cd /news/large-language-models/show-hn-hw-sw-co-design-blueprint-foโ€ฆ ยท home โ€บ topics โ€บ large-language-models โ€บ article
[ARTICLE ยท art-62796] src=github.com โ†— pub= topic=large-language-models verified=true sentiment=ยท neutral

Show HN: HW/SW co-design blueprint for fault recovery in LLM training

A developer released a hardware-software co-design prototype for fault recovery in LLM training, featuring zero-copy memory bypass and dynamic hot-plugging for HBM bank failures. The project targets NVIDIA Ampere and Hopper architectures to mitigate compilation stalls and hardware jitter in distributed clusters.

read5 min views1 publishedJul 16, 2026
Show HN: HW/SW co-design blueprint for fault recovery in LLM training
Image: source

Experimental PIM-HBM Hardware Co-Design Subsystem exploring 0ns Framework Memory View Fusion & Fault Telemetry

This project is a hardware-software co-design prototype engineered to investigate and resolve software abstraction fragmentation barriers within next-generation accelerator infrastructure environments.

By interlocking the low-level physical cache-line alignment mechanisms directly with the upper high-performance framework (JAX/XLA) address interfaces into a unified algebraic pipeline, this subsystem explores the structural viability of stably mitigating static compilation stalls and hardware timing jitters during distributed cluster operations.

0ns Memory Copy Bypass: Leverages the__cuda_array_interface__

specification to directly link low-level C++ physical address lines to the JAX tensor bus, structurally eliminating host-device hardware memory copy overhead.Pure Branchless Loop: Completely eradicates conditional branches (if

), instead deploying precise ternary operations and register-resident data reuse patterns to guarantee the compiler forces conditional move primitives (SEL/PRMT

).Warp-level Dynamic Bounds: Intercepts potential out-of-bound memory faults (SegFault

) inside ragged tail grids by driving a warp-level binary tree reduction firewall via__shfl_down_sync

to compute active maximum surviving address offsets.Algebraic Insulation Gate: Captures hardware fault signals and floating-point divergence anomalies behind the JAX runtime using dedicatedstop_gradient

circuits, isolating backpropagation chain contamination at the physical layer.Dynamic Hot-Plugging Recovery: Upon physical HBM bank failure during active cluster runs, this mechanism attempts real-time 64-bit address wire hot-swapping for the corrupted device slot without causing collective communication (NCCL) stalls or graph recompilation.

: Declares legal safeguards and patent retaliation defense clauses under the Apache License 2.0 specification.LICENSE

: The build orchestrator that automatically tracks system hardware architecture topologies andCMakeLists.txt

pybind11

compilation paths to emit the final shared object (.so

) libraries.: The core branchless mathematical acceleration kernel implementingpim_hbm_core.cu

alignas(32)

cache-line matching,__activemask()

dynamic address firewalls, and__ldg

high-speed read rails (C++/CUDA).: The pre-warming and backpropagation chain insulation layer utilizingpim_hardware_gate.py

ShapeDtypeStruct

virtual abstract tracers to lock XLA compiler machine code while maintaining 0MB of physical VRAM footprint (Python/JAX).: The macro-level topology control tower that intercepts per-node VRAM physical address lines to establish zero-copytopology_sharding.py

NamedSharding

global distributed matrix views (Python/JAX).: The real-time, high-availability hot-plugging swap engine that handles background fault scans within the distributed weight matrices and switches routing to the pre-reserved emergency backup address pool (Python/JAX).hardware_fault_recovery.py

: The collective scanning and recovery engine engineered for large-scale infrastructure, leveraging wire-level NCCL All-Reduce fusion andhardware_fault_recovery_distributed.py

np.flatnonzero

vectorized fault extraction (Python/JAX).: A transformer layer adapter plugin that conducts direct 0ns address ingestion matching Llama-3-8B dimensions (4096 / 14336) alongside branchless fault-tolerant forward execution buses (Python/JAX).llama3_layer_adapter.py

Execute the following commands sequentially within a high-performance cluster terminal configured with NVIDIA Ampere (A100) or Hopper (H100/H200) environments to ignite the memory view fusion mode and the hardware fault-tolerant emulation engine.

mkdir build && cd build

cmake ..

make -j\$(nproc)

cp pim_hbm_bridge_core*.so .. && cd ..

python3 pim_hardware_gate.py

python3 topology_sharding.py

python3 hardware_fault_recovery_distributed.py

python3 llama3_layer_adapter.py

This project is distributed under the terms of the Apache License 2.0. You are free to modify and distribute the software, provided that the original copyright notice and license disclosure obligations are fully preserved.

