Skip to content

🧠 APU & Unified Memory

For sovereign enterprise AI deployment, unified memory (a single memory shared by every component of the chip) is one of the most important hardware shifts of the decade.

On a conventional PC, the model is first loaded into RAM (the processor’s memory), then copied into VRAM (the graphics card’s memory) over the PCIe bus, the “pipe” that connects the two. Unified-memory machines eliminate this copy. Their chip is a SoC (system on a chip: nearly a whole computer on a single chip), which AMD calls an APU. The central processor (CPU), the graphics processor (GPU) that computes the AI and an AI accelerator (NPU) share the same memory, of the LPDDR5X type (low-power memory soldered next to the chip).

This memory goes up to 192 GB on AMD Ryzen AI Max PRO 400 platforms, and up to 512 GB on the Mac Studio M5 Ultra 1234. The latter reaches 1.2 TB/s of memory bandwidth, the data rate between the memory and the chip 34. Unified memory is the reference architecture for running a model with 70 billion parameters (“70B”) or more on a quiet desktop workstation, without a datacenter GPU server 53. This requires quantizing it, that is, compressing it by reducing the precision of its numbers: Llama 3.1 70B in Q4_K_M thus weighs ~40 GB of weights 53.


⚔️ The Hardware Landscape: Apple Silicon, AMD Gorgon Halo, and NVIDIA DGX Spark

Three ecosystems share the high-performance unified memory market as of Q4 2026:

1. Apple Silicon (Mac Studio M5, August 2026)

Apple integrates CPU, GPU, and NPU on a single package and houses the unified memory inside the chip’s package, a design introduced with the M1 chip 6. The Mac Studio M5 was announced on 2026-08-25 4.

  • Mac Studio M5 Max (18c CPU / 40c GPU): up to 128 GB unified, 614 GB/s bandwidth (460 GB/s with the 32-core GPU), from US$2,499 34.
  • Mac Studio M5 Ultra (up to 36c CPU / 80c GPU): 96, 256, or 512 GB unified, 1.2 TB/s, from US$5,499; the 512 GB option, reserved for the M5 Ultra with 36c CPU / 80c GPU, has been orderable since 28 October 2026 from US$22,299 ($6,799 + $15,500) and €25,079 incl. VAT in France, with shipping lead times reported at 13 to 15 weeks in early November 2026 347.
  • Strengths: very high bandwidth, mature MLX (Apple’s AI framework for its chips) and Metal (its GPU compute interface) ecosystems, silence and low power draw 54.
  • Limits: macOS (operating system imposed), non-upgradable soldered memory, high price at 128 GB+ configurations 3.

2. AMD Ryzen AI Max PRO 400 (“Gorgon Halo”)

Professional refresh of the Strix Halo platform (Zen 5 + RDNA 3.5); as of Q4 2026 the flagship Max+ PRO 495 is listed by AMD and shipping in the 192 GB Framework Desktop (deliveries from November 2026) 128.

  • Architecture: up to 16 Zen 5 cores, Radeon 8065S integrated GPU (iGPU) with 40 compute units (CUs) on the flagship model Max+ PRO 495, 256-bit LPDDR5X-8533 memory bus 12.
  • Capacity: up to 192 GB unified (+50% vs the 300 series at 128 GB) 12.
  • Bandwidth: ~273 GB/s (vs ~256 GB/s theoretical on Strix Halo 128 GB — ~7% gain from memory clock) 12.
  • GPU allocation: up to 160 GB reservable as iGPU VRAM, 32 GB left for the system 12.
  • Strengths: open x86 PC architecture (native Linux / Windows / Docker), capacity/price ratio still attractive but deteriorating fast: the 128 GB Ryzen AI Halo development platform sold for US$3,999 (July 2026) and the 128 GB Framework Desktop is listed at €3,889 incl. VAT and out of stock as of 2026-10-09 — the LPDDR5X price surge is hitting every capacity tier 8910.
  • Limits: bandwidth roughly 2.2× lower than the M5 Max (614 GB/s) and ~4.4× lower than the M5 Ultra (1.2 TB/s) — decoding dense 70B models stays memory-bound, that is, limited by memory (~5 tok/s measured on Strix Halo 395, similar profile expected on PRO 400) 2311.

