🖥️ AI & GPU Industry Weekly Recap: March 2–8, 2026

🔑 Key Highlights

  • Nvidia’s driver saga continues: The GeForce 595.59 driver was recalled for a catastrophic fan-control bug, replaced by 595.71 — which then introduced voltage-limiting issues on RTX 40- and 50-series GPUs, sparking renewed criticism of AI-assisted driver development quality.
  • Nvidia invests $2B each in Lumentum and Coherent, signaling a major strategic pivot toward co-packaged optics (CPO) and optical circuit switching for next-generation AI cluster interconnects, with multibillion-dollar purchase commitments attached.
  • Broadcom’s Q1 FY2026 revenues hit $19.31B (up 29.5% YoY), reinforcing its emergence as a serious counterbalance to Nvidia in AI compute through custom XPU ASICs and an expanding rackscale systems business.
  • Apple unveiled the M5 Pro and M5 Max with a new “Fusion Architecture” dual-die design, featuring an 18-core CPU and up to a 40-core GPU with 4x AI compute improvements over the prior generation.
  • AMD’s Linux ecosystem advances broadly — from EPYC Turin benchmark deep-dives and SEV-SNP security improvements to DCN 4.2 / GFX 12.1 AMDGPU driver staging for Linux 7.1, hinting at next-gen RDNA hardware on the horizon.

🤖 AI & Machine Learning

AMD VP Uses AI to Build Python GPU Driver

AMD’s VP of AI Software, Anush Elangovan, used Anthropic’s Claude Code (claude-opus-4-6) to build a pure-Python AMD GPU user-space driver from scratch — reportedly without opening a code editor once. The driver interfaces directly with /dev/kfd and /dev/dri/renderD* via ctypes ioctls, bypassing the ROCm/HIP stack entirely. It already supports multi-GPU configurations, SDMA copy engines, PM4 compute packets, and passed 130 tests on an MI300X (gfx942). The project was inspired by Tinygrad’s user-space GPU driver approach and is positioned for ROCm stress-testing and debugging.

AI-Assisted Kernel Driver Raises LKML Controversy

A developer submitted an AMD DPTCi (Dynamic Power and Thermal Configuration Interface) Linux kernel driver aimed at improving power/thermal tuning for Ryzen-powered gaming handhelds (GPD, AYANEO, OneXPlayer, AOKZOE, OrangePi). However, the driver was partially written by AI without disclosure, triggering pushback on the Linux Kernel Mailing List (LKML). A rewrite is expected. The episode illustrates the growing friction around AI-generated code in safety-critical open-source infrastructure.

Intel OpenVINO & Rendering Toolkit on Panther Lake Xe3

Benchmarks published this week showed Intel’s Arc B390 (Xe3) GPU on the Core Ultra X7 358H Panther Lake delivering meaningful AI inference performance gains via OpenVINO and OpenVINO GenAI versus Lunar Lake (Xe5) and Meteor Lake (Xe) predecessors. The Intel Rendering Toolkit (OSPRay Studio, OpenVKL, Open Image Denoise) also showed generation-over-generation improvements in performance-per-watt.

Apple M5 Pro/Max: 4x AI Compute Leap

Apple’s new M5 Pro and M5 Max, built on a “Fusion Architecture” with two dies combined into a single SoC, feature an 18-core CPU with “super cores” and up to a 40-core GPU delivering over 4x peak GPU compute for AI compared to the M4 generation. Linux support via Asahi Linux remains distant, with M3/M4 still being upstreamed.

⚡ GPU & Hardware

Nvidia Driver Chaos: 595.59 → 595.71 → More Problems

The week was defined by Nvidia driver turbulence:

  • 595.59 was recalled after a bug caused GPU fans to stop spinning on RTX 30, 40, and 50-series cards — a potentially hardware-damaging failure.
  • Replacement 595.71 introduced new problems: artificial voltage limits on RTX 40- and 50-series GPUs, costing overclockers up to 200MHz of headroom (e.g., an Asus TUF RTX 5090 lost 65mV and ~171MHz of boost). RTX 5080 owners reported drops from 3,100–3,200MHz to ~2,955MHz.
  • Not all SKUs were affected — some Gigabyte Aorus Master RTX 5090 and PNY Epic OC users reported no issues, suggesting the bug may be silicon-specific or board-partner dependent.
  • 595.71 also added game-ready support for Marathon (DLSS Super Resolution + Nvidia Reflex) and optimizations for Resident Evil Requiem.
  • On Linux, NVIDIA 595.45.04 beta launched with long-awaited VK_EXT_descriptor_heap, VK_EXT_present_timing Vulkan extensions, DRI3 v1.2 support, modesetting enabled by default for nvidia-drm, and improved GPU reset handling.

