Svashasan Logo

An Automotive Intelligence SoC Designed Around Svashasan OS

Purpose: How Svashasan Labs proposes to develop its own automotive compute chip, Svashasan S1, by designing it around the measured workloads of Svashasan OS, and how that chip anchors a wider ecosystem of software, silicon, intelligence and developer tools.

Status note. Svashasan S1 is a proposed SoC described in an internal product requirements document (v1.0). Every hardware figure on this page is a preliminary target pending workload profiling, IP selection, architecture validation and OEM requirements. No silicon has been designed or fabricated, and the chip alone does not determine an SAE autonomy level.

Measure

Workloads first

Profile Svashasan OS on existing chips before drawing a single block.

Design

A chip shaped by the OS

NPU, memory bandwidth and safety island sized to measured demand.

Gate

No premature silicon

Seven conditions must hold before any tapeout spend.

Grow

An ecosystem, not a part

OS, HAL, silicon, intelligence and SDKs reinforcing each other.


1. Why Our Own Silicon

Software-defined vehicles are consolidating dozens of separate controllers into a few central computers. The company that defines both the operating system and the compute it runs on can tune them to each other, reduce OEM integration effort and execute perception, decisions and vehicle services locally, without depending on the cloud.

Svashasan platform visionSvashasan Developer EcosystemSDKs · APIs · tools · OEM integration infrastructureSvashasan Intelligencelocalised intelligence and decision supportSvashasan OSthe vehicle operating system — the primary productSvashasan Hardware Abstraction Layerruns on third-party SoCs and on Svashasan siliconSvashasan S1 · S2 · S3proprietary automotive compute siliconOne integrated platformTodayOS + HALProposedsilicon
Figure 1. The long-term platform. Svashasan OS stays the product; the hardware abstraction layer lets it run on existing automotive chips today, and Svashasan silicon is the proposed layer that makes it run best.

2. Designed Around Svashasan OS

S1 starts from the workloads, not the transistor count. The PRD defines five computational domains Svashasan OS must run, and each block on the chip exists to serve one or more of them.

Workload to hardware mappingPerceptiondetection · lanes · signs · segmentation · multi-cameraSensor fusioncamera/radar/LiDAR · tracking · SLAM · GNSS/IMUDecision-makingprediction · path planning · trajectory · ACC/AEB requestsVehicle operationsECU comms · diagnostics · BMS analytics · OTA · telemetryOn-device intelligencevoice · local language models · predictive maintenanceISPcamera pre-processingNPU300–400 INT8 TOPS targetDSPradar + signal filteringGPUparallel compute · HMIApplication CPU8 cores · Svashasan OSReal-time CPU8 cores · deterministicSafety islandlockstep · monitors allEvery block on the chip exists because a Svashasan OS workload needs it — measured first, designed second.
Figure 2. Designed around Svashasan OS. The five workload domains in the S1 requirements map onto specific compute units; the dark blocks carry the heaviest load. The safety island watches everything and runs none of it.

How the fit is achieved

Hardware-software co-design loopRun Svashasan OSon existing platformsProfileCPU · NPU · memory · bandwidthSet requirementsfrom measurementsArchitecture + IPchoose and size blocksVirtual prototype+ FPGA emulationRe-measureOS on the modelLoop until themodel matchesthe workload
Figure 3. How “built for Svashasan OS” actually happens. The chip is sized from measured workloads, then the OS runs on a virtual and FPGA model of the chip, and the loop repeats until silicon would match real demand.