3. NVIDIA Grace Blackwell (DGX Spark)

The DGX Spark is NVIDIA’s first product combining an ARM SoC (a processor architecture that comes from smartphones) and a Blackwell GPU in a desktop form factor 12.

  • Architecture: Grace Blackwell SoC (GB10) — 20-core ARM CPU (Grace) + Blackwell GPU, co-developed with MediaTek, manufactured in 3 nm by TSMC (the Taiwanese foundry) 12.
  • Unified LPDDR5x memory: 128 GB on DGX Spark (≈ $6,950 as of 2026-10-02); a 64 GB variant (same GB10, models up to ~100B parameters according to NVIDIA, ~200B with two units linked through their QSFP ports: a high-speed network link) from $4,999, sold exclusively by OEMs (Acer, ASUS, Dell, Gigabyte, HP, MSI) starting 2026-10-23 121314. DGX Station scales to 748 GB for models > 400B 12.
  • Bandwidth: ~273 GB/s (LPDDR5x) — close to AMD Gorgon Halo 12.
  • Power draw: 240 W power supply, GB10 TDP (power to dissipate) ~140 W (DGX Spark) — versus > 1,000 W for a dual discrete-GPU workstation 12.
  • Native FP4 (Blackwell): numbers encoded on 4 bits, a quantization format. Unlike the Ada Lovelace architecture (RTX 4090 — emulated FP4), Blackwell implements FP4 in silicon — without software overhead 15.
  • ConnectX-7 200 Gbps scale-out: ConnectX-7 is NVIDIA’s built-in network card; two QSFP ports link up to four DGX Spark units in a ring, without an external switch (512 GB aggregated, models up to ~700B according to NVIDIA). Since 22 October 2026, NVIDIA also officially supports clusters of eight nodes connected through a switch (1 TB aggregated, enough to load the full GLM 5.3, 753B) 1216.
  • NVIDIA Sync software: full CUDA (the software platform of NVIDIA GPUs) environment pre-installed to reduce startup friction.

NVIDIA and Microsoft are marketing the RTX Spark, distinct from the DGX Spark: Blackwell RTX GPU with up to 6,144 cores, Grace CPU with up to 20 cores, up to 128 GB of unified memory and a theoretical 1 PFLOPS FP4, running Windows with full CUDA, from Acer, ASUS, Dell, Gigabyte, HP, Lenovo, Microsoft and MSI (laptops first; compact desktops available since 2026-11-18 from ASUS, Dell and MSI, with those from Acer, Gigabyte and HP announced as “coming soon”) 1718.

On the desktop side, the Surface RTX Spark Dev Box (128 GB) costs $5,999 (United States only, deliveries in November 2026); Microsoft specifies that the 1 PFLOPS FP4 assumes sparsity and that the GPU addresses only part of the 128 GB, and no memory bandwidth is published as of 2026-10-10 19.


🏎️ Why Unified Memory Removes the “PCIe Bottleneck”

The RAM → PCIe → VRAM copy on a PC with a discrete GPU, and its elimination with unified memory, are detailed, with diagrams, in ⚔️ Unified Memory vs RAM vs VRAM; the trade-off: a shared bandwidth lower than dedicated GDDR7 VRAM, but a much larger capacity per machine 20.


🛠️ Practical Guide: Configuring GPU Memory Allocation

Operating systems (OS) cap the share of unified RAM usable by the GPU by default — you must raise this ceiling for heavy LLMs.

1. AMD Side (Ryzen AI Max 300 / PRO 400)

On Strix Halo and Gorgon Halo, allocation is mainly via OEM BIOS/firmware 12:

  1. Reboot → BIOS/UEFI (the boot firmware).
  2. Advanced menu → UMA Frame Buffer Size, i.e. the share of unified memory (UMA) reserved for the GPU (label varies by OEM).
  3. On a 192 GB system, AMD documents up to 160 GB for the GPU and 32 GB for the OS 12.

The first PRO 400 systems have been on pre-order since Q4 2026 (192 GB Framework Desktop; HP, Lenovo and ASUS announced); Strix Halo 395 guides remain the practical reference for llama.cpp/Vulkan/ROCm (AMD’s GPU software stack, the equivalent of CUDA), and SGLang has published a ROCm image for gfx1151 (Strix Halo) since v0.5.20 11824.

