Anthropic 上市路演把家底摊开了讲:Claude 同时跑在 AWS、Google、NVIDIA 和 AMD 四种芯片上,接入 AMD 只花 3 天还是自家智能体 Mythos 干的活。想了解大厂怎么压低算力成本、以及“算力乘数”这个概念的,这篇值得点进去看原始转录。
#AWS Trainium
The non-obvious takeaway is that Trainium's GPSIMD cores can be programmed outside the official NKI path, unlocking hardware AWS's own toolchain does not yet expose. Engineers porting sparse or hash-based workloads to Trainium 3, and teams comparing it against GPU-FPGA setups for DeepSeek-style memory modules, can reuse the injection technique directly. AWS infrastructure buyers evaluating Trainium against NVIDIA GPUs gain a concrete data point on heterogeneity advantages.
要在 AWS Trainium 上跑模型的可以看看,Hugging Face 把 torch-neuronx 支持的模型整理成查询方式,不用再自己一个个试。
The notable point is that one config setup now covers three different training backends on Trainium, so teams can switch between them without rewriting launch scripts or breaking checkpoint compatibility. Engineers choosing between NVIDIA GPUs and AWS silicon can use this to evaluate Trainium for fine-tuning open-source models like Qwen3. The guide also shows how NKI MoE kernels fit into the same workflow.
在 Trainium 上跑 PyTorch 但不知道慢在哪的朋友,这篇教你怎么用 Neuron Profiler 找瓶颈,很实用。
Amazon 把 PyTorch 代码直接编译到自家 Trainium 芯片上跑,开发者一行代码不用改,还开源了。
PyTorch 现在能在 Amazon 自研的 Trainium 芯片上直接跑,代码一行不用改,想省算力成本的可以关注。
AWS adopting community extension points means Trainium code can look like ordinary PyTorch rather than a separate SDK, lowering the barrier to running training workloads on AWS's own silicon. Teams benchmarking GPU alternatives for cost-sensitive training jobs can now test Trainium without rewriting code. The announcement lands ahead of the PyTorch Conference, where other hardware vendors are expected to show similar integrations.
PyTorch 直接跑在 AWS 自研 Trainium 上了,GPU 之外多了个训练选择,10 月大会还有现场 demo。
Trainium becoming a standard PyTorch device means teams can apply familiar torch.profiler workflows to custom silicon instead of NVIDIA tooling. Practitioners running training or inference on AWS should note the session's path from profiler capture to hardware cycle debugging when deciding whether to adopt Trainium for cost-sensitive workloads.