Skymizer at SC26
Trillion-Parameter-Class AI in One Air-Cooled Server.
See HTX301 — the world’s first decode-first chip purpose-built for on-premise AI inference — at the MiTAC Computing booth #4615: eight HTX301 PCIe cards in a MiTAC G4520G6, and a video of the same server class running Kimi K3 (2.8T) on 2 TB of onboard DRAM.
- Nov 17–19, 2026SC26 exhibit days
- ChicagoMcCormick Place West · SC26, The International Conference for HPC, Networking, Storage and Analysis
- MiTAC Computing booth #4615Skymizer HTX301 with MiTAC G4520G6
Architecture
Heterogeneous Prefill / Decode Disaggregation
Prefill is typically compute-intensive. Decode repeatedly generates tokens and is increasingly constrained by memory bandwidth, capacity, and data movement. In a heterogeneous setup, a GPU system handles prefill and hands off to a server like the MiTAC G4520G6 with HTX301 for decode — each phase on the infrastructure suited to it, targeting trillion-parameter-class models.
- User Request
- Prefill: GPU system
- KV Transfer / Handoff
- Decode: MiTAC G4520G6 + HTX301
- Token Stream
Heterogeneous prefill/decode is one supported architecture option, not the only deployment mode.
In the booth
8× HTX301 PCIe Cards in MiTAC G4520G6
- Accelerator
- Skymizer HTX301 LPU PCIe card (four HTX301 chips per card)
- At the booth
- Static display: MiTAC G4520G6 with 8× HTX301 PCIe cards (128 GB onboard DRAM each)
- Demo video
- Kimi K3 (2.8T) on 8× HTX301 PCIe cards (256 GB onboard DRAM each, 2 TB total)
- Configuration
- 8 cards, 128 or 256 GB onboard DRAM each, up to 2 TB per server
- Deployment
- Air-cooled, standard rack, no cluster
- Software
- LISA™ toolchain, one ISA from 4B to trillion-scale
- Architecture
- Decode layer for heterogeneous prefill/decode
HTX301 platform
Built Around the Workload That Runs Every Day
- 128 / 256 GBOnboard DRAM per HTX301 PCIe card
- Up to 2 TBOnboard DRAM across 8 cards in one air-cooled MiTAC G4520G6
- Decode-FirstPurpose-built for large-model serving economics
- PCIe-BasedFrom enterprise systems to Mini Data Center scale
Maximum memory, model capacity, power, performance, and form factor depend on hardware configuration, model architecture, precision / quantization, software release, and system design.
Meet us at SC26
Book a Meeting at SC26
Bring your model and deployment requirements. We’ll walk you through the prefill/decode demo and the HTX301 configuration that fits your workload.
- Live walkthrough at the MiTAC Computing booth #4615
- HTX301 configuration & memory requirement for your model
- Software status & evaluation path
Prefer email? [email protected]