2. Apple Silicon Side (macOS)

By default, macOS caps the GPU Metal working set at about 75% of unified RAM (via recommendedMaxWorkingSetSize) — on 128 GB, only ~96 GB is usable without tweaks 2526.

To raise the limit (e.g. ~120 GB on a 128 GB machine), the community-documented llama.cpp/MLX command is:

Fenêtre de terminal
# Value in MEGABYTES (e.g. 120 GB → 120 × 1024 = 122880)
sudo sysctl iogpu.wired_limit_mb=122880
  • Verify: sysctl iogpu.wired_limit_mb (0 = default policy).
  • Revert to default: sudo sysctl iogpu.wired_limit_mb=0 2526.
  • Leave 8–16 GB for the system to avoid memory pressure / swap 2526.
  • The change is volatile (lost on reboot); for persistence, use a LaunchDaemon or equivalent — not officially supported by Apple 2526.

📊 Economic and Performance Trade-offs

Comparison of unified workstations for Llama 3.1 70B Q4_K_M (~40 GB of weights). Speeds = decode (generation), order-of-magnitude measured or published — vary by backend (MLX vs llama.cpp), context, and build.

CriterionMac Studio (M5 Max 128 GB)Mac Studio (M5 Ultra 256–512 GB)AMD Ryzen AI (Halo / PRO 400)NVIDIA DGX Spark (64 / 128 GB)
Chip (LLM config)M5 Max 18c/40c GPU 3M5 Ultra 36c/80c GPU 3Ryzen AI Max+ PRO 495 12Grace Blackwell SoC (GB10) 12
Max unified RAM128 GB 3512 GB (96 / 256 / 512 GB tiers) 34192 GB 1264 or 128 GB LPDDR5x 1213
Max allocatable GPU VRAM~120 GB (sysctl, 128 GB machine) 25~75% by default; up to RAM − 8–16 GB via sysctl (~240 GB on 256 GB, ~496 GB on 512 GB) 2526160 GB (BIOS max) — in practice ~130–140 GB 12~64 / ~128 GB (direct CUDA access) 12
Bandwidth614 GB/s (460 GB/s with the 32-core GPU) 341.2 TB/s 34~273 GB/s 12~273 GB/s 12
70B Q4 throughput (decode)≤ ~15 tok/s (ceiling 614 GB/s ÷ 40 GB; M5 Max MLX measurements not published) 3≤ ~30 tok/s (ceiling 1.2 TB/s ÷ 40 GB; M5 Ultra MLX measurements not published) 3~4.5–5 tok/s (Strix Halo 395) 11not officially published 12
Native FP4❌❌❌✅ Blackwell 15
Scale-out✅ Thunderbolt 5 RDMA (macOS 26.2+, up to 4 Mac Studios) 4✅ Thunderbolt 5 RDMA (macOS 26.2+, up to 4 Mac Studios) 4❌✅ ConnectX-7 200 Gbps, up to 4 nodes in a ring, 8 with a switch (since 2026-10-22) 1216
OSmacOSmacOSLinux / Windows 12Linux (CUDA) 12
Indicative pricefrom US$2,499 (36 GB base); 128 GB config: custom-configuration (CTO) price to be checked 4from US$5,499 (96 GB); 512 GB: from US$22,299 (M5 Ultra 36c/80c required); 256 GB: custom-configuration (CTO) price to be checked 47~€3,900 (128 GB, out of stock as of 2026-10) · ~$6,800–7,450 (192 GB, pre-order, shipping Nov. 2026) 8927≈ $6,950 (128 GB, as of 2026-10-02) · $4,999 (64 GB, OEM, from 2026-10-23) 1213

📋 Architect’s Takeaway

For sovereign on-premise deployment:

