🖥️ AI & GPU Industry Weekly Recap: January 12–18, 2026

🔑 Key Highlights

  • NVIDIA & Eli Lilly announce a landmark $1B AI co-innovation lab for drug discovery at the J.P. Morgan Healthcare Conference, powered by Lilly’s NVIDIA DGX SuperPOD with DGX B300 systems
  • AMD and Tata Consultancy Services (TCS) formalize a strategic AI collaboration, targeting enterprise-scale GenAI deployment across life sciences, manufacturing, and BFSI verticals using AMD Instinct GPUs, EPYC CPUs, and Ryzen AI PCs
  • NVIDIA DLSS 4.5 Super Resolution exits beta, rolling out to all NVIDIA app users with a second-generation transformer model supporting 400+ titles — using 5× the compute of the first-gen model
  • AMD’s ROCm team publishes major MI300X compute partitioning results, demonstrating up to 2.47× speedups for GROMACS molecular dynamics and 1.42× speedup for REINVENT4 AI drug design workflows using CPX mode
  • AMD releases Athena-PRM, a data-efficient multimodal Process Reward Model that achieves +10.2 points on WeMath and sets state-of-the-art on VisualProcessBench, running on AMD Instinct GPUs

🤖 AI & Machine Learning

AMD Athena-PRM: Multimodal Reasoning Breakthrough

AMD Research published Athena-PRM, a Process Reward Model (PRM) for enhancing reasoning in Large Vision-Language Models (LVLMs). Key technical innovations include:

  • Weak-Strong Consistency filtering: Uses both weak and strong completers to generate higher-quality step-level labels, dramatically reducing computational overhead vs. Monte Carlo estimation methods like Math-Shepherd
  • ORM Initialization + Negative Up-sampling: Two training strategies that further boost PRM performance
  • Results: With Qwen2.5-VL-7B as the policy model, Athena-PRM achieves +10.2 points on WeMath and +7.1 points on MathVista under Best-of-N (N=8) test-time scaling
  • Athena-PRM outperforms VisualPRM-8B across multiple benchmarks and sets new SoTA on VisualProcessBench (+3.9 F1-score over previous best)
  • Athena-7B, fine-tuned from Qwen2.5-VL-7B via reward-ranked fine-tuning, was trained entirely on AMD Instinct GPUs
  • The model was trained on just 5,000 high-quality labeled examples, making it notably data-efficient

NVIDIA + Lilly: $1 Billion AI Drug Discovery Lab

Jensen Huang (NVIDIA) and Dave Ricks (Lilly) announced a first-of-its-kind AI co-innovation lab at the J.P. Morgan Healthcare Conference:

  • Up to $1 billion joint investment over 5 years in talent, infrastructure, and compute
  • Lab will use a scientist-in-the-loop framework connecting agentic wet labs with computational dry labs in a continuous learning loop
  • Lilly’s AI supercomputer — an NVIDIA DGX SuperPOD with DGX B300 systems — serves as the foundation for training large-scale biomedical models
  • NVIDIA also announced BioNeMo platform expansions including: NVIDIA Clara open models for RNA structure prediction, BioNeMo Recipes for accelerating foundation model training, and nvMolKit (GPU-accelerated cheminformatics library)
  • Huang gifted DGX Spark systems to ~12 AI biology pioneers at the event, including leaders from Isomorphic Labs (AlphaFold), Chai Discovery, Recursion (OpenPhenom), and the Arc Institute (Evo 2)

⚡ GPU & Hardware

AMD MI300X Compute Partitioning: Benchmarked

A detailed ROCm blog post by David Björelind demonstrated major performance gains from using CPX (Core Partitioned X-celerator) mode on the AMD MI300X:

  • The MI300X’s 8 XCDs (Accelerator Complex Dies) can be exposed as 8 independent logical GPUs via a single amd-smi command
  • GROMACS (molecular dynamics): CPX mode delivered 1.75× to 2.47× speedups over SPX (single partition) across 1–8 GPU configurations. At 8 GPUs: 8,026 ns/day (CPX) vs. 4,507 ns/day (SPX)
  • REINVENT4 (drug design ML): Concurrent partitioned execution of 8 jobs achieved a 1.42× speedup (191 min vs. 272 min), with a crossover advantage beginning at 5+ concurrent jobs
  • Key takeaway: CPX mode is ideal for naturally parallel, multi-trial workloads (HPO, ensemble MD, parameter sweeps)

NVIDIA DLSS 4.5 Super Resolution GA Release

  • DLSS 4.5 Super Resolution — announced at CES 2026 — officially exited beta and began auto-rolling out to all NVIDIA app users
  • Features a second-generation transformer model using 5× the compute of the first-gen version
  • Supports 400+ games and applications across all GeForce RTX GPU generations (including RTX 20 and 30 series, with caveats)
  • Reported improvements: reduced shimmering on static surfaces, eliminated ghosting/after-image artifacts, richer particle/lighting effects
  • Note: RTX 20/30 series users may see performance overhead of 20%+; RTX 40/50 series Tensor Cores handle the workload more efficiently