RTX 5070 Mobile May Get 12GB GDDR7 Upgrade

Multiple OEM listings (Lenovo Yoga Pro 7i, Asus ROG Zephyrus G14 2026, HP OMEN, Asus ROG Strix G16/G18) suggest an unannounced GeForce RTX 5070 Mobile 12GB GDDR7 SKU, matching the desktop version’s memory capacity. The current mobile variant ships with only 8GB, a 33% reduction. Amid a global GDDR7 memory shortage, the upgrade — if real — would be significant for mobile gamers, though some listings have already been updated or removed, and Nvidia hasn’t officially confirmed the SKU.

Nvidia Goes Optical: $2B Each into Lumentum & Coherent

Nvidia is making $2 billion strategic investments in both Lumentum and Coherent, accompanied by multibillion-dollar purchase commitments for laser components and optical networking products. This move is driven by:

  • The need for co-packaged optics (CPO) on next-gen switches (Quantum-X InfiniBand, Spectrum-X Ethernet) and eventually on GPUs and NVSwitch itself.
  • Exploration of optical circuit switches (OCS) — Lumentum’s 300×300-port R300 (MEMS-based) and Coherent’s DLX (liquid crystal, up to 512×512 ports) — as spine-layer interconnects for massive AI clusters.
  • Potential cluster power savings of up to 65% and 5–10x lower latency vs. electrical Ethernet switching at 100,000+ XPU scale.
  • Speculation that Nvidia’s upcoming “Rubin Ultra” / “Kyber” rack generation may shift to a torus or dragonfly interconnect topology using OCS spines.

AMD EPYC Turin: Zen 5 vs. Zen 5C at 128 Cores

Phoronix published a deep-dive comparison of the EPYC 9755 (128 Zen 5 cores, 500W TDP, 512MB L3, $7,200) vs. EPYC 9745 (128 Zen 5C cores, 400W TDP, 256MB L3, ~$7,200) using a Gigabyte MZ33-AR1 server. The 9745 supports cTDP down to 320W for power-constrained platforms. With similar pricing, the 9745 is primarily differentiated by platform TDP compatibility and energy efficiency rather than price, with the reduced L3 cache and lower boost clocks (3.45GHz all-core vs. 4.1GHz) being the key trade-offs.

AMD DCN 4.2 and GFX 12.1 Staging for Linux 7.1

AMD began staging AMDGPU driver updates for the Linux 7.1 kernel cycle, including:

  • DCN 4.2 (Display Core Next) IP enablement for upcoming AMD hardware.
  • Further GFX 12.1 (updated RDNA4 graphics engine) updates with 57-bit scratch memory addressing via 5-level page tables.
  • GCN 1.1 / Sea Islands APU display engine (DC) enabled by default, contributed by Valve’s Linux graphics driver team.
  • UserQ updates, SDMA improvements, and ring reset enhancements.

🏭 Industry & Market

Broadcom: A $19B Quarter and the Nvidia Challenger Thesis

Broadcom posted $19.31 billion in Q1 FY2026 revenue, up 29.5% YoY, with operating income up 36.8% to $8.56 billion. The analysis from The Next Platform argues that Broadcom’s combination of custom AI XPU ASICs (for hyperscalers building their own silicon), networking chips, and an emerging rackscale systems business positions it as the most credible long-term counterbalance to Nvidia — potentially outpacing AMD’s datacenter ambitions. CEO Hock Tan’s software acquisitions (CA Technologies, VMware) continue to fund the AI chip build-out, and the company appears to be moving toward end-to-end rack integration, threatening traditional OEMs and ODMs.