  1. x86 sovereignty + Docker (AMD Halo / PRO 400): if your stack relies on Linux, Docker, and Python, the AMD platform is the most rational: 160 GB allocatable VRAM, open ecosystem, lower price than an equivalent-capacity Mac Studio 1228. Accept ~5 tok/s on a dense 70B — favor MoE models (Qwen3.5-A3B, etc.) for interactivity 11.
  2. Speed and comfort (Mac Studio): for the best feel on a dense 70B (memory-bound ceiling ~30 tok/s on M5 Ultra, measurements not published), the Mac Studio M5 Ultra 256–512 GB (1.2 TB/s) remains the unified-bandwidth reference as of Q4 2026; the M5 Max 128 GB (614 GB/s) is an excellent compromise if 128 GB is enough 34. Budget for macOS and containers.
  3. CUDA + FP4 + scale-out (NVIDIA DGX Spark): if your team is already in the NVIDIA ecosystem (CUDA, TensorRT, vLLM), DGX Spark offers 128 GB LPDDR5x natively accessible via CUDA, native Blackwell FP4, and the ability to interconnect up to four enclosures in a ring via ConnectX-7 200 Gbps without a switch, and up to eight with a switch since October 2026 1216. It is not the best choice for pure inference on models ≤ 34B: a discrete-GPU workstation (RTX 5090, RTX PRO 6000) remains faster for that use case — at a price now comparable or higher as of Q4 2026 (see the “DGX Spark positioning” callout).
  4. Size at purchase time: LPDDR5X is soldered — no post-purchase upgrade. Plan margin for weights + KV cache + OS + load page cache (see foundation chapters). On a 192 GB AMD system, allocating 160 GB to the GPU leaves 32 GB for the system — sufficient at idle, but tight on first load of a model ≥ 100 GB.

📚 Sources and References

Footnotes

  1. AMD, AMD Ryzen AI Max+ PRO 495 — product sheet (16 Zen 5 cores, Radeon 8065S 40 CU, 192 GB LPDDR5X-8533, 256-bit bus, cTDP 45–120 W), accessed 2026-10-09 · AMD, AAI 2026: AMD Delivers Full-Stack Compute for the Agentic AI Era (Halo PRO 400 platforms “later this year”), 2026-07-23. https://www.amd.com/en/products/processors/laptop/ryzen-pro/ai-max-pro-400-series/amd-ryzen-ai-max-plus-pro-495.html · https://ir.amd.com/news-events/press-releases/detail/1294/aai-2026-amd-delivers-full-stack-compute-for-the-agentic-ai-era ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7 ↩8 ↩9 ↩10 ↩11 ↩12 ↩13 ↩14

  2. ServeTheHome, AMD Ups Ante With 192GB Ryzen AI Max PRO 400 Chips for AI Systems, May 2026. https://www.servethehome.com/amd-reveals-ryzen-ai-max-pro-400-series-192gb-ram-for-ai-systems/ ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7 ↩8 ↩9 ↩10 ↩11 ↩12 ↩13 ↩14 ↩15

  3. Apple, Mac Studio — Technical Specifications (M5 Max 614 GB/s / M5 Ultra 1.2 TB/s, 96 / 256 / 512 GB RAM tiers), re-read on 2026-10-09. https://www.apple.com/mac-studio/specs/ ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7 ↩8 ↩9 ↩10 ↩11 ↩12 ↩13 ↩14 ↩15 ↩16 ↩17

  4. Apple Newsroom, Apple introduces new Mac Studio with M5 Max and M5 Ultra (M5 Max 614 GB/s, M5 Ultra 1.2 TB/s, 96 / 256 / 512 GB, US starting prices, 512 GB option announced for late October 2026, Thunderbolt 5 + RDMA clustering under macOS 26.2 up to 4 Mac Studios), August 25, 2026. https://www.apple.com/newsroom/2026/08/apple-introduces-new-mac-studio-with-m5-max-and-m5-ultra/ ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7 ↩8 ↩9 ↩10 ↩11 ↩12 ↩13 ↩14

  5. Apple Newsroom, Apple introduces M4 Pro and M4 Max, October 2024 (546 GB/s, 128 GB max M4 Max). https://www.apple.com/newsroom/2024/10/apple-introduces-m4-pro-and-m4-max/ ↩ ↩2 ↩3

  6. Apple Newsroom, Apple unleashes M1 (“a unified memory architecture that brings together high-bandwidth, low-latency memory into a single pool within a custom package”, with no copying between multiple memory pools), 10 November 2020. https://www.apple.com/newsroom/2020/11/apple-unleashes-m1/ ↩ ↩2

