, partial-array self-refresh, clock gating of idle channels, power gating of unused banks.
FPGA prototyping experience (Xilinx Vivado/Vitis) with DDR MIG or HBM subsystem IP integration.
SVA assertions for JEDEC protocol compliance (command sequencing, timing parameter checking, training state machines).
Prior IP building and delivery experience for DDR/LPDDR controllers, HBM controllers, or cache subsystem IPs.
Performance modeling: experience building or using cycle-accurate memory system simulators (e.g., DRAMSim, Ramulator) to validate microarchitectural decisions.
Domain-specific research contributions: publications or patents in memory systems, memory scheduling algorithms, or memory-centric compute architectures for ML workloads.
You’ll join a founding team building the future of chip design at the intersection of AI and silicon. Your memory subsystem expertise will directly shape production ASICs — enabling the bandwidth and efficiency that ML workloads demand — and influence how AI transforms hardware development from spec to tapeout.