Here is the Technical Intelligence Report for 2026-03-23.

Executive Summary

  • AMD Software Advancements: AMD officially released the FSR “Redstone” SDK 2.2, introducing ML-powered FSR Upscaling 4.1 and Ray Regeneration 1.1 specifically optimized for the new RDNA 4 architecture (Radeon RX 9000 Series).
  • Linux Kernel Optimization: AMD engineers submitted RFC patches for “pghot,” a new Linux kernel hot-page tracking subsystem that significantly improves memory promotion/demotion on EPYC Zen 5 servers utilizing CXL memory tiers.
  • Competitor Software ecosystem: Intel successfully pressured Crimson Desert developer Pearl Abyss into reversing their decision to drop Intel Arc GPU support, highlighting the increasing competitiveness in the discrete and integrated GPU market.
  • Hardware Durability: A viral hardware repair story highlighted an Nvidia RTX 4060 (titled as 5060) surviving a house fire with an intact PCB, showcasing the physical resilience of modern GPU board designs.

🤖 ROCm Updates & Software

[2026-03-23] Welcome to the AMD FSR SDK 2.2, now available on GPUOpen

Source: AMD GPUOpen

Key takeaway relevant to AMD:

  • This release marks a critical transition for AMD’s rendering ecosystem by fully integrating machine learning (ML) capabilities for RDNA 4 hardware while maintaining analytical fallbacks for older architectures, ensuring developers can target the entire AMD hardware stack with a single implementation.
  • AMD is retiring the “FidelityFX” naming convention, unifying all rendering features under the AMD FSR brand to simplify integration and consumer recognition.

Summary:

  • AMD launched FSR SDK 2.2 (codenamed “Redstone”), updating its suite of ML-based rendering technologies including Upscaling, Frame Generation, Ray Regeneration, and Radiance Caching.
  • The update primarily features FSR Upscaling 4.1 and FSR Ray Regeneration 1.1, bringing improved image quality, new debug modes, and memory optimizations to supported titles like Crimson Desert.

Details:

  • Architecture Support: The ML-powered features are heavily optimized for AMD RDNA 4 architecture (Radeon RX 9000 Series). An analytical fallback mode (FSR 3) is provided for RDNA 3.5 and earlier graphics cards.
  • FSR Upscaling 4.1 Changes: Introduces an updated inference model resulting in sharper images, particularly in-motion. It features specific improvements to ultra-performance and dynamic resolution scaling (DRS) modes.
  • FSR Ray Regeneration 1.1 Changes: Adds new debug view modes and implements quality and memory improvements. Note for developers: Version 1.1 contains ABI breaking changes from version 1.0.
  • Unreal Engine Integration: The current AMD FSR Unreal Engine 5 plugin currently utilizes FSR Upscaling 4.0.3 and will be updated to v4.1 at a later date.
  • Hardware Context: Example captures used a Radeon RX 9070 XT combined with a Ryzen 7 9800X3D running Crimson Desert at 4K Cinematic settings using FSR Upscaling 4.1 (Native AA) + FSR Ray Regen 1.1.

[2026-03-23] AMD Posts Latest “pghot” Code For Overhauling Linux Hot Page Tracking & Promotion

Source: Phoronix (AMD Linux)

Key takeaway relevant to AMD:

  • This Linux kernel update directly benefits AMD’s enterprise hardware ecosystem, specifically modern AMD EPYC servers utilizing CXL and multi-tier memory.
  • By upstreaming “pghot,” AMD is actively ensuring that Linux environments can natively and efficiently leverage EPYC’s advanced memory topologies without performance bottlenecks.

Summary:

  • AMD engineer Bharata Rao posted Request for Comments (RFC) patches for “pghot,” a new hot-page tracking and promotion subsystem for the Linux kernel.
  • The framework unifies hot page detection, centralizes promotion logic, and efficiently manages memory migrations between different tiers (e.g., NUMA nodes and CXL).

Details:

  • Performance Metrics: Initial benchmark testing on an AMD EPYC Zen 5 server (configured with two CPU NUMA nodes and one CXL node) demonstrated confirmed time savings during page promotion and in over-committed top-tier memory scenarios (mixed promotion/demotion).
  • Data Structure: Hotness parameters are maintained in a per-PFN hotness record within the mem_section data structure.
  • Default Mode Constraints: Uses 1 byte (u8) per hotness record. 5 bits are allocated to store time, utilizing a bucketing scheme that represents up to 4 seconds of total access time (at HZ=1000). The default top-tier NID is 0.
  • Precision Mode Constraints: Uses 4 bytes (u32) per record. 14 bits are allocated to store time, representing approximately 16 seconds (at HZ=1000). This mode additionally tracks the accessing NUMA node ID (NID).
  • Migration Logic: Pages breaching configurable thresholds are flagged with a “ready bit” (using the MSB of the hotness record). Per-lower-tier-node kmigrated threads routinely scan for this bit to execute batched migrations.

