Back to articles
Compute & Chips

At WAIC 2026, Qiymor Puts Domestic GPU Direct Connect on Display

2 min read

Qiymor’s first appearance at WAIC 2026 was less about a single product launch and more about showing how domestic AI infrastructure may finally connect into a coherent system.

What was unveiled

  • a domestic full-stack supernode interconnect solution
  • an IBGDA demo connecting a domestic GPU directly to a domestic RDMA NIC
  • a CPO white paper for next-generation optical interconnect
  • a joint initiative to build a domestic supernode ecosystem

Why it matters

The bottleneck in domestic AI is increasingly system-level rather than chip-level. Even when individual chips improve, performance can still be limited by communication overhead, fragmented architectures, and inconsistent software-hardware integration. This becomes especially visible in MoE workloads, where frequent small-message exchanges make CPU-mediated communication expensive in both latency and overhead.

Qiymor’s IBGDA demo targets exactly this pain point. By letting the GPU initiate communication more directly, the design reduces the need for CPU handoffs and helps preserve GPU throughput in communication-heavy scenarios. The article describes the result as a clear gain in throughput and a substantial reduction in latency compared with the traditional approach.

Just as important, the solution is presented as part of an open Ethernet-based ecosystem. That matters because domestic AI clusters need performance, but they also need deployability, cost control, and compatibility across suppliers.

Bigger picture

The CPO white paper and the OSU-related roadmap point to the next stage of interconnect evolution: bringing optical links closer to the compute core. That is a sign that the industry is thinking beyond incremental fixes and toward a longer-term architecture for scale-up systems.

The ecosystem initiative is equally significant. As supernodes become the new battleground for AI infrastructure, no single vendor can solve the entire stack alone. Standards, protocols, software adaptation, and system RAS capabilities all need to move together.

In short, Qiymor’s WAIC showing suggests that domestic AI infrastructure is shifting from isolated component wins to integrated, ecosystem-driven competitiveness.

Source: 量子位

Comments

Checking sign-in status...

Loading comments...

Related articles

CCTest · Blog
Intellifusion’s AI Inference Roadmap Shifts the Focus From Peak FLOPS to Token Cost
Compute & Chips
cctest.ai
Compute & Chips

Intellifusion’s AI Inference Roadmap Shifts the Focus From Peak FLOPS to Token Cost

Intellifusion outlined a new inference-chip roadmap built around three DeepVerse processors optimized for Prefill, Decode and Decode FFN workloads. The strategy points to a broader industry shift toward heterogeneous clusters and lower token-generation costs.

Read more