  7. MacRumors, Mac Studio With M5 Ultra Now Available With 512GB RAM for $22,299 (512 GB option reserved for the M5 Ultra with 36c CPU / 80c GPU at $6,799, $15,500 surcharge), 28 October 2026 · MacGeneration, Mac Studio M5 Ultra : l’option 512 Go de RAM coûte plus cher qu’un second Mac Studio (€12,429 for 256 GB + €12,650 for 512 GB, i.e. €25,079 incl. VAT; lead times of 13 to 15 weeks), 4 November 2026 · Apple Support, Mac Studio (M5 Max or M5 Ultra) — Tech Specs (“256GB or 512GB (M5 Ultra with 36‑core CPU and 80‑core GPU)”). https://www.macrumors.com/2026/10/28/mac-studio-m5-ultra-512gb-ram/ · https://www.macg.co/mac/2026/11/mac-studio-m5-ultra-loption-512-go-de-ram-coute-plus-cher-quun-second-mac-studio-312149 · https://support.apple.com/en-us/128107 ↩ ↩2

  8. Framework, The 192GB Framework Desktop is open for pre-order (Ryzen AI Max+ PRO 495, 192 GB, pre-orders from 2026-09-30, deliveries November 2026, warning about rising memory prices), September 30, 2026. https://frame.work/blog/192gb-framework-desktop-open-for-pre-order ↩ ↩2 ↩3 ↩4 ↩5

  9. Framework, Framework Desktop (AMD Ryzen AI Max 300) — product page (128 GB: €3,889 incl. VAT, “Out of stock”), accessed 2026-10-09. https://frame.work/products/desktop-diy-amd-aimax300 ↩ ↩2

  10. Tom’s Hardware, AMD challenges Nvidia’s DGX Spark with $3,999 Ryzen AI Halo (Ryzen AI Max+ 395 development platform, 128 GB, Micro Center, available July 10, 2026), July 2026. https://www.tomshardware.com/desktops/mini-pcs/amd-challenges-nvidias-dgx-spark-with-usd3-999-ryzen-ai-halo-with-windows-11-support-strix-halo-desktop-undercuts-nvidia-by-usd700-packs-128gb-of-unified-memory ↩

  11. ignasivt, Strix Halo Guide (llama.cpp benchmarks Llama 3.1 70B Q4 ~4.7–4.9 tok/s), 2026. https://github.com/ignasivt/strix-halo-guide ↩ ↩2 ↩3 ↩4

  12. NVIDIA, DGX Spark — official product page (Grace Blackwell SoC GB10, 128 GB LPDDR5x, ~273 GB/s, ConnectX-7 200 Gbps and cluster of up to 4 systems, 64 GB variant “Coming Soon” from OEMs, “RTX Spark” pre-order link), re-read on 2026-10-09. https://www.nvidia.com/en-us/products/workstations/dgx-spark/ ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7 ↩8 ↩9 ↩10 ↩11 ↩12 ↩13 ↩14 ↩15 ↩16 ↩17

  13. ServeTheHome, NVIDIA DGX Spark 64GB Launched and Big 128GB GB10 Price Increases (128 GB ≈ $6,950, 64 GB $4,999 OEM from 2026-10-23), October 3, 2026. https://www.servethehome.com/nvidia-dgx-spark-64gb-launched-and-big-128gb-gb10-price-increases/ ↩ ↩2 ↩3

  14. NVIDIA Blog, NVIDIA DGX Spark 64GB Gives Developers More Ways to Build and Scale Local AI (same GB10, from $4,999, models up to 100B parameters, OEMs Acer, ASUS, Dell, Gigabyte, HP, MSI from Friday 23 October; two units linked over QSFP = 128 GB, models up to 200B), 2 October 2026. https://blogs.nvidia.com/blog/local-ai-dgx-spark-64gb-sync/ ↩

  15. NVIDIA, NVIDIA Blackwell Architecture Technical Brief (native FP4 Tensor Cores vs emulated FP4 Ada Lovelace). https://resources.nvidia.com/en-us-blackwell-architecture ↩ ↩2

  16. NVIDIA Technical Blog, Running Frontier-Class AI Models Across 8 NVIDIA DGX Spark Nodes (official support for 8-node clusters, 1 TB; switchless ring recommended up to 4 nodes, switch required at 8; 8-node support in NVIDIA Sync planned for late 2026), 22 October 2026. https://developer.nvidia.com/blog/running-frontier-class-ai-models-across-8-nvidia-dgx-spark-nodes/ ↩ ↩2 ↩3

