Minimum system RAM needed
64 GB kit
45.2 GB at 32K. Left after the model: 42.8 GB on 64 GB, 106.8 GB on 128 GB.
GPU directory · Ryzen AI Max · RDNA 3.5 · gfx1151
Cited, not measured on the local desktop
Ryzen AI Max / Strix Halo
Strix Halo is an APU: CPU and GPU live on one chip and share one memory pool. There is no discrete VRAM sticker. Mini-PCs and laptops built on Ryzen AI Max+ 395 with Radeon 8060S (gfx1151) are the shopper box. AMD lists max system memory 128 GB (256-bit LPDDR5x). We size two kits — 64 GB and 128 GB — as calculator estimates. Phoenix / Hawk Point 780M and Strix Point 880M / 890M are a different sheet at /gpus/ryzen-igpu.
Strix Halo APU. Official on the ROCm 10.1.0 Ryzen table as gfx1151. VRAM is Dynamic + carveout — shared system memory, not a discrete sticker. Official on the current ROCm matrix as gfx1151.
What fits on this card (256K context)
Shopper verdict at 4K, then follow the pair page for 32K and full context. Q4 and up. No printed prices.
| Model | Verdict |
|---|---|
| Qwen3.6-35B on Strix Halo | Fits fully on GPU |
| Qwen 3.8 Flash on Strix Halo | Does not fit |
| gpt-oss 20B on Strix Halo | Fits fully on GPU |
| Gemma 3 27B on Strix Halo | Fits fully on GPU |
| Llama 3.3 70B on Strix Halo | Fits with RAM offload |
| GLM 5.3 Flash on Strix Halo | Does not fit |
| Qwen3.6 27B on Strix Halo | Fits fully on GPU |
| Qwen3.8 27B on Strix Halo | Fits fully on GPU |
Open 64 GB in calculator
Open 128 GB in calculator
Out of VRAM
R9700 measured page
780M / 890M iGPU
Ubuntu setup
How we test — Measured cells cite official amdai-bench/run@1 cards. Estimates stay labeled. Q3 is estimate-only. Q1/Q2 stay off.
What it is
BIOS UMA / VRAM carve-out is a reserved slice the firmware hands the iGPU at boot. GTT / dynamic allocation lets the driver grow into the rest of the same kit. AMD’s ROCm 10.1.0 GPU table prints VRAM for this SKU as Dynamic + carveout — not a 32 GB card. A large carve-out leaves less for the OS; a tiny carve-out leans on GTT. Check the board menu on /bios. We do not pin a megabyte number.
Go deeper UMA carve-out versus GTT
The calculator’s unified default is 70% of the kit as the GPU slice — an estimate, not a BIOS reading. A 64 GB kit yields 44.8 GB; a 128 GB kit yields 89.6 GB. Left for the OS after that slice is 19.2 GB and 38.4 GB. Change the slice in the calculator if your BIOