News: 2026-02-22
February 22, 2026 · Generated 12:26 PM PT
Here is the Technical Intelligence Report for 2026-02-22.
Executive Summary
- Linux 7.0 Zen 6 Enablement: The Linux 7.0 kernel has merged initial support for AMD Zen 6 (Family 1Ah) performance events and metrics, preparing the ecosystem for next-generation profiling tools.
- Legacy Hardware Optimization: Valve’s open-source team has resolved critical
amdgpudriver bugs affecting legacy GCN 1.1 GPUs (specifically Radeon R9 M380) in Apple hardware, enabling modern driver features like Vulkan on aging systems.
🤖 ROCm Updates & Software
[2026-02-22] AMD Zen 6 Performance Events & Metrics Merged For Linux 7.0
Source: Phoronix
Key takeaway relevant to AMD:
- Next-Gen Preparation: Crucial for systems engineers and developers who need to profile and optimize software for the upcoming Zen 6 architecture immediately upon silicon availability.
- Architecture ID: Confirms Zen 6 identification within the kernel as “AMD Family 1Ah Model 50h-57h.”
Summary:
- The Linux 7.0 kernel merge window (closing today) includes the
perfsubsystem tooling changes required for Zen 6. - The update adds vendor event handling for Zen 6 core and uncore events, metrics, and mapping changes.
- The patch set also includes a retroactive fix for Zen 5 MAC allocation events.
Details:
- Version Context: These changes are part of the
perftool code for Linux 7.0. - Hardware Identification: The code explicitly supports AMD Family 1Ah Model 50h-57h Processors (Zen 6).
- Performance Counters: The merged code enables monitoring for:
- Branch prediction.
- L1 and L2 cache activity.
- TLB (Translation Lookaside Buffer) activity.
- Uncore events, specifically UMC (Unified Memory Controller) command activity.
- Additional Tooling:
- Introduction of
perf sched stats: A new tool for record/report/diff workflows using scheduler statistics. - Data type profiling fixes and improvements.
- Introduction of
- Implications: While described as “routine pre-launch enablement,” this integration ensures that the standard Linux profiling suite is ready for Zen 6 server and client deployment, allowing for immediate performance tuning without waiting for post-launch kernel patches.
[2026-02-22] Yet Another Fix Coming For Older AMD GPUs On Linux - Thanks To Valve Developer
Source: Phoronix
Key takeaway relevant to AMD:
- Driver Longevity: Demonstrates the continued viability of older GCN (Graphics Core Next) architectures on modern Linux stacks, driven by third-party (Valve) contributions rather than AMD directly.
- User Base Expansion: Enables owners of older hardware (specifically Apple iMacs with Radeon dGPUs) to utilize the modern
amdgpudriver (and thus Vulkan/RADV) rather than the deprecatedradeondriver.
Summary:
- Timur Kristóf (Valve) has developed fixes for the open-source
amdgpudriver targeting Sea Islands (GCN 1.1) GPUs, specifically the Radeon R9 M380 found in older iMacs. - The fix resolves long-standing boot issues and kernel errors that prevented these GPUs from using the modern driver stack.
Details:
- Target Hardware: Radeon R9 M380 (Sea Islands / GCN ~1.1) specifically within Intel-powered Apple iMacs.
- The Issue:
- Users reported boot failures and kernel errors (
AtomBIOSerrors,hw_initfailures,failed to send message 10a). - The issue had been open for over 1.5 years.
- Previously forced users to fallback to the legacy
radeonkernel driver, losing performance and Vulkan support.
- Users reported boot failures and kernel errors (
- Technical Root Cause:
- The
amdgpudriver did not implement voltage dependency on the display clock. - The default voltage supplied to the hardware was insufficient to drive the display controller properly.
- The issue was linked to the AMDGPU DC (Display Core) code.
- The
- The Fix:
- Disable MCLK DPM: Patches disable Memory Clock Dynamic Power Management on the problematic Sea Islands GPUs.
