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]

@Qbb7680 那我呢。。。。

[Robin Lu @RobinLu2001]

@_WEARIDE_ FF外か��失礼します、自分の経験から、　7900 XTX のおすすめの実行環境は ROCm 6.4.3 + PyTorch (Multi-GPU 2.6?, Single-GPU >2.7?) , FAv2 は https://t.co/k67KJ2edBQ と思います。この場合は RTX 3090 級の性能が出ています。

[Robin Lu @RobinLu2001]

@Qbb7680 这件外套好熟���x

[Robin Lu @RobinLu2001]

@OriaChiuan 我感觉我身上保留传统士大夫腐朽的地方还是太多了（x

[Robin Lu @RobinLu2001]

两朝贰臣，悖前主，负后主，真个里外不��人

[Robin Lu @RobinLu2001]

也许已经铺完了路，但我没力气走下去了

[Robin Lu @RobinLu2001]

我会在遥远地方等你 直到你已经不再悲伤 I want your freedom like a bird

[Robin Lu @RobinLu2001]

@arielheleneto 该想的，该说的，都说了，油尽灯枯了

[Robin Lu @RobinLu2001]

@mizu_bai 是我

[Robin Lu @RobinLu2001]

而解开这个绳结，不就是我们的工作吗？

[Robin Lu @RobinLu2001]

https://t.co/Hl0WJqX3vO

(Custom / PyTorch-native) Operator misbehaving, Reduce a+b+c != c+b+a problem in non-NVIDIA training.

[Robin Lu @RobinLu2001]

@RajaXg And High-Flyer has at least 10000 A100 + 2048 H800 working simultaneously, that might be already among the largest NVIDIA clusters in China...

[Robin Lu @RobinLu2001]

@RajaXg But the scaling law(?) and memory pattern in DeepSeek doesn't really change, (imagine we provide high-flyer with GH200/MI300, it would even accelerate or enhance their work with a larger model) Maybe still need LPDDR+SRAM+Locality to change the game?

[Robin Lu @RobinLu2001]

"Less Artificial Involvement, (with First Principles Thinking), More Intelligence" - Robin Lu, (2023)

[Robin Lu @RobinLu2001]

@Redstone_Bi @regymm0 yep still you can use PCIe P2P, or some board with NVLink Bridge

[Robin Lu @RobinLu2001]

@regymm0 @Redstone_Bi On the roadmap.

[Robin Lu @RobinLu2001]

Some publicly available benchmarks on Chinese domestic solutions. (information source can be find on the photo) Huawei Ascend 910C has been released (and performance has improved), #DeepSeek has official support for Ascend.

[Robin Lu @RobinLu2001]

And actually that's where US sanction already restrict China, (imagine High-Flyer with same amount of GH200, they can afford that), #DeepSeek could be way further than current performance.

Will talk about China domestic design in AI training accelerator later.

[Robin Lu @RobinLu2001]

When I was in Tongji, I tried to apply for a intern in High-Flyer. They was working on "AGI" as second priority but still quant as first priority. At that time they was recruiting NCCL and other GPU cluster internship and seems to freeze quant HC. (1/1) #DeepSeek

[Robin Lu @RobinLu2001]

Thus I simply have no idea why #DeepSeek is astonishing to others, like if your employees are brilliant as Jump or Optiver, and you have more resources (10,000 A100 instead of China domestic design) than other company in China (https://t.co/5NGYRPDfFT), likely you'll succeed too.