🤼‍♂️ Market & Competitors

[2026-03-23] Crimson Desert devs apologize for ‘confusion’ over Intel GPU FAQ — backtracks over prior dismissive language regarding Arc graphics support

Source: Tom’s Hardware (GPUs)

Key takeaway relevant to AMD:

  • Intel is aggressively lobbying developers to support its Arc GPU and iGPU architectures, meaning AMD must maintain its strong developer relations program to ensure Radeon optimization remains a priority.
  • With Intel’s Core Ultra 300H series (Arc B390) acting as a minimum-spec competitor, AMD’s own Ryzen AI and integrated RDNA graphics will face stiffer competition in the mobile gaming sector.

Summary:

  • Developer Pearl Abyss officially apologized for a previous FAQ stating that Crimson Desert would not support Intel Arc GPUs and advising Intel users to seek refunds.
  • Following intervention and offers of technical assistance from Intel, the studio confirmed they are now actively working on compatibility and optimization for Arc GPU systems.

Details:

  • Competitor Hardware Performance: The article highlights that Intel’s lower-wattage Core Ultra 300H series mobile processors with Arc B390 integrated graphics can match the performance of a classic Nvidia GTX 1060.
  • Game Requirements: The Nvidia GTX 1060 is noted as the official minimum specification to run Crimson Desert.
  • Industry Dynamics: Intel stated they reached out “many times” to offer collaborative efforts to Pearl Abyss to ensure support, demonstrating Intel’s reliance on proactive engineering outreach to secure market share for its emerging graphics division.

[2026-03-23] RTX 5060 in a house fire suffers melted shroud and fans, but survives with PCB intact — scorched GPU just needed a cleanup and a new cooler for full restoration

Source: Tom’s Hardware (GPUs)

Key takeaway relevant to AMD:

  • While primarily a human-interest hardware story about a competitor, it highlights the extreme thermal resilience of modern multi-layer GPU PCBs, a metric relevant to all GPU manufacturers including AMD.

Summary:

  • A user’s Asus graphics card (noted as an RTX 5060 in the title but identified as an RTX 4060 internally) survived a severe room fire in China.
  • Despite melted plastic, oxidized metal, and heavy soot, the underlying PCB and core remained entirely functional after basic cleaning and a cooler replacement.

Details:

  • Damage Assessment: The left side of the fan shroud melted, the backplate suffered severe oxidation burns, and the internal casing warped.
  • Restoration Process: The PCB was scrubbed with liquid soap and a toothbrush under tap water to remove smoke residue.
  • Testing Metrics: Fitted with a modded RTX 4060 Ti cooler, the restored card was tested in Furmark, peaking at an excellent thermal operating temperature of roughly 56 degrees Celsius with no internal shorts detected.

📈 GitHub Stats

Category Repository Total Stars 1-Day 7-Day 30-Day
AMD Ecosystem AMD-AGI/GEAK-agent 80 +1 +4 +15
AMD Ecosystem AMD-AGI/Primus 82 0 0 +8
AMD Ecosystem AMD-AGI/TraceLens 64 0 +1 +5
AMD Ecosystem ROCm/MAD 32 0 +1 +1
AMD Ecosystem ROCm/ROCm 6,275 +1 +25 +95
Compilers openxla/xla 4,107 +3 +29 +105
Compilers tile-ai/tilelang 5,415 +6 +44 +183
Compilers triton-lang/triton 18,737 +18 +70 +278
Google / JAX AI-Hypercomputer/JetStream 417 0 +2 +7
Google / JAX AI-Hypercomputer/maxtext 2,184 +2 +13 +40
Google / JAX jax-ml/jax 35,191 +12 +88 +275
HuggingFace huggingface/transformers 158,296 +58 +385 +1521
Inference Serving alibaba/rtp-llm 1,073 0 +6 +24
Inference Serving efeslab/Atom 336 0 0 0
Inference Serving llm-d/llm-d 2,673 +13 +50 +157
Inference Serving sgl-project/sglang 24,913 +47 +284 +1288
Inference Serving vllm-project/vllm 74,070 +140 +770 +3225
Inference Serving xdit-project/xDiT 2,571 -1 +4 +27
NVIDIA NVIDIA/Megatron-LM 15,770 +9 +98 +534
NVIDIA NVIDIA/TransformerEngine 3,233 0 +21 +64
NVIDIA NVIDIA/apex 8,936 -1 +6 +10
Optimization deepseek-ai/DeepEP 9,062 +2 +15 +70
Optimization deepspeedai/DeepSpeed 41,881 +12 +62 +238
Optimization facebookresearch/xformers 10,386 +2 +17 +40
PyTorch & Meta meta-pytorch/monarch 998 +2 +9 +23
PyTorch & Meta meta-pytorch/torchcomms 351 0 +2 +14
PyTorch & Meta meta-pytorch/torchforge 653 +1 +9 +32
PyTorch & Meta pytorch/FBGEMM 1,547 +2 +4 +12
PyTorch & Meta pytorch/ao 2,741 +1 +10 +47
PyTorch & Meta pytorch/audio 2,846 +1 +3 +15
PyTorch & Meta pytorch/pytorch 98,529 +42 +216 +885
PyTorch & Meta pytorch/torchtitan 5,175 +4 +29 +92
PyTorch & Meta pytorch/vision 17,584 +2 +18 +60
RL & Post-Training THUDM/slime 4,910 +17 +117 +630
RL & Post-Training radixark/miles 1,006 +5 +32 +114
RL & Post-Training volcengine/verl 20,136 +32 +195 +842