Search AI Chips News Today

Chipmakers

Meta MTIA & AMD Instinct MI325X to MI355X: Custom AI Chips and 2025–2026 GPU Roadmaps

Inside Meta AI chips (MTIA v2/v3) and AMD Instinct AI chips from late-2025 MI325X (256GB HBM3E) to 2026 MI355X CDNA 4 and MI400 rack architectures.

A80386DX-25 SX218. Photo: Mister rf (CC BY-SA 4.0, Wikimedia Commons).

For engineers tracking Meta AI chips news today and AMD news today on AI chips, the partnership between hyperscale infrastructure operators and merchant GPU architectures illustrates how large-scale AI fleets are actually built across 2025 and 2026. Rather than relying on a single processor family for every task, Meta operates a dual-track silicon architecture: deploying its in-house Meta Training and Inference Accelerator (MTIA) for high-volume ranking, recommendation and targeted generative inference while scaling massive clusters of merchant GPUs — including AMD Instinct MI300X, MI325X and MI355X alongside NVIDIA Blackwell — for Llama foundation model training and long-context serving.

Meta AI chips news today: MTIA v2 and next-generation MTIA v3 custom silicon

Meta’s internal silicon engineering team designed the MTIA (Meta Training and Inference Accelerator) family to eliminate the utilization mismatch between merchant GPUs and deep learning recommendation models (DLRMs):

  • MTIA v2 (5nm TSMC process): Built around an 8×8 grid of processing elements with 128 MB of on-chip SRAM and up to 128 GB of LPDDR5 memory operating at a 90 W thermal design power (TDP). Compared with first-generation MTIA v1, MTIA v2 more than tripled dense and sparse compute bandwidth while integrating into 72-accelerator rack chassis running PyTorch 2.x via the TorchInductor compiler.
  • MTIA v3 for generative AI & HBM integration: As generative AI inference and reasoning workloads grow across Meta’s platforms, next-generation MTIA silicon expands into HBM-backed package designs co-engineered with foundry and custom ASIC partners on 3nm nodes, targeting both low-latency Llama token decoding and distributed training phases.

AMD AI chips roadmap: from late-2025 / December 2025 MI325X to 2026 MI355X and MI400

Readers following AMD AI chips news today across 2025, December 2025 and 2026 have watched AMD transition onto an aggressive annual cadence designed to out-capacity competing accelerators on High Bandwidth Memory (HBM) per socket:

  • 2025 & December 2025 milestone — AMD Instinct MI300X and MI325X: Throughout 2025 and into the Q4 / December 2025 production ramp, AMD’s Instinct MI325X upgraded the OAM (OCP Accelerator Module) footprint to 256 GB of HBM3E memory delivering 6.0 TB/s of bandwidth. That memory advantage allowed hyperscalers including Meta and Microsoft to serve large Llama 405B and Mixture-of-Experts (MoE) models on fewer nodes without aggressive tensor splitting.
  • CDNA 4 architecture — AMD Instinct MI355X: Fabricated on TSMC’s 3nm (N3P) process node, the Instinct MI355X introduces native hardware support for FP4 and FP6 low-precision datatypes alongside 288 GB of HBM3E memory at up to 8.0 TB/s bandwidth within a 1,400 W direct-to-chip liquid-cooled envelope. AMD reports up to a 35× generational leap in low-precision AI inference throughput compared with CDNA 3.
  • 2026 rack-scale transition — AMD Instinct MI400 & UALink / Pollara fabrics: In 2026, AMD’s Instinct MI400 series shifts focus from standalone 8-GPU baseboards to integrated rack-scale architectures pairing next-generation CDNA compute chiplets, HBM4 memory stacks, AMD EPYC “Venice” host CPUs, Pensando Pollara 400G/800G AI NICs and open UALink / Ultra Ethernet scale-up and scale-out fabrics.
Accelerator Generation Deployment Window Memory Subsystem & Bandwidth Key Architectural Milestone
Meta MTIA v2 / v3 2025–2026 Hyperscale Fleet 128 MB on-die SRAM + LPDDR5 / HBM3E tiers Custom 8×8 PE mesh optimized for DLRM ranking & Llama inference via PyTorch TorchInductor.
AMD Instinct MI325X (CDNA 3) Late 2025 / Dec 2025 Ramp 256 GB HBM3E @ 6.0 TB/s Drop-in OAM socket compatibility expanding single-node LLM KV-cache capacity by 33% over MI300X.
AMD Instinct MI355X (CDNA 4) 2025–2026 Volume Production 288 GB HBM3E @ 8.0 TB/s 3nm compute chiplets + native FP4/FP6 matrix engines up to 1.4 kW liquid-cooled TDP.
AMD Instinct MI400 Series 2026 Rack-Scale Roadmap HBM4 (2,048-bit interface stacks) Full rack-scale co-design with UALink scale-up and Ultra Ethernet scale-out networking.

Takeaways for AI hardware architects and enterprise buyers

  • HBM capacity per GPU directly reduces tensor-parallel overhead: Moving from 192 GB to 256 GB (MI325X) or 288 GB (MI355X) keeps larger weight matrices and long-context KV caches local to fewer GPUs, cutting inter-chip communication latency.
  • ROCm software maturity is now validated in production: Upstream support in vLLM, SGLang, PyTorch and OpenAI Triton allows engineering teams to benchmark AMD Instinct and NVIDIA Blackwell side-by-side on identical model checkpoints.

Compare measured AMD vs NVIDIA workload telemetry in our InferenceMAX v1 benchmark analysis, explore the Chipmakers coverage hub, or review our guide to choosing HBM memory for AI workloads.