Executive Summary

  • AMD Tooling & Software Advances: AMD released “Smoldr,” a new open-source scripting tool that bypasses C++ requirements for testing DirectX 12 shaders, and launched MLIR-AIE v1.3, which introduces a high-performance C++ compiler for running non-AI/ML and DSP workloads on Ryzen AI NPUs.
  • AMD Hardware Ecosystem: System76 debuted the Thelio Mira desktop, confirming high-end Linux market adoption of AMD’s Zen 5 architecture (Ryzen 9 9950X/X3D) and next-generation Radeon RX 9070 XT GPUs.
  • Nvidia’s Aggressive Data Center Expansion: At GTC 2026, Nvidia unveiled a rapidly accelerating hardware roadmap featuring the 2028 “Feynman” GPU (utilizing die stacking and custom C-HBM4E) and “Rosa” CPUs. Nvidia confirmed a $20 billion acquihire of Groq to natively integrate low-latency LPUs into the Vera-Rubin platform to dominate agentic AI inference.
  • Telecom & Spatial Computing Pushes: Nvidia is fortifying its enterprise moat by partnering with major telecoms to build “AI Grids” using Blackwell GPUs at the network edge, and by integrating CloudXR 6.0 natively into Apple Vision Pro for spatial computing.
  • DLSS 5 Backlash & Hardware Demands: Nvidia faced community pushback over its neural rendering tech (DLSS 5), which CEO Jensen Huang defended as “generative control at the geometry level.” Hands-on previews revealed massive computational demands, requiring dual RTX 5090s to run the tech during demos.

🤖 ROCm Updates & Software

[2026-03-17] Enhancing DirectX Testing with AMD Smoldr

Source: AMD GPUOpen

Key takeaway relevant to AMD:

  • AMD is significantly lowering the barrier to entry for graphics developers to test DirectX 12 shaders on AMD hardware by providing a lightweight, script-based alternative to heavy C++ boilerplates.

Summary:

  • AMD released Smoldr, an open-source command-line scripting tool designed to run DX12 shaders on GPUs using simple text input files.
  • The tool handles the compilation of HLSL shaders, pipeline creation, resource allocation, and compute dispatch without requiring traditional C++ application development.

Details:

  • Technical Implementation: Compiles HLSL shaders through DXIL. Currently optimized for compute shaders and fully supports raytracing pipelines.
  • Upcoming Features: Support for mesh and amplification shaders is in development to handle modern graphics rendering pipelines.
  • Integration: Supports the Microsoft Agility SDK for experimental HLSL features.
  • Profiling Capabilities: Integrates seamlessly with the AMD Radeon GPU Profiler by using the --window parameter to spawn a window and execute scripts frame-by-frame.
  • Developer Impact: Greatly speeds up rapid prototyping and debugging on AMD hardware, providing specific error pointers for typos and pipeline failures.

[2026-03-17] AMD MLIR-AIE Releases New AIECC C++ Compiler To Help Bring New Workloads To Ryzen AI NPUs

Source: Phoronix (AMD Linux)

Key takeaway relevant to AMD:

  • AMD is expanding the utility of its Ryzen AI NPUs far beyond traditional AI inference, opening the hardware to digital signal processing (DSP) and custom edge workloads via a versatile LLVM/MLIR toolchain.

Summary:

  • AMD released MLIR-AIE v1.3, a compiler toolchain for AMD AI Engine devices (Ryzen AI NPUs and AMD-Xilinx accelerators).
  • The update introduces aiecc, a new C++ compiler that replaces the project’s existing Python-based tooling for improved performance and robust feature support.

Details:

  • New Compiler Architecture: The aiecc tool acts as the main AIE compiler driver. It utilizes the LLVM CommandLine library, executes MLIR transformation pipelines, handles core compilation (xchesscc/peano), and generates NPU instructions and CDO/PDI/xclbin files.
  • Standards & Compatibility: Supports C++17 and brings early native Windows support alongside existing Linux workflows.
  • Vector Improvements: Introduces specific AIE2 vector enhancements for BF16, FMA (Fused Multiply-Add), reductions, and tanh functions.
  • Ecosystem Synergy: The MLIR route allows cross-vendor versatility and cross-hardware targeting (e.g., Radeon/Instinct via ROCm, CPUs via LLVM).

🔲 AMD Hardware & Products

[2026-03-17] System76 New Thelio Mira Linux Desktop Running Strong - Powered By AMD Ryzen 9000 Series

Source: Phoronix (AMD Linux)

Key takeaway relevant to AMD:

  • AMD’s Zen 5 and next-gen RDNA architecture continue to be the hardware of choice for premium, pre-built Linux workstation vendors, reinforcing AMD’s strong positioning in the Linux desktop ecosystem.