3. Inside S1

Svashasan S1 block diagramSvashasan S1 · proposed block diagram · 5–6 nm target · automotive BGAI/O: 16-lane MIPI CSI-2 · 2×10GbE + 4×2.5GbE · 8× CAN FD · 4× CAN XL · LIN · PCIe Gen4 x8 · USB 3 · displayNPU300–400 INT8 TOPSINT8 · INT4 · FP16 · BF16 · CNN + transformerconcurrent pipelines · hardware priority schedulingApplication CPU8 × 64-bit Arm-compatible2.0–2.5 GHz · virtualisationReal-time CPU8 × Cortex-R52 classsplit-lock / lockstepGPUFP32/FP16HMI · mapsISP4–6 Gpx/sHDR · denoiseDSPradaraudioNetwork-on-chip interconnect · QoS · bandwidth allocation · hardware partitioningOn-chip SRAM48–64 MB low-latency buffersMemory controllersLPDDR5X · 64 GB (128 GB option) · 250–300 GB/sHSMroot of trust · AES · SHA · keysPower management · DVFS · independent power domains · thermal sensors45–65 W nominal · ~75 W maximum · AEC-Q100 Grade 2 minimumSafety islanddual lockstep MCUindependent reset
Figure 4. Svashasan S1, block by block. The NPU gets the most silicon because perception dominates the workload; the safety island is fenced off with its own cores, memory and reset so it survives when the main domains do not. All figures are preliminary targets from the S1 PRD.
BlockProposed specificationWhy Svashasan OS needs it
NPU300–400 INT8 TOPS · INT4/FP16/BF16Perception, BEV and fusion networks dominate compute
Application CPU8 × 64-bit Arm-compatible · 2.0–2.5 GHzRuns the Svashasan OS kernel, planning and orchestration
Real-time CPU8 × Cortex-R52 class · lockstep capableDeterministic ECU communication and control coordination
ISP / DSP4–6 Gpx/s · radar and audio processingClean, synchronised sensor data before the NPU sees it
Safety islanddual lockstep MCU · ECC · independent resetMonitors everything and survives when the main domains fail
Memory64–128 GB LPDDR5X · 250–300 GB/s · 48–64 MB SRAMKeeps the NPU fed; bandwidth is a top-three priority

One frame through the chip

On-chip data path8–12 camerasRAW10/12 · CSI-2ISPHDR · denoise · syncNPUdetection · BEV · lanesFusion+ radar via DSP · LiDARApp CPUplanning · decisionsRT CPUvehicle bus · CAN FDShared LPDDR5X + SRAM: zero-copy hand-offs, DMA engines, ECC on safety-relevant memorySafety islanddeadline monitoring of every stageTarget: sensor fusion ≤ 50 ms (preferred ≤ 30 ms) · planning 20–50 Hz · sensor sync ≤ 100 µsCamera to vehicle command, on one chip
Figure 5. One frame through the chip. Each stage hands data to the next through shared memory rather than copying it, which is where much of the latency budget is won or lost.

4. Feeding the NPU: Bandwidth and Memory

The PRD names three subsystems that will decide whether S1 is useful: NPU performance, memory bandwidth and the functional-safety architecture. The second is the one most often underestimated.

Sensor bandwidth calculation8cameras×8,000,000pixels each×30frames/s×12bits/pixel=23.04Gbps raw…before radar (up to 5), LiDAR (1–2), ultrasonics, GNSS/IMU and protocol overhead — every bit moves through memory at least once.Why it matters:Sensor bandwidth sets the minimum memory bandwidth and interconnect capacity before a single neural network runs.That is why the PRD ranks memory bandwidth alongside NPU compute in its top three priorities.
Figure 6. The arithmetic that sizes the chip. Raw camera payload alone is 23 Gbps for the baseline configuration in the S1 PRD.
Memory hierarchyL1 / L2 cacheper core · fastestShared L3inter-core communicationOn-chip SRAM · 48–64 MBsensor and tensor buffersLPDDR5X · 64–128 GB · 250–300 GB/smain working memoryNVMe / UFS · 512 GB–2 TBmaps, logs, software, modelsfaster, smallerlarger, slowerRequirements: zero/low-copy paths · hardware partitioning · bandwidth QoS · ECC · predictable allocation for real-time tasks
Figure 7. The memory hierarchy. On-chip SRAM is the deliberate middle layer: big enough to hold sensor frames and active model tensors so the NPU is not waiting on external memory.
Why TOPS is not throughputArithmetic intensity (operations per byte moved from memory)Usable performancecompute ceiling (peak TOPS)memory-bandwidth limitmore bandwidthBEV / transformerperceptionmulti-cameradetectiondense CNNlayersHigh TOPS alone will notguarantee throughput.Many workloads sit left ofthe ridge, waiting on memory.
Figure 8. Why bandwidth matters as much as TOPS. A qualitative roofline: workloads that move a lot of data per operation hit the memory limit long before the compute ceiling. Placement of workloads is illustrative; Phase 1 profiling will measure where Svashasan OS actually sits.

5. Safe and Secure by Design

Workload isolationHardware: CPU, NPU, memory partitions · bandwidth QoS · IOMMU-style isolationHardware-assisted hypervisorSvashasan OScoreAutonomyapplicationsVehiclecontrolservicesThird-partyOEM appsInfotainmentExternalconnectivityCloudcommunicationsafety-relevantnon-criticaluntrusted edgeA misbehaving infotainment app or a compromised network stack cannot starve or reach the driving pipeline.
Figure 9. Isolation by design. The S1 requirements call for secure separation of seven domains; non-critical workloads must never be able to starve autonomous-driving pipelines of compute or memory.

