DigitalOcean、AMD 和 Character.ai 将 AI 推理性能提升一倍，重新定义云 AI

DigitalOcean、AMD 和 Character.ai 携手实现了 AI 推理吞吐量的 2 倍显着提升，大幅降低了成本，并为大规模、低延迟的 AI 应用树立了新标准。

A Quantum Leap in AI Efficiency

人工智能效率的巨大飞跃

In a landscape where AI innovation is moving at warp speed, the efficiency of inference—the process by which AI models make predictions—is paramount. DigitalOcean, a leading cloud provider, has teamed up with semiconductor giant AMD and AI entertainment platform Character.ai to deliver a significant breakthrough. Their joint effort has resulted in a staggering twofold increase in production inference throughput for Character.ai’s applications, all while substantially reducing operational costs.

在人工智能创新飞速发展的背景下，推理效率（人工智能模型进行预测的过程）至关重要。领先的云提供商 DigitalOcean 与半导体巨头 AMD 和人工智能娱乐平台 Character.ai 合作，实现了重大突破。他们的共同努力使 Character.ai 应用程序的生产推理吞吐量实现了惊人的两倍增长，同时大幅降低了运营成本。

Character.ai, serving approximately 20 million users globally, faced the quintessential challenge of scaling AI with demanding low-latency requirements. Their quest for optimized GPU performance and cost efficiency led them to DigitalOcean and AMD. What ensued was a deep, multi-team technical collaboration that harnessed the raw power of AMD Instinct™ MI300X and MI325X GPU platforms hosted on DigitalOcean's robust infrastructure.

Character.ai 为全球约 2000 万用户提供服务，面临着扩展 AI 并满足低延迟要求的典型挑战。他们对优化 GPU 性能和成本效率的追求促使他们选择了 DigitalOcean 和 AMD。随之而来的是深入的多团队技术合作，利用了在 DigitalOcean 强大的基础设施上托管的 AMD Instinct™ MI300X 和 MI325X GPU 平台的原始能力。

The Art of Optimization: A Technical Deep Dive

优化的艺术：技术深入探讨

Achieving this 2x performance gain wasn't merely a matter of throwing more hardware at the problem. It was a testament to sophisticated engineering and a meticulous approach to software-hardware co-design. The teams focused on platform-level optimizations, including:

实现 2 倍的性能提升并不仅仅是投入更多硬件来解决问题。它证明了复杂的工程和细致的软硬件协同设计方法。这些团队专注于平台级优化，包括：

• Clever parallelization strategies tailored for large Mixture-of-Experts (MoE) models, a technique critical for handling complex AI architectures.
• The implementation of efficient FP8 execution paths, leveraging AMD Instinct GPUs’ native support for this precision to reduce VRAM usage by approximately 50% and enhance throughput.
• The integration of optimized kernels through AITER (AI Tensor Engine for ROCm), AMD's high-performance AI operator library, ensuring peak hardware efficiency.
• Topology-aware GPU allocation and production-ready Kubernetes orchestration via DigitalOcean Kubernetes (DOKS), simplifying deployment and management of intensive GPU workloads.

Specifically, the optimization of the Qwen3-235B Instruct FP8 model saw a transition from generic, non-optimized setups to advanced vLLM recipes like DP2 / TP4 / EP4 configurations. This nuanced approach, balancing distributed serving, tensor parallelism, and expert parallelism, proved to be about 91% more efficient in throughput compared to prior, less optimized deployments, directly translating to a substantial reduction in cost-per-token.

具体来说，Qwen3-235B Instruct FP8 模型的优化实现了从通用、非优化设置到高级 vLLM 配方（如 DP2 / TP4 / EP4 配置）的转变。这种微妙的方法平衡了分布式服务、张量并行性和专家并行性，事实证明，与之前优化程度较低的部署相比，吞吐量效率提高了约 91%，这直接导致了每个代币成本的大幅降低。

Beyond Benchmarks: The New AI Systems Paradigm

超越基准：新的人工智能系统范式

This collaboration underscores a crucial evolution in AI infrastructure. The findings highlight a "New AI Systems Paradigm" where success hinges on several foundational shifts:

此次合作凸显了人工智能基础设施的重要演变。研究结果强调了“新人工智能系统范式”，其中成功取决于几个基本转变：

• Multi-Dimensional Optimization: Performance now demands a delicate balance across cost, latency, throughput, and concurrency, with strategic architectural choices driving down expenses while boosting capabilities.
• Hardware-Software Co-Design: Peak efficiency isn't accidental. It requires a precise alignment between low-level system architecture—GPU interconnects, memory bandwidth, FLOPs utilization—and the AI model serving stack. The detailed work on vLLM configurations and AITER is a prime example of this synergy.
• Infrastructure Reinvention: Deploying large-scale models in data centers is a distinct challenge from traditional web services. It necessitates a "full-stack" re-evaluation of deployment strategies, from managed Kubernetes (like DOKS, which now offers pre-baked GPU drivers and device plugins) to smart caching solutions for massive model weights.

For Character.ai, this optimized infrastructure means predictable scaling without increased operational burden, a factor so compelling it led to a multi-year, eight-figure annual agreement with DigitalOcean for GPU services. It's a win-win, proving that cutting-edge AI doesn't have to break the bank or strain engineering teams.

对于 Character.ai 来说，这种优化的基础设施意味着可预测的扩展，而不会增加运营负担，这一因素非常引人注目，以至于与 DigitalOcean 就 GPU 服务达成了为期多年、八位数的年度协议。这是双赢的结果，证明尖端人工智能不必倾家荡产或给工程团队带来压力。

The Future Looks Bright (and Fast)

未来看起来光明（而且快速）

As DigitalOcean continues to build out its inference cloud, partnerships like these with industry titans AMD and innovative platforms like Character.ai are charting a course for the future of AI. The message is clear: deep technical collaboration, meticulous optimization, and a holistic view of the AI stack can unlock extraordinary performance. It seems the digital ocean is getting both deeper and a good deal faster, making complex AI tasks feel as breezy as a walk through Central Park. We're all for it.

随着 DigitalOcean 继续构建其推理云，与行业巨头 AMD 和 Character.ai 等创新平台的合作正在为人工智能的未来制定路线。传达的信息很明确：深入的技术协作、细致的优化以及人工智能堆栈的整体视图可以释放非凡的性能。数字海洋似乎变得越来越深，速度也越来越快，使得复杂的人工智能任务感觉就像在中央公园散步一样轻松。我们都支持它。