Summary:

  • System76 officially launched the newly redesigned Thelio Mira Linux desktop chassis, featuring AMD’s latest Ryzen 9000 series processors.
  • The system offers improved thermals and highly configurable, top-tier AMD CPU and GPU options tailored for power users and developers.

Details:

  • Processor Support: Configurable up to the AMD Ryzen 9 9950X3D (16-core, 3D V-Cache) and Ryzen 9 9950X (Zen 5).
  • Graphics Support: The “Elite” configuration explicitly ships with the newly mentioned Radeon RX 9070 XT graphics card.
  • System Specs: Supports up to 192GB of DDR5-5600 memory and up to 28TB of PCIe NVMe solid-state storage.
  • Chassis Improvements: Features a new “Precision Industrialism” design with front panel I/O, improved repairability, and an internal layout that reduces operating temperatures by 13 degrees.

🤼‍♂️ Market & Competitors

[2026-03-17] Nvidia updates data center roadmap with Rosa CPU and stacked Feynman GPUs

Source: Tom’s Hardware (GPUs)

Key takeaway relevant to AMD:

  • Nvidia is accelerating its CPU development cycle to two years to match AMD, and is transitioning to die stacking and co-packaged optical interconnects by 2028, significantly raising the bandwidth and scalability bar AMD’s Instinct line will have to meet.

Summary:

  • At GTC 2026, Nvidia updated its data center roadmap to include the 2026 Vera Rubin platform, 2027 Rubin Ultra, and the newly announced 2028 Feynman GPU and Rosa CPU architectures.
  • The roadmap officially integrates Groq LPUs into Nvidia’s standard rack architecture to accelerate low-latency inference.

Details:

  • 2026 (Vera Rubin): Features Vera CPU, Rubin GPU, Groq LP30 inference accelerator, BlueField-4 DPU, and NVLink-6. (Rubin CPX processors were dropped in favor of Groq LPUs).
  • 2027 (Rubin Ultra): Features four compute chiplets, 1 TB of HBM4E memory, Groq LP35 (with NVFP4 data format support), and NVLink 7. Kyber NVL144 racks will pack 144 Rubin Ultra GPUs.
  • 2028 (Feynman & Rosa): The Feynman GPU will transition to die stacking and custom high-bandwidth memory (C-HBM4E) exceeding 1 TB per package.
  • Rosa CPU: An in-house processor heavily focused on single-thread performance, signaling Nvidia’s shift to a 2-year CPU cadence.
  • Optical Scaling: Introduction of NVLink switches with co-packaged optics (CPO), allowing rack-scale systems to scale up to 1152 GPU packages (Kyber chassis).

[2026-03-17] Nvidia Finally Admits Why It Shelled Out $20 Billion For Groq

Source: The Next Platform

Key takeaway relevant to AMD:

  • Nvidia recognized it was vulnerable in the low-latency, agentic AI inference market and bought its way out via Groq. AMD must aggressively leverage its relationships with remaining inference specialists (like Cerebras) to maintain a competitive counter-offering.

Summary:

  • Nvidia acquired Groq for $20 billion (structured as an “acquihire”) to rapidly integrate Language Processing Units (LPUs) into its Vera-Rubin platform.
  • The LP30 inference accelerators will replace the previously announced Rubin CPX, pairing traditional GPU threshing capabilities with LPU speed-demon latency.

Details:

  • Hardware Shift: General-purpose GPUs (like the R200) are excellent for batching through HBM, but statically scheduled, SRAM-heavy LPUs (like Groq’s LP30) are vastly superior for low-latency token generation required by agentic AI.
  • Performance Metrics: The R200 boasts 21X the peak FP8 theoretical performance of the LP30, and up to 42X if leveraging FP4 formats. However, the LP30’s SRAM architecture drastically reduces latency and cost-per-token for single-batch interactivity.
  • System Architecture: Nvidia’s VR200 NVL72 rackscale systems enable 72-way memory sharing. Introducing Groq LPUs allows Nvidia to offer hybrid “premium/ultra tiers” tailored strictly for agent-to-agent inference.
  • Market Implication: Intel missed SambaNova, and Nvidia secured Groq. The article notes AMD’s close ties with Cerebras as the next logical industry maneuver.

[2026-03-17] Jensen Huang says gamers are ‘completely wrong’ about DLSS 5 — Nvidia CEO responds to DLSS 5 backlash

Source: Tom’s Hardware (GPUs)

Key takeaway relevant to AMD:

  • Nvidia is facing consumer friction over generative AI altering artistic intent in games. AMD has an opportunity to market FSR as a high-fidelity, non-intrusive alternative that preserves original developer aesthetics without generative hallucination.

Summary:

  • Nvidia CEO Jensen Huang dismissed user criticism that DLSS 5 (neural rendering) makes games look homogenous or like “AI slop.”
  • Huang argued that DLSS 5 provides geometry-level generative control rather than just frame-level post-processing, keeping artistic control entirely in the hands of developers.