Experimental PIM-HBM Hardware Co-Design Subsystem exploring 0ns Framework Memory View Fusion & Fault Telemetry

๋ณธ ํ”„๋กœ์ ํŠธ๋Š” ์ฐจ์„ธ๋Œ€ ๊ฐ€์†๊ธฐ ์ธํ”„๋ผ ํ™˜๊ฒฝ์—์„œ ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ๋Š” ์†Œํ”„ํŠธ์›จ์–ด ๊ณ„์ธต ๊ฐ„ ํŒŒํŽธํ™” ์žฅ๋ฒฝ์„ ์—ฐ๊ตฌํ•˜๊ธฐ ์œ„ํ•ด ์„ค๊ณ„๋œ ํ•˜๋“œ์›จ์–ด-์†Œํ”„ํŠธ์›จ์–ด ๊ณต๋™ ์„ค๊ณ„(Co-design) ํ”„๋กœํ† ํƒ€์ž…์ž…๋‹ˆ๋‹ค.

์ €์ˆ˜์ค€์˜ ๋ฌผ๋ฆฌ ์บ์‹œ๋ผ์ธ ์ •๋ ฌ ๋งค์ปค๋‹ˆ์ฆ˜๊ณผ ์ƒ์œ„ ๊ณ ์„ฑ๋Šฅ ํ”„๋ ˆ์ž„์›Œํฌ(JAX/XLA) ๊ฐ„์˜ ์ฃผ์†Œ์„  ์ธํ„ฐํŽ˜์ด์Šค๋ฅผ ๋‹จ์ผ ๋Œ€์ˆ˜ ํŒŒ์ดํ”„๋ผ์ธ์œผ๋กœ ์—ฐ๊ฒฐํ•˜์—ฌ, ๋ถ„์‚ฐ ํด๋Ÿฌ์Šคํ„ฐ ๊ฐ€๋™ ์ค‘ ์œ ๋ฐœ๋˜๋Š” ์ •์  ์ปดํŒŒ์ผ ๋ ‰ ๋ฐ ํ•˜๋“œ์›จ์–ด ์ง€ํ„ฐ๋ฅผ ์•ˆ์ •์ ์œผ๋กœ ์ œ์–ดํ•  ์ˆ˜ ์žˆ๋Š” ๊ฐ€๋Šฅ์„ฑ์„ ํƒ๊ตฌํ•ฉ๋‹ˆ๋‹ค.

0ns Memory Copy Bypass:__cuda_array_interface__

๊ทœ๊ฒฉ์„ ํ™œ์šฉํ•ด C++ ๊ธฐ๊ณ„์–ด ์ฃผ์†Œ์„ ์„ JAX ํ…์„œ ๋ฒ„์Šค์— ์ง๊ฒฐํ•จ์œผ๋กœ์จ, ํ˜ธ์ŠคํŠธ-๋””๋ฐ”์ด์Šค ๊ฐ„ ๋ฌผ๋ฆฌ ๋ณต์‚ฌ ์˜ค๋ฒ„ํ—ค๋“œ๋ฅผ ๊ตฌ์กฐ์ ์œผ๋กœ ํ•ด์†Œํ•ฉ๋‹ˆ๋‹ค.Pure Branchless Loop: ์กฐ๊ฑด ๋ถ„๊ธฐ๋ฌธ(if

)์„ ์™„์ „ํžˆ ๋ฐฐ์ œํ•˜๊ณ  ์‚ผํ•ญ ์—ฐ์‚ฐ ๋ฐ ๋ ˆ์ง€์Šคํ„ฐ ์ƒ์ฃผ ๋ฐ์ดํ„ฐ ์žฌ์‚ฌ์šฉ ๊ตฌ์กฐ๋ฅผ ์ •๋ฐ€ ๊ตฌ์„ฑํ•˜์—ฌ, ์ปดํŒŒ์ผ๋Ÿฌ ์ˆ˜์ค€์˜ ์กฐ๊ฑด๋ถ€ ์ด๋™ ๋ช…๋ น์–ด(SEL/PRMT

) ์ถœ๋ ฅ์„ ์œ ๋„ํ•ฉ๋‹ˆ๋‹ค.Warp-level Dynamic Bounds: ๋งˆ์ง€๋ง‰ ๊ทธ๋ฆฌ๋“œ ์žํˆฌ๋ฆฌ ์˜์—ญ์—์„œ ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ๋Š” ๋ฉ”๋ชจ๋ฆฌ ์ฐธ์กฐ ์˜ค๋ฅ˜(SegFault)๋ฅผ ๋ฐฉ์ง€ํ•˜๊ธฐ ์œ„ํ•ด,__shfl_down_sync