Lilly’s NVIDIA DGX B300 SuperPOD

  • Eli Lilly’s dedicated AI supercomputer — described as “the biopharma industry’s most powerful AI factory” — is based on NVIDIA DGX B300 systems in a SuperPOD configuration, purpose-built for training biomedical foundation models

🏭 Industry & Market

AMD × TCS Strategic Enterprise AI Alliance

  • Tata Consultancy Services (TCS) (~590,000 employees, $30B+ revenue) and AMD announced a broad strategic collaboration targeting enterprise AI deployment at scale
  • TCS will integrate AMD Ryzen CPUs for AI PC workplace transformation and leverage AMD EPYC CPUs + AMD Instinct GPUs for hybrid cloud HPC modernization
  • AMD’s embedded portfolio (adaptive SoCs, FPGAs) will power TCS edge computing and industrial digitalization offerings
  • Industry-specific GenAI frameworks planned for: life sciences (drug discovery), manufacturing (cognitive quality engineering), BFSI (intelligent risk management)
  • TCS will upskill and certify associates on AMD hardware/software stacks — a meaningful software ecosystem development move for AMD

NVIDIA’s Healthcare Ecosystem Push

  • At the world’s largest healthcare investment conference (8,000+ attendees), NVIDIA positioned itself as the central compute platform for AI-driven biology, spotlighting partnerships with Thermo Fisher (autonomous lab infrastructure) and Multiply Labs (robotic cell therapy manufacturing)
  • The breadth of the BioNeMo announcements signals NVIDIA’s intent to own the full stack from GPU silicon to biology-specific AI tooling

AMD HIP on Europe’s LUMI Supercomputer

  • AMD’s ROCm team published a guide for installing HIP-enabled GROMACS on LUMI, Europe’s most powerful supercomputer — reinforcing AMD Instinct’s growing footprint in European HPC infrastructure

🛠️ Developer Ecosystem

AMD GPU DRA Driver (Beta) for Kubernetes

AMD published and then clarified the AMD GPU Dynamic Resource Allocation (DRA) Driver for Kubernetes — a significant step toward cloud-native GPU orchestration:

  • Enables AMD Instinct GPUs to be published as attribute-aware ResourceSlices in Kubernetes
  • Allows declarative ResourceClaims for specific GPU models and partition profiles (e.g., requesting two MI300X partitions on the same PCIe root)
  • Replaces the traditional Device Plugin framework with a richer, lifecycle-aware allocation model
  • Currently in beta — APIs and behaviors subject to change
  • Roadmap includes: dynamic workload-driven partitioning for fractional GPU allocations, cross-driver orchestration with NICs, and cluster-level resource managers for rack-scale deployments

ROCm Blog Activity: Developer Documentation Surge

AMD’s ROCm team had an exceptionally active week publishing developer content:

  • Athena-PRM blog: Full technical deep-dive on multimodal PRM training (GitHub repo: AMD-AGI/Athena-PRM)
  • GPU Compute Partitioning guide: Detailed MI300X CPX/SPX tutorial with GROMACS and REINVENT4 benchmarks
  • LUMI Supercomputer HIP-GROMACS guide: HPC installation tutorial for AMD ROCm on LUMI
  • GPU DRA Driver blog: Kubernetes-native GPU allocation using Dynamic Resource Allocation

NVIDIA BioNeMo Platform Expansion

New developer tools released/announced under the NVIDIA BioNeMo umbrella:

  • NVIDIA Clara open models: RNA structure prediction + synthesizability scoring for AI-designed drugs
  • BioNeMo Recipes: Accelerated biological foundation model training and deployment pipelines
  • nvMolKit: GPU-accelerated cheminformatics library for molecular design

📊 Key Takeaways

This week underscored the intensifying convergence of AI and life sciences as both NVIDIA (via the Lilly $1B co-innovation lab and BioNeMo expansions) and AMD (via TCS partnership, Athena-PRM, and MI300X drug discovery benchmarks) made major moves to capture the lucrative pharmaceutical AI compute market. AMD’s ROCm ecosystem showed notable maturation, with compelling MI300X CPX partitioning results (up to 2.47× speedups for HPC workloads) and new Kubernetes-native GPU orchestration tooling addressing real enterprise deployment gaps. Meanwhile, NVIDIA’s DLSS 4.5 Super Resolution GA launch — bringing a second-generation transformer model to 400+ titles — reinforces its dominance in consumer GPU software, even as both companies race to expand their AI infrastructure footprints at the hardware and platform level.

Sources: NVIDIA Blog, AMD Press Releases, ROCm Tech Blog, Tom’s Hardware | Week of Jan 12–18, 2026