Details:

  • Technological Approach: Huang clarified that DLSS 5 fuses the game’s actual geometry and textures with generative AI, meaning developers can fine-tune the model to match specific stylistic visions (e.g., a “toon shader”).
  • Community Reaction: Vocal critics have cited updated appearances of characters (like Resident Evil’s Leon Kennedy) as looking unnatural.
  • Release Window: DLSS 5 is slated to officially launch in the fall of 2026.

[2026-03-17] We got a first look at Nvidia’s DLSS 5 and the future of neural rendering at GTC

Source: Tom’s Hardware (GPUs)

Key takeaway relevant to AMD:

  • DLSS 5 requires massive, brute-force AI compute to function (running on dual RTX 5090s). If neural rendering becomes an industry standard, AMD will need to drastically scale its matrix math acceleration (AI accelerators) on consumer RDNA cards.

Summary:

  • Tom’s Hardware previewed DLSS 5 across five games, noting highly impressive, photorealistic lighting and geometry enhancements, but occasionally jarring “uncanny valley” results on older character models.
  • The technology operates deep within the pipeline, using inputs like color buffers and motion vectors to infer how elements should look in reality.

Details:

  • Compute Requirements: The GTC demos were running on systems equipped with dual RTX 5090 GPUs—one for the game rendering, and one dedicated solely to accelerating the DLSS 5 AI model.
  • Visual Enhancements: Added realistic rim lighting, ambient occlusion, accurate shadows under objects, and corrected native engine rendering errors (e.g., Assassin’s Creed Shadows and Hogwarts Legacy).
  • SDK Integration: Implemented via Nvidia’s existing Streamline SDK, offering developers controls for color grading, intensity, and masking.
  • Flaws: Over-rendering exaggerated facial features in older games (like Oblivion Remastered) resulted in uncomfortable, un-stylized realism.

[2026-03-17] NVIDIA RTX-Accelerated Computers Now Connect Directly to Apple Vision Pro

Source: NVIDIA Blog

Key takeaway relevant to AMD:

  • Nvidia is successfully bridging its workstation GPU monopoly with Apple’s high-end spatial computing ecosystem, securing the enterprise XR and digital twin markets.

Summary:

  • Nvidia announced that its CloudXR 6.0 suite natively integrates with Apple Vision Pro via visionOS 26.4.
  • The integration allows professionals to stream uncompromised 4K, low-latency spatial computing workflows directly from RTX PRO workstations or GeForce RTX GPUs.

Details:

  • Streaming Tech: Utilizes “dynamic foveated streaming” to optimize rendering resolution based on the user’s approximate gaze point, maximizing bandwidth efficiency while protecting privacy (gaze data isn’t exposed to the app).
  • Use Cases: Used for Autodesk VRED design reviews (Kia, BMW, Volvo), pharmaceutical biofluid lab simulations (Roche), and factory floor walkthroughs (Foxconn).
  • Developer Access: CloudXR 6.0 is available as a native streaming framework for Swift within Apple’s Xcode IDE.

[2026-03-17] NVIDIA, Telecom Leaders Build AI Grids to Optimize Inference on Distributed Networks

Source: NVIDIA Blog

Key takeaway relevant to AMD:

  • Nvidia is decentralizing AI compute by turning global telecom infrastructure into an extended AI moat. AMD must target telecom providers with its EPYC and Instinct platforms to prevent being locked out of the lucrative edge-inference market.

Summary:

  • Nvidia is partnering with leading telecoms (AT&T, Comcast, Spectrum, Akamai, T-Mobile, Indosat) to build “AI Grids” using existing distributed network edge data centers.
  • These grids push AI inference to the network edge, drastically lowering latency and cost-per-token for real-time applications.

Details:

  • Hardware Stack: Built on the NVIDIA AI Grid Reference Design using NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs and AI-RAN integration.
  • Performance Metrics: Enables conversational AI agents to achieve sub-500ms end-to-end latency and over 50% lower cost-per-token (Personal AI), and video generation streams with sub-12ms network latency (Decart).
  • Scale: Telecoms possess roughly 100,000 distributed data centers globally with spare power capable of adding 100+ gigawatts of new AI capacity over time.

💬 Reddit & Community

[2026-03-17] I want buy a new GPU from AMD under 350$

Source: Reddit AMDGPU

Key takeaway relevant to AMD:

  • Despite the high-end focus of the current market news cycle, there remains consistent consumer demand for AMD’s mid-range, value-oriented graphics cards.

Summary:

  • A community discussion was initiated regarding recommendations for a new AMD GPU priced under the $350 threshold. (Note: Full thread content was blocked by network policy restrictions during extraction).

Details:

  • The title indicates active consumer interest in AMD’s budget/mid-range tier, an area historically dominated by cards like the RX 7600/7600 XT, which are vital for maintaining overall market share against Nvidia’s lower-end hardware.