  17. NVIDIA Blog, NVIDIA, Microsoft Kick Off a New Beginning for Windows PCs With RTX Spark and AI Agents (Blackwell RTX GPU with up to 6,144 cores, Grace CPU with up to 20 cores, up to 128 GB of unified memory, 1 PFLOPS FP4; Acer, ASUS, Dell, HP, Lenovo, Microsoft, MSI, Gigabyte; laptops available for pre-order on 7 October and available on 16 October, compact form factors in November), 7 October 2026. https://blogs.nvidia.com/blog/local-ai-rtx-spark-microsoft-windows-event/ ↩

  18. NVIDIA Blog, NVIDIA RTX Spark Compact Desktops Out Now (compact desktops from ASUS, Dell and MSI available; mini-PCs from Acer, GIGABYTE and HP and the Surface RTX Spark Dev Box “coming soon”; 20-core Grace CPU, Blackwell RTX GPU with 6,144 CUDA cores, up to 128 GB of unified memory), 18 November 2026. https://blogs.nvidia.com/blog/rtx-spark-desktop-local-ai-available-now/ ↩

  19. Microsoft Devices Blog, Pre-order our most powerful Surface devices ever (Surface RTX Spark Dev Box: 128 GB of unified memory, $5,999 suggested price, pre-order on Microsoft.com in the United States only, deliveries from November; theoretical 1 PFLOPS FP4 with sparsity; GPU-addressable memory lower than the total), 7 October 2026. https://blogs.windows.com/devices/2026/10/07/pre-order-our-most-powerful-surface-devices-ever/ ↩

  20. NVIDIA Technical Blog, Mastering LLM Techniques: Inference Optimization (memory bottlenecks, quantization), November 2023. https://developer.nvidia.com/blog/mastering-llm-techniques-inference-optimization/ ↩

  21. Framework, Introducing the Framework Desktop (LPDDR5x soldered to reach the Ryzen AI Max’s memory bandwidth; modular memory judged not technically feasible at high throughput on the 256-bit bus), 25 February 2025. https://frame.work/blog/introducing-the-framework-desktop ↩

  22. NVIDIA, RTX PRO 6000 Blackwell Workstation Edition — product page (96 GB GDDR7 ECC, 1,792 GB/s), accessed 2026-10-12. https://www.nvidia.com/en-us/products/workstations/professional-desktop-gpus/rtx-pro-6000/ ↩

  23. NVIDIA, H200 Tensor Core GPU — product page (141 GB HBM3e, 4.8 TB/s), accessed 2026-10-12. https://www.nvidia.com/en-us/data-center/h200/ ↩

  24. SGLang Project, Release v0.5.20 (ROCm image for gfx1151 Strix Halo; CUDA 13 required on the NVIDIA side), September 18, 2026. https://github.com/sgl-project/sglang/releases/tag/v0.5.20 ↩

  25. ggml-org/llama.cpp, Issue #16646 (iogpu.wired_limit_mb), 2025. https://github.com/ggml-org/llama.cpp/issues/16646 ↩ ↩2 ↩3 ↩4 ↩5 ↩6

  26. ivanopcode, Override macOS Metal VRAM cap (iogpu.wired_limit_mb, tables by RAM), 2025. https://github.com/ivanopcode/devnote-override-macos-metal-vram-cap ↩ ↩2 ↩3 ↩4 ↩5

  27. Phoronix, Framework Desktop With AMD Ryzen AI Max+ PRO 495 “Gorgon Halo” Now Available For Pre-Order (192 GB: $6,799 DIY / $7,449 assembled), September 30, 2026. https://www.phoronix.com/news/Framework-Desktop-Gorgon-Halo ↩

  28. TweakTown, AMD launches Ryzen AI Max PRO 400 — up to 192GB unified memory, May 2026. https://www.tweaktown.com/news/111752/amd-launches-the-ryzen-ai-max-pro-400-series-of-cpus-up-to-16-cores-with-192gb-of-unified-memory/index.html ↩

Give my opinion

Everything is optional: one answer is enough. This site sends nothing: you send it, the way you choose.

Your noteNo note
Your opinion is about
About what? (several choices possible)
What will be sent
Send by e-mail

Private, no account: opens your mail program with the message ready, for ia-on-prem@damien.becherini.fr.

Open a GitHub issue ↗

Public, with a GitHub account. You read the issue again before you publish it.

To paste it wherever you want.

Give at least a note, a subject or a message to be able to send.