When something fails

Safety island fault handlingDetectmissed NPU deadline1Identifyaffected domain2Isolatefaulty domain3Reportfailure event4Degrademinimum-risk response5Shed loadrestart non-essentials6Scenario from the S1 PRD: the main NPU stops responding while the vehicle is movingThe safety island keeps running when the application CPU or NPU does not.Lockstep cores alone do not make a chip ASIL-D: that takes the full ISO 26262 process, a fault model and quantified diagnostic coverage.
Figure 10. What happens when the NPU fails. The independent safety island detects, isolates and reports; the vehicle's safety architecture, not the chip alone, decides the degraded-driving response.

Hardware security

Hardware securityHardware root of trustunique identity · HSM keysSecure bootverified boot chainSigned firmwareanti-rollbackHypervisorpartition integritySvashasan OSattestedApplicationssigned, isolatedChain of trust — each layer verifies the nextAES-256 + SHA-256/384acceleration in the HSMAuthenticated debugno open debug in releaseDevice attestationprove hardware + software stateMACsec / TLS accelerationwhere appropriateSecure OTA + rollbacklinks to the QNu-protected update path
Figure 11. Security starts in silicon. A hardware root of trust anchors a verified chain up to Svashasan OS and its applications, protecting both vehicle data and Svashasan's proprietary software. Built to align with ISO/SAE 21434 and support OEM UN R155/R156 obligations.

Power and thermal

Power and thermal envelope0 W15 W30 W45 W65 W75 W45–65 W nominal≈75 W maxS1 power envelopeDVFS per domainindependent power domainsthermal-aware schedulingcontrolled throttlingstandby · sleep · always-onThermal limits decide sustained performance: a chip that throttles in an Indian summer delivers less than its peak TOPS.
Figure 12. The power budget. Sustained NPU performance under thermal load is one of the PRD's mandatory validation tests, because peak numbers mean little inside a hot vehicle.

6. The Software Stack

S1 must ship with a full software ecosystem: secure bootloader, Linux and QNX compatibility, AUTOSAR integration, ROS 2, DDS and SOME/IP middleware, a hypervisor, ONNX-compatible inference, a dedicated NPU compiler and runtime, drivers and profiling tools.

Software stack and HALLayer 1 · Svashasan OS applicationsnavigation · decisions · vehicle services · diagnostics · energyLayer 2 · Autonomy and intelligence runtimeperception · SLAM · fusion · prediction · localisation · planningLayer 3 · Svashasan Hardware Abstraction LayerCPU · NPU · GPU · ISP · DSP · memory · safety APIsLayer 4 · hardware and firmware — swappable underneath the same HALThird-party automotive SoCtoday: development + early deploymentsPartner silicone.g. RISC-V co-design pathSvashasan S1proposed: tuned for these workloadsSame OS, same applications: the HAL is what makes custom silicon an option rather than a dependency.
Figure 13. The hardware abstraction layer is the strategy. Svashasan OS ships on existing chips now; S1 becomes the best place to run it, never the only one.

7. Development Roadmap

Development roadmap with fabrication gatePhase 1Workloadprofilingbenchmark reportsensor specfinal PRDPhase 2Architecture+ IPblock diagramIP selectionpower budgetPhase 3FPGA + virtualprototypeOS on pre-siliconHALdriversPhase 4SilicondevelopmentRTLverificationtapeouteval boardPhase 5AutomotivequalificationsafetysecurityreliabilityOEM designFabrication gate — proceed only after:actual workload profilingOEM requirements validationIP availability assessmentpreliminary cost analysisarchitecture simulationsoftware compatibility validationfunding + manufacturing feasibilityPhases 1–3 aresoftware-led andcapital-light.Phase 4 is wherethe money is spent;the gate guards it.
Figure 14. The development roadmap. Profiling, architecture and pre-silicon prototyping come first; the PRD's “avoid premature fabrication” rule puts a hard gate in front of tapeout.

Acceptance benchmarks

Acceptance benchmarksminimum targetpreferred targetAI performance300 TOPS400 TOPSMemory bandwidth250 GB/s300 GB/sRAM64 GB128 GBCamera processing8 × 8 MP @ 3012 streamsPerception throughput30 FPS60 FPSSensor-fusion latency≤ 50 ms≤ 30 msLocalisation rate50 Hz100 HzPath-planning rate20 Hz50 HzPreliminary engineering acceptance targets — to be validated against the real software stack and sensor set.
Figure 15. What S1 must achieve. Dark bars are the minimum, the dashed outline the preferred target; for latency, lower is better, so the bar shows how far minimum is from preferred.

