Performance Engineer, GPU Kernels at Sarvam AI (Bengaluru)

Job Description

Sarvam is hiring a senior Performance Engineer to own its GPU kernel layer, writing custom CUDA, DSL-based and PTX kernels where stock libraries such as cuBLAS, FlashAttention or out-of-the-box Triton leave performance on the table. The posting lists Bengaluru or Chennai, hybrid or on-site.

At a glance

• Role: Performance Engineer, Kernels (senior)
• Where: Bengaluru or Chennai, India (hybrid or on-site, per the posting)
• Experience: 5+ years in ML systems, with 2+ years writing production CUDA kernels
• Hardware: H100, H200 and B200 GPUs
• Apply: open until filled (re-checked 12 October 2026)

About the kernels role

• Author custom CUDA, CUTLASS/CuTe DSL and PTX kernels that beat stock libraries on real production workloads
• Own the explanation of every performance change your kernels make to production latency
• Work on attention kernels and other hot paths across the serving stack
• Collaborate with the inference team, which integrates kernels into the serving system

The two performance postings form a vertical stack: the kernels team writes microsecond-level GPU code, and the inference team integrates it into a running serving stack and owns system-level numbers. Sarvam serves several model families, including LLMs, mixture-of-experts models, Indic speech and multimodal models, on a multi-node fleet of NVIDIA Hopper and Blackwell GPUs.

What Sarvam is looking for

• 5+ years in ML systems, including 2+ years authoring production CUDA kernels, with a kernel in production that beat its baseline
• CUDA at kernel-authoring level: block sizing, shared-memory layout, warp primitives, async copies (cp.async, TMA) and MMA selection
• CUTLASS/CuTe DSL at modify-and-extend level, and PTX at debug-and-modify level
• Nsight Compute and Systems fluency, including reading roofline plots
• Hands-on attention kernel work (FlashAttention-family, paged, MLA, sliding-window or sparse) and awareness of Hopper-to-Blackwell changes

Nice to have

• Communication kernels: NCCL/NVSHMEM, custom collectives, expert-parallel dispatch or KV transfer
• Open-source kernel contributions, which the posting calls the highest-yield signal
• tcgen05, TMA, CTA clusters, distributed shared memory, async pipelining and FP4/microscaling paths
• Host-path optimisation on GH200/GB200

Pay and location

Sarvam’s posting does not state a salary. It lists the location as Bengaluru or Chennai, hybrid or on-site.

How to apply

Apply through the official Sarvam job posting. Sarvam’s posting does not give a closing date, so the role is open until filled; ResearchJobs.in will re-check this listing on 12 October 2026. Check the posting for location and work-authorisation details before you apply.

See all our artificial intelligence jobs, or browse more research jobs on ResearchJobs.in.