Robin Lu

(They/Them) Working on HPC and Computational Neuroscience. Bachelor@Tongji, MSc @UTokyo_GSFS (2023.10-2025.9), @Cadence 中文/English/日本語 云门挂锁难留客，一片青山了此身。

[Robin Lu @RobinLu2001]

斜月如旧，西风又拂，
回首尘缘，几许愁梦，
恰似一江春水东流。

[Robin Lu @RobinLu2001]

Also https://t.co/8dCBaqz9YN for a kind working triton

[Robin Lu @RobinLu2001]

https://t.co/6mmlNNNsuJ some ongoing interesting vibe coding experiments to try to bring-up LLVM on MN-Core (waiting for my server to be back)

[Robin Lu @RobinLu2001]

@256_daisuki FF外から失礼します、https://t.co/RNFHpnAGan と GitHub SX-Aurora + sx-aurora-dev を参考にしてみるといいかもしれません。同じ趣味を持つ人たちのDiscordグループもあります。

[Robin Lu @RobinLu2001]

Implement virglrender and Wayland compositor on Harmony for Linux container

[Robin Lu @RobinLu2001]

https://t.co/q73mGT5swo A Triton 3.2 LTS on Volta/Turing and AMD consumer cards, correctness on performance.
Current Effort: V100 tl.trans (fixed), Turing tl.dots INT8 (fixing), GCN5/RDNA1/CDNA and Vega20 support (fixing)
Will do: Kunlun K200, full Volta/Turing tuning…

[Robin Lu @RobinLu2001]

https://t.co/RqIpN6deEn Flash Attention in Triton for Volta and Turing (PyTorch <=2.6.0, Triton<=3.2.0), better than xformers. Optimal forward on 2080Ti, T4 (limited by bandwidth). Backward requires further tuning on register pressure. V100 requires tuning on MMA occupancy.

[Robin Lu @RobinLu2001]

Updated with a reasonable AMS C35 and TSMC D35 implementation (with SPICE calibration, PEX and Field), and a initial implementation for x-fab xh035 (main problem is missing of open SPICE model)

[Robin Lu @RobinLu2001]

Working on a modernized version of MOSCS(_SUBM) / NCSU CDK, based on klayout/magic, with stdcell library auto generation, and "newer" or available tech support (X-Fab XH018, XH035, AMS C35, CNM25). Salute to Conway. https://t.co/Cn1GCCioDt

[Robin Lu @RobinLu2001]

We could fix for a Triton 3.1/3.2+torch 2.5.1/2.6.0 and CUDA 12.4 for a LTS support of Volta and Turing, making them affordable for Ai4s.

[Robin Lu @RobinLu2001]

https://t.co/RqIpN6deEn Flash Attention in Triton for Volta and Turing (PyTorch <=2.6.0, Triton<=3.2.0), better than xformers. Optimal forward on 2080Ti, T4 (limited by bandwidth). Backward requires further tuning on register pressure. V100 requires tuning on MMA occupancy.

[Robin Lu @RobinLu2001]

https://t.co/RqIpN6deEn Flash Attention in Triton for Volta and Turing (PyTorch <=2.6.0, Triton<=3.2.0), better than xformers. Optimal forward on 2080Ti, T4 (limited by bandwidth). Backward requires further tuning on register pressure. V100 requires tuning on MMA occupancy.

[Robin Lu @RobinLu2001]

@KhushPatil786 Just added it by Luna https://t.co/T1xGovpQZU

[Robin Lu @RobinLu2001]

Working on a modernized version of MOSCS(_SUBM) / NCSU CDK, based on klayout/magic, with stdcell library auto generation, and "newer" or available tech support (X-Fab XH018, XH035, AMS C35, CNM25). Salute to Conway. https://t.co/Cn1GCCioDt

[Robin Lu @RobinLu2001]

Update: Official rednote account reply to me that there will be SPICE and other parts finally

[Robin Lu @RobinLu2001]

lol, they’ve released their official one after we claimed a clean-room rewritten version, which is definitely great for analog sign-off. They still didn’t release SPICE though, maybe we still need to do it by ourselves before they want to release them.

[Robin Lu @RobinLu2001]

lol, they’ve released their official one after we claimed a clean-room rewritten version, which is definitely great for analog sign-off. They still didn’t release SPICE though, maybe we still need to do it by ourselves before they want to release them.

[Robin Lu @RobinLu2001]

https://t.co/L8O43Six9g rewrite klayout DRC/LVS/PEX, still tuning SPICE MOS model (based on HVT stdcells) Also ckdur has turned it into LibreLane-capable digital flow.

[Robin Lu @RobinLu2001]

https://t.co/L8O43Six9g rewrite klayout DRC/LVS/PEX, still tuning SPICE MOS model (based on HVT stdcells) Also ckdur has turned it into LibreLane-capable digital flow.

[Robin Lu @RobinLu2001]

Digital Sign-Off Flow done, Analog PDK and new Digital PDK (targeting better PPA) on the way.

[RegYMM レジム @regymm0]

Ready for @Open_SUSI TR-1um open-source PDK #ASIC #tapeout? In 3 weeks I went from first-timer xschem/klayout to a range of designs: ring oscillator, voltage-controlled oscillator, time-to-digital converter, and a librelane RTL-to-GDS flow developed by @RobinLu2001

[RegYMM レジム @regymm0]

Ready for @Open_SUSI TR-1um open-source PDK #ASIC #tapeout? In 3 weeks I went from first-timer xschem/klayout to a range of designs: ring oscillator, voltage-controlled oscillator, time-to-digital converter, and a librelane RTL-to-GDS flow developed by @RobinLu2001

[Robin Lu @RobinLu2001]

@mov_axbx is that MLU270-S4? I’m not sure but likely you need to sign some NDA or something for their SDK and torch 2.x

[Robin Lu @RobinLu2001]

@yangyuc 以及 910B2/3 本身也还是个 Ampere 级的东西？非要黑屁的话 910C 也更像双芯 Ampere （不

[Robin Lu @RobinLu2001]

现在很多工作，相比挖掘存量卡的潜力，更倾向于去chase新卡，是有问题的。尤其Flash模型，完全应该考虑INT8 W8A8（这样Turing和Ampere还能吃到红利，910C也可以），以及在互联受限的平台上更好的推理。另外，稀疏MoE的模型也有助于Spark和Mac推理的发展。资源有限的组，做这种事情实际上更有价值。