Qualification

Qualification and mandatory validationISO 26262functional safetyISO/SAE 21434cybersecurityUN R155 / R156OEM obligationsAEC-Q100Grade 2 min.1Concurrent multi-camera perception2Full-load multi-sensor fusion3Sustained NPU performance4Thermal throttling analysis5Memory bandwidth validation6ECC fault injection7Worst-case execution time8CPU/NPU/memory isolation9Secure boot validation10OTA rollback validation11Power-interruption recovery12Automotive temperature testing13EMC testing14Vibration + reliability15Safety fault detection + recovery
Figure 16. Qualification. Fifteen mandatory validation tests from the PRD, under the standards an automotive chip is judged against. Qualifying the chip does not by itself establish compliance of the vehicle.

8. Competitive Position

Competitive reference pointsHigh AIcomputeDeterministicreal-timeSafetyislandWorkloadisolationOwn OS+ HALNVIDIA DRIVE Thorup to 1,000 INT8 TOPS · 64 GB · 273 GB/sNXP S32N5 / S32N5516 × R52 · lockstep M7 · 48 MB SRAMNXP S32N7central SDV computeSvashasan S1 (target)combine all fiverecorded in the PRDpresent, scale not statednot covered by the PRD — not the same as absentS1 values are targets; the reference platforms are shipping products.
Figure 17. Where S1 aims to sit. The PRD benchmarks against NVIDIA DRIVE Thor for AI compute and NXP's S32N family for deterministic, safety-oriented central compute; S1's thesis is to combine both around its own OS. Dots show only what the PRD records — check current vendor datasheets before any external comparison.

9. Building the Ecosystem

Own versus licenseSvashasan ownsSvashasan licensesOverall SoC architectureProven CPU cores (Arm-compatible)System requirementsMemory controllers + PHYsWorkload optimisationSecurity IP (HSM building blocks)Hardware/software integrationInterface IP (CSI-2, Ethernet, PCIe, CAN)Proprietary accelerators, where justifiedGPU, ISP and DSP IPSvashasan OS integration + original IPEDA tools, design services, foundry
Figure 18. Own the system, license the commodity. The value is in the architecture and how tightly it fits Svashasan OS, not in re-inventing a CPU core or a DDR controller.

Partners

Partner ecosystemSvashasan S1architecture ownerSemiconductor IPCPU · NPU · GPU · ISP · DSPe.g. Arm, Synopsys, CadenceASIC design housesRTL · verification · tapeoutFoundriesautomotive-qualified fabe.g. TSMC, Samsung FoundrySoftware partnersBSP · compilers · driversOEMs + Tier 1srequirements · integration · validation
Figure 19. Nobody builds an automotive SoC alone. Named organisations are potential partners listed in the PRD, not agreements; each relationship must be scoped as licensing, co-development, services or manufacturing.

Product generations

Product generationsS1Central automotive compute300–400 TOPS64 GB8+8 coressafety islandSDV + L2+/L3 computeS2High-performance autonomy600–800+ TOPSmore bandwidthbetter perf/Wadvanced autonomyS3Unified vehicle intelligencechipletsscalable acceleratorsdomain consolidationautonomous mobility platforms
Figure 20. Three generations. S1 earns the right to build S2; chiplets in S3 let one design scale across vehicle classes. Generation targets are aspirations, each gated on the one before.

The ecosystem flywheel

Ecosystem flywheelOS on partner chipsOEM programmes shipReal workload datafrom deploymentsBetter silicon fitS1 sized to demandBest perf per wattfor Svashasan OSMore OEMs + developersSDK and APIsMore workloadsnew modulesSoftware pullssilicon along
Figure 21. The wider ecosystem. The loop starts with the OS shipping on other people's chips; real deployment data is what makes custom silicon worth building and what makes it fit.

10. Key Engineering Decisions

DecisionWhat it means
Build around Svashasan OSThe OS stays the primary product; silicon accelerates it through the HAL
Prioritise three subsystemsNPU performance, memory bandwidth and functional safety decide usefulness
License proven IPOwn the architecture and integration; license CPUs, memory, security and interface IP
Avoid premature fabricationNo tapeout before profiling, OEM validation, IP, cost, simulation, software and funding checks pass
Svashasan S1 is the chip Svashasan OS would design for itself: sized by measured workloads, fenced by an independent safety island, and optional by design, so the software never waits for the silicon.