RTX 5070 Tops Steam Hardware Survey — With Caveats

The February 2026 Steam Hardware & Software Survey showed the GeForce RTX 5070 jumping to 9.42% market share (a 6.55 percentage-point MoM leap), displacing the RTX 4060 as the top GPU. However, analysts flag significant data quality concerns: Chinese users represented 54.6% of Steam’s surveyed population (+30.74% MoM) during the Chinese New Year period, with high-end gaming cafes likely causing survey duplication. Separately, Windows 11 reportedly fell 10.43% among Steam users while Windows 10 gained 12.46%, despite Win 10 reaching end-of-support in October 2025.

AMD Mobile RDNA 4 Still MIA

Despite RDNA 4 desktop cards having been on the market for roughly a year, AMD has yet to launch any mobile RDNA 4 GPUs. The Radeon RX 9000M series (RX 9080M, 9070M XT, 9070M, 9060M) remains on the horizon with no firm launch date, leaving laptop buyers with only RDNA 3 options. This vacuum is likely influencing Nvidia’s RTX 5070 Mobile 12GB refresh consideration.

🛠️ Developer Ecosystem

Linux 7.0 Delivers Database Performance Gains on AMD EPYC

Linux 7.0 benchmarks on both large (EPYC 9755, 128-core Turin) and small (EPYC 4585PX, 16-core) AMD servers show consistent, statistically significant gains for database workloads: PostgreSQL, MariaDB, Redis, and CockroachDB all benefited. No regressions were observed. This is particularly noteworthy as Linux 7.0 is slated to be the default kernel in Ubuntu 26.04 LTS.

Linux 7.0: IBPB-On-Entry Security for AMD SEV-SNP VMs

A patch targeting the Linux 7.0 “x86/urgent” branch (marked for stable backport) enables IBPB-On-Entry for AMD SEV-SNP guest VMs on Zen 5 EPYC processors. This hardware security feature — Indirect Branch Predictor Barrier on VM entry — helps prevent speculative execution attacks. It was an oversight that the feature wasn’t activated since Zen 5 launch.

GNOME Mutter 50.rc: Major NVIDIA Linux Performance Boost

GNOME Mutter 50.rc shipped with significant improvements for NVIDIA Linux users via the R590 driver: frame block time reduced from milliseconds to microseconds. Additional highlights include SDR-Native color mode, Wayland color management v2, HDR screen sharing, VRR + commit timing support, and 10-bit-per-channel scanout via FBOs on NVIDIA. GNOME 50 stable is targeted for two weeks out.

NVIDIA R595 Linux Beta: New Vulkan Extensions

The NVIDIA 595.45.04 Linux beta adds VK_EXT_descriptor_heap and VK_EXT_present_timing Vulkan extensions — long-requested capabilities for the Linux gaming ecosystem. Modesetting is now enabled by default in the nvidia-drm kernel module. The minimum Wayland version is bumped to v1.20.

Framework 16 / 13 Getting Coreboot + AMD openSIL

9elements is porting Coreboot + AMD openSIL (open Silicon Initialization Library) to the first-generation Framework 16 (Ryzen 7000 / Phoenix SoC), with the Framework 13 Gen1 to follow. The effort is still in early stages (memory training, early init). Complementary work by 3mdeb targets a Gigabyte EPYC 9005 server board and an MSI AM5 Ryzen board. AMD’s openSIL is expected to reach production status with the upcoming Zen 6 platform.

📊 Key Takeaways

Nvidia dominated headlines for the wrong reasons this week — back-to-back driver releases with serious bugs (fan failures, voltage suppression) across RTX 30/40/50 cards have amplified community concern about software quality, with some critics pointing fingers at AI-assisted code development. Meanwhile, Nvidia’s $4 billion combined investment in Lumentum and Coherent represents a far more strategic story: the company is clearly laying the groundwork for optically interconnected AI supercomputers at scales that current electrical switching cannot economically support, potentially reshaping cluster architecture as early as the “Rubin Ultra” generation. On the competitive front, Broadcom’s accelerating custom XPU momentum and AMD’s continued EPYC/RDNA ecosystem maturation — particularly the strong Linux 7.0 database gains and next-gen GFX 12.1/DCN 4.2 hardware hints — signal that the AI hardware landscape is diversifying rapidly, even as Nvidia retains its commanding lead in deployed GPU compute.