๊ธฐ๋ฐ˜ ์›Œํ”„ ๋‚ด 2์ง„ ํŠธ๋ฆฌ ์ตœ๋Œ€ ์ƒ์กด ์ฃผ์†Œ ๋™์  ๋ฆฌ๋•์…˜ ๋ฐฉํ™”๋ฒฝ์„ ๊ตฌ๋™ํ•ฉ๋‹ˆ๋‹ค.Algebraic Insulation Gate: JAX ๋Ÿฐํƒ€์ž„์—์„œ ํ•˜๋“œ์›จ์–ด ๊ฒฐํ•จ ์‹ ํ˜ธ ๋ฐ ์ˆ˜์น˜ ๋ฐœ์‚ฐ ์˜ค์ฐจ๋ฅผstop_gradient

ํšŒ๋กœ๋กœ ํฌํšํ•˜์—ฌ, ๋ฏธ๋ถ„ ์‚ฌ์Šฌ ์˜ค์—ผ์„ ํ”ผ์ง€์ปฌ ๋ ˆ๋ฒจ์—์„œ ๊ฒฉ๋ฆฌ ๋ฐ ์ ˆ์—ฐํ•ฉ๋‹ˆ๋‹ค.Dynamic Hot-Plugging Recovery: ๊ฐ€์†๊ธฐ ํด๋Ÿฌ์Šคํ„ฐ ๊ตฌ๋™ ์ค‘ ํŠน์ • HBM ๋ฑ…ํฌ์— ๋ฌผ๋ฆฌ์  ๊ฒฐํ•จ ๋ฐœ์ƒ ์‹œ, ๊ธ€๋กœ๋ฒŒ ํ†ต์‹ (NCCL) ์ค‘๋‹จ ๋ฐ ๊ทธ๋ž˜ํ”„ ์žฌ์ปดํŒŒ์ผ ์—†์ด ์˜ค์ง ๋ถˆ๋Ÿ‰ ์žฅ์น˜์˜ 64๋น„ํŠธ ์ฃผ์†Œ์„ ๋งŒ ์‹ค์‹œ๊ฐ„์œผ๋กœ ์šฐํšŒ ์Šค์™€ํ•‘(Hot-Swapping)ํ•˜๋Š” ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ์‹œ๋„ํ•ฉ๋‹ˆ๋‹ค.

: Apache License 2.0 ์˜๊ฑฐ ๋ฒ•์  ๋ฐฉํ™”๋ฒฝ ๋ฐ ํŠนํ—ˆ ๋ณดํ˜ธ ์กฐํ•ญ ๋ช…์‹œLICENSE

: ์‹œ์Šคํ…œ ๊ฐ€์†๊ธฐ ์•„ํ‚คํ…์ฒ˜ ํ™˜๊ฒฝ๊ณผCMakeLists.txt

pybind11

์ปดํŒŒ์ผ ํŒจ์Šค๋ฅผ ์ž๋™ ์ถ”์ ํ•˜์—ฌ ๊ณต์œ  ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ(.so

)๋ฅผ ์‚ฌ์ถœํ•˜๋Š” ๋นŒ๋“œ ์˜ค์ผ€์ŠคํŠธ๋ ˆ์ดํ„ฐ:pim_hbm_core.cu

alignas(32)

์บ์‹œ๋ผ์ธ ์ผ์น˜ ๋ ˆ์ด์•„์›ƒ,__activemask()

๋™์  ์ฃผ์†Œ ๋ฐฉํ™”๋ฒฝ ๋ฐ__ldg

๊ฐ€์† ๋ ˆ์ผ์ด ์ฃผ์ž…๋œ ๋ฌด๋ถ„๊ธฐ ์ˆ˜ํ•™ ๊ฐ€์† ์ปค๋„ ์ฝ”์–ด (C++/CUDA):pim_hardware_gate.py

ShapeDtypeStruct

๊ฐ€์ƒ ์ถ”์ƒํ™” ํŠธ๋ ˆ์ด์„œ๋ฅผ ํ™œ์šฉํ•ด ์‹ค์žฌ VRAM ์ ์œ  0MB ์ƒํƒœ๋กœ XLA ์ปดํŒŒ์ผ๋Ÿฌ ๊ธฐ๊ณ„์–ด๋ฅผ ๊ณ ์ •ํ•˜๋Š” ์˜ˆ์—ด ๋ฐ ๋ฏธ๋ถ„ ์‚ฌ์Šฌ ์ ˆ์—ฐ ๋ ˆ์ด์–ด (Python/JAX): ๋Œ€๊ทœ๋ชจ ํด๋Ÿฌ์Šคํ„ฐ ๋…ธ๋“œ๋ณ„ VRAM ๋ฌผ๋ฆฌ ์ฃผ์†Œ์„ ์„ ๊ฐ€๋กœ์ฑ„์–ด ์ œ๋กœ์นดํ”ผtopology_sharding.py