- Force High Clock: The driver now forces the usage of the highest memory clock when DPM is disabled to ensure voltage sufficiency.
- Status: Patches are currently in a Git branch, expected to be reviewed and upstreamed to the Linux kernel shortly.
📈 GitHub Stats
| Category | Repository | Total Stars | 1-Day | 7-Day | 30-Day |
|---|---|---|---|---|---|
| AMD Ecosystem | AMD-AGI/GEAK-agent | 65 | 0 | +2 | +9 |
| AMD Ecosystem | AMD-AGI/Primus | 74 | 0 | 0 | +8 |
| AMD Ecosystem | AMD-AGI/TraceLens | 59 | 0 | +1 | +3 |
| AMD Ecosystem | ROCm/MAD | 31 | 0 | 0 | 0 |
| AMD Ecosystem | ROCm/ROCm | 6,183 | +3 | +12 | +83 |
| Compilers | openxla/xla | 4,005 | +3 | +18 | +87 |
| Compilers | tile-ai/tilelang | 5,241 | +9 | +49 | +446 |
| Compilers | triton-lang/triton | 18,461 | +2 | +35 | +238 |
| Google / JAX | AI-Hypercomputer/JetStream | 411 | +1 | +4 | +8 |
| Google / JAX | AI-Hypercomputer/maxtext | 2,145 | +1 | +7 | +40 |
| Google / JAX | jax-ml/jax | 34,923 | +7 | +57 | +246 |
| HuggingFace | huggingface/transformers | 156,823 | +48 | +328 | +1232 |
| Inference Serving | alibaba/rtp-llm | 1,049 | 0 | 0 | +19 |
| Inference Serving | efeslab/Atom | 336 | 0 | 0 | +2 |
| Inference Serving | llm-d/llm-d | 2,518 | +2 | +26 | +125 |
| Inference Serving | sgl-project/sglang | 23,652 | +27 | +121 | +997 |
| Inference Serving | vllm-project/vllm | 70,910 | +65 | +567 | +2589 |
| Inference Serving | xdit-project/xDiT | 2,544 | 0 | +5 | +33 |
| NVIDIA | NVIDIA/Megatron-LM | 15,244 | +8 | +34 | +248 |
| NVIDIA | NVIDIA/TransformerEngine | 3,169 | 0 | +6 | +64 |
| NVIDIA | NVIDIA/apex | 8,926 | 0 | +8 | +27 |
| Optimization | deepseek-ai/DeepEP | 8,993 | +1 | +9 | +76 |
| Optimization | deepspeedai/DeepSpeed | 41,647 | +4 | +23 | +277 |
| Optimization | facebookresearch/xformers | 10,346 | 0 | +8 | +55 |
| PyTorch & Meta | meta-pytorch/monarch | 976 | +1 | +9 | +23 |
| PyTorch & Meta | meta-pytorch/torchcomms | 337 | 0 | +5 | +15 |
| PyTorch & Meta | meta-pytorch/torchforge | 621 | 0 | 0 | +21 |
| PyTorch & Meta | pytorch/FBGEMM | 1,534 | -1 | 0 | +15 |
| PyTorch & Meta | pytorch/ao | 2,695 | +1 | +7 | +53 |
| PyTorch & Meta | pytorch/audio | 2,831 | 0 | +3 | +17 |
| PyTorch & Meta | pytorch/pytorch | 97,680 | +36 | +256 | +825 |
| PyTorch & Meta | pytorch/torchtitan | 5,082 | -1 | +11 | +88 |
| PyTorch & Meta | pytorch/vision | 17,523 | -1 | +13 | +55 |
| RL & Post-Training | THUDM/slime | 4,296 | +16 | +123 | +803 |
| RL & Post-Training | radixark/miles | 896 | +4 | +17 | +131 |
| RL & Post-Training | volcengine/verl | 19,310 | +16 | +85 | +672 |