Local AI on a Ryzen iGPU (Radeon 780M, 880M, 890M)

Which chips

CU counts and gfx IDs come from AMD product pages (Graphics Core Count) and the ROCm 10.1.0 GPU specifications table. We do not invent a shader or TMU count. 740M / 860M / 840M exist on lower SKUs; they are not the shopper chips on this page.

Strix Halo is not in that table. Ryzen AI Max / Strix Halo is Radeon 8060S, gfx1151.

Shared memory

BIOS UMA / VRAM carve-out is a reserved slice the firmware hands the iGPU at boot. GTT / dynamic allocation lets the driver map more of the same kit after boot. AMD’s ROCm 10.1.0 GPU table prints VRAM for these SKUs as Dynamic + carveout. The official RDNA 3.5 system-optimization page (gfx1150 / 1151 / 1152) says memory is mapped rather than physically partitioned, GTT is the user-process pool, and the default GTT limit is about 50 percent of system RAM. That page recommends keeping the dedicated BIOS reservation small (example: 0.5 GB) and raising the kernel TTM page limit at /sys/module/ttm/parameters/pages_limit if you need a larger map. We do not pin a megabyte number or a ttm.pages_limit value — read the board menu on /bios and the sysfs file on your box.

llama.cpp Vulkan allocates GPU buffers through Mesa RADV. Official llama.cpp build.md: cmake -DGGML_VULKAN=ON, then --list-devices at runtime. How much of the kit the process can actually hold is the GTT / TTM map plus whatever the OS still needs — not a second physical pool. The calculator’s 70% GPU slice is a house estimate so the desktop keeps a cut, not an amdgpu reading.

Backend

llama.cpp Vulkan is the AMDai first token. It does not need ROCm. Official build.md covers the Vulkan backend on Linux. HIP / ROCm is second, and only after rocminfo lists your gfx. ROCm 10.1.0 names Hawk Point 780M / 760M as gfx1103 and Strix Point 880M / 890M as gfx1150. Desktop Phoenix 8700G / 8600G use the same 780M / 760M marketing names and the GPU-specs table lists Phoenix 7840U 780M as gfx1103, but those 8000G desktop SKUs are not named on the 10.1 compatibility matrix — do not claim official ROCm for 8700G. HSA_OVERRIDE_GFX_VERSION is unofficial. Do not spoof gfx1100 on gfx1103; Vulkan does not need the override.

• ROCm 10.1 matrix
• llama.cpp
• gfx IDs
• Ubuntu setup

What fits

Qwen 3.6 and Qwen 3.8 only. Q6 / Q5 / Q4 are calculator estimates. UD-Q3_K_XL is a labeled estimate (Hugging Face listing for Qwen3.6-35B-A3B-UD-Q3_K_XL.gguf (16.8 GB)). Q1/Q2 stay off. Minimum system RAM needed and RAM needed at full context below are the same two shopper labels as the rest of AMDai, adapted for one shared pool.

The calculator’s unified default is 70% of the kit as the GPU slice — an estimate, not a BIOS reading. A 32 GB kit yields 22.4 GB; 64 GB yields 44.8 GB; 96 GB yields 67.2 GB. Left for the OS after that slice is 9.60 GB, 19.2 GB, and 28.8 GB.

Qwen3.6-35B-A3B at 32K

Qwen 3.8 Flash at 32K

Minimum system RAM needed for Qwen3.6-35B-A3B on llama.cpp at Q4_K_M, 32K. One shared pool: model total plus the calculator OS/desktop (8 GB) and headroom (16 GB). Estimate until a measured card overlays the kit. AMD lists max memory 256 GB on the cited 8700G and HX 370 pages — we size 32 / 64 / 96 GB kits.

Speeds

No cited tok/s on this page. local desktop has no 780M / 890M card, and this repo has no third-party Ryzen iGPU speed row we can quote. Memory type and official max speed (DDR5-5200 on 8700G; DDR5-5600 / LPDDR5X-8000 on HX 370) are the bandwidth story — we do not invent a GB/s figure. Empty stays empty. Record a card with amdai-bench / Stack Lab on your own box.

House launch

llama.cpp Vulkan, sane 32K context, f16 KV. Confirm --list-devices before you paste this. A 35B Q5 file wants a kit that the calculator says fits — the command is a shape, not a measured pin.

# Shape only. Not a measured iGPU pin. Size the file on /calculator first.
# llama-cli --list-devices  → pick RADV GFX1103 (780M/760M) or RADV GFX1150 (880M/890M).
./build-vk/bin/llama-server \
  -m ~/models/Qwen3.6-35B-A3B-UD-Q5_K_XL.gguf \
  -ngl 99 \
  -c 32768 \
  --cache-type-k f16 \
  --cache-type-v f16 \
  --host 127.0.0.1 --port 8080

When to add a dGPU

Add a discrete card when the file plus KV no longer fit the kit, or when you want the measured 35B in-VRAM path. The R9700 sheet is the measured 32 GB discrete page. Used and budget boards live on /gpus, /budget, /calculator, and /can-run. Strix Halo is the larger shared-memory APU (AMD max 128 GB on that SKU), not a faster 890M. Two discrete cards are a different layout.

Radeon Stack Lab house profiles are the Ubuntu R9700 recipes. There is no Ryzen iGPU package and no public download. Open in Stack Lab is not on this page because those profiles do not apply. Record an amdai-bench card if you measure one.

• GPU directory
• R9700
• Strix Halo
• Budget
• Calculator
• Can my Radeon run it?
• Compare GPUs

Sources

• AMD — Ryzen 7 8700G product page (Radeon 780M)
• AMD — Ryzen 5 8600G product page (Radeon 760M)
• AMD — Ryzen AI 9 HX 370 product page (Radeon 890M)
• AMD — Ryzen AI 9 365 product page (Radeon 880M)
• AMD — ROCm 10.1.0 GPU specifications (Ryzen APUs; Dynamic + carveout)
• AMD — ROCm 10.1.0 compatibility matrix (gfx1103 / gfx1150)
• AMD — RDNA 3.5 system optimization (GTT / UMA / TTM)
• llama.cpp — build.md (Vulkan backend)

Compare with other sheets

Head-to-head pages use the same VRAM, bandwidth, and calculator fit already on the sheets. Q3 is a labeled estimate. No invented prices. Open the GPU picker to choose two sheets.

• Compare GPUs