NamedSharding

๊ธ€๋กœ๋ฒŒ ๋ถ„์‚ฐ ๋งคํŠธ๋ฆญ์Šค ๋ทฐ๋ฅผ ์ˆ˜๋ฆฝํ•˜๋Š” ๊ฑฐ์‹œ์  ํ† ํด๋กœ์ง€ ๊ด€์ œํƒ‘ (Python/JAX): ๋ถ„์‚ฐ ๊ฐ€์ค‘์น˜ ํ–‰๋ ฌ ๋‚ด ๋ถˆ๋Ÿ‰ ๋ฑ…ํฌ ๋ฐฑ๊ทธ๋ผ์šด๋“œ ์Šค์บ” ๋ฐ ๋น„์ƒ ๋ฐฑ์—… ํ’€ ์ฃผ์†Œ์„ ์„ ํ™œ์šฉํ•œ ์‹ค์‹œ๊ฐ„ ๋ฌด์ค‘๋‹จ ํ•ซํ”Œ๋Ÿฌ๊น… ์Šค์™€ํ”„ ์—”์ง„ (Python/JAX)hardware_fault_recovery.py

: ์ดˆ๋Œ€ํ˜• ์ธํ”„๋ผ๋ฅผ ์œ„ํ•œ NCCL All-Reduce ์™€์ด์–ด ๋ ˆ๋ฒจ ์œตํ•ฉ ์ง‘์‚ฐ(Collective) ์Šค์บ” ๋ฐhardware_fault_recovery_distributed.py

np.flatnonzero

๋ฒกํ„ฐํ™” ๊ฒฐํ•จ ์ ์ถœ ๋ณต๊ตฌ ์—”์ง„ (Python/JAX): Llama-3-8B ๊ณ ์œ  ์ฐจ์›(4096 / 14336) ์งํ†ต 0ns ์ฃผ์†Œ ์ธ์ž… ๋ฐ ๋ฌด๋ถ„๊ธฐ ๊ฒฐํ•จ ํ—ˆ์šฉ ์ˆœ๋ฐฉํ–ฅ ํ›ˆ๋ จ/์ถ”๋ก  ๋ฒ„์Šค ์–ด๋Œ‘ํ„ฐ ํ”Œ๋Ÿฌ๊ทธ์ธ (Python/JAX)llama3_layer_adapter.py

NVIDIA Ampere(A100) ๋˜๋Š” Hopper(H100/H200) ํ™˜๊ฒฝ์ด ๊ตฌ์ถ•๋œ ๊ณ ์„ฑ๋Šฅ ํด๋Ÿฌ์Šคํ„ฐ ํ„ฐ๋ฏธ๋„์—์„œ ๋‹ค์Œ ๋ช…๋ น์„ ์ˆœ์ฐจ ๊ฐ€๋™ํ•˜์—ฌ ๋ฐ”์ดํŒจ์Šค ๋ชจ๋“œ ๋ฐ ํ•˜๋“œ์›จ์–ด ๊ฒฐํ•จ ํ—ˆ์šฉ(Fault-Tolerant) ์—”์ง„ ์—๋ฎฌ๋ ˆ์ด์…˜์„ ๊ธฐํญํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

mkdir build && cd build

cmake ..

make -j\$(nproc)

cp pim_hbm_bridge_core*.so .. && cd ..

python3 pim_hardware_gate.py

python3 topology_sharding.py

python3 hardware_fault_recovery_distributed.py

python3 llama3_layer_adapter.py

๋ณธ ํ”„๋กœ์ ํŠธ๋Š” Apache License 2.0 ์˜๊ฑฐํ•˜์—ฌ ๋ฐฐํฌ๋ฉ๋‹ˆ๋‹ค. ์ž์œ ๋กœ์šด ์ˆ˜์ • ๋ฐ ๋ฐฐํฌ๊ฐ€ ๊ฐ€๋Šฅํ•˜๋‚˜ ์ €์ž‘๊ถŒ ๋ฐ ๋ผ์ด์„ ์Šค ๊ณ ์ง€ ์˜๋ฌด๊ฐ€ ์ˆ˜๋ฐ˜๋ฉ๋‹ˆ๋‹ค.

โ”€โ”€ more in #large-language-models 4 stories ยท sorted by recency
โ”€โ”€ more on @nvidia 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain โ€” perfect for shipping the agent you just read about.

$git push zahid main
โ†’ Live at https://your-agent.zahid.host โœ“
Get free account โ†’ Pricing
from โ‚ฌ0/mo ยท no card required
LIVE [news/show-hn-hw-sw-co-desโ€ฆ] indexed:0 read:5min 2026-07-16 ยท โ€”