AMD HPC and Sovereign AI VP Scott Tease argues that AI will not replace traditional high-performance computing, since frontier research in climate modeling, nuclear energy and materials still depends on FP64 double-precision numerics. He advocates a heterogeneous future spanning CPUs, GPUs, AI accelerators and quantum systems, with AMD and IBM piloting integration of classical compute stacks and IBM's 129-qubit Quantum Nighthawk at Oak Ridge National Laboratory.
Heterogeneous Computing
异构计算
Using different processor architectures (e.g., CPU + GPU) together to improve task processing efficiency.
Curated Intelligence
AMD, FuriosaAI, and MangoBoost have signed an MoU to build the first open heterogeneous AI reference cluster in Seoul by next spring. The initiative integrates AMD CPUs/GPUs, FuriosaAI NPUs, and MangoBoost DPUs to optimize hardware pairing for AI workloads and improve overall compute efficiency.
Open-source heterogeneous computing framework TornadoVM released version 7.3, refactoring its backend by removing JVMCI dependencies and adopting the FFM API to deepen integration with the Nvidia CUDA ecosystem. The update supports direct calls to cuBLAS/CUTLASS, multi-GPU execution, and hybrid APIs for Apple Metal, offering Java developers a more efficient acceleration solution for machine learning workloads.
Qujing Tech, in collaboration with Zhipu AI and Moore Threads, launched a 'Domestic Heterogeneous Inference High-Quality AI Token Factory' solution at the GAI+C conference to reduce computing costs and enhance inference efficiency on standard hardware via system-wide heterogeneous coordination. Concurrently, Qujing announced a Series B funding round exceeding 1.3 billion RMB, bringing its total annual fundraising to 2.4 billion RMB, highlighting strong investor interest in domestic AI infrastructure.
Anthropic's S-1 filing reveals a unified compute resource pool integrating chips from Amazon, Google, Nvidia, and AMD. This strategy uses software scheduling to distribute tasks across platforms, aiming to reduce vendor dependency and strengthen bargaining power. While unlikely to disrupt Nvidia's training dominance in the short term, it may erode pricing power and accelerate compute commodification long-term.
Arm, alongside partners like Meta and OpenAI, proposes a new CPU performance paradigm shifting from token throughput to end-to-end task completion. In AI agent scenarios, tool invocation latency and orchestration efficiency become key metrics, driving data center compute allocation from GPU-centric models toward heterogeneous collaborative architectures.
Open-source heterogeneous computing framework TornadoVM has released version 7.3, enabling Java applications to directly leverage GPUs for high-performance computing. The tool works alongside existing JVMs without replacement, automatically offloading code to hardware backends and managing memory to provide underlying acceleration for cloud-native and machine learning scenarios.
Synaptics has launched the Astra SR200 series MCUs, integrating Google's Coral NPU and high-bandwidth custom DRAM to enable multimodal edge AI on microcontroller-class hardware. With up to 400 GOPS of compute power, the chip targets next-gen wearables, industrial vision, and battery-powered sensors requiring larger on-device models.
SpacemiT has released the Vital Stone V100, claiming it is the first chip fully compliant with RISC-V server specifications. The processor features a heterogeneous design with 6 Xiangshan performance cores and 40 energy-efficient cores, supporting 8-channel DDR5-4800 memory and PCIe Gen5. It is fully compatible with the RVA23 Profile and includes server-grade virtualization and security capabilities.
Moonshot AI revealed that Kimi K3 achieves per-GPU throughput parity with NVIDIA HGX B300 on AMD Instinct MI355X accelerators using the AgentX framework. The report details bottleneck analysis and optimization decisions, demonstrating the viability of non-NVIDIA hardware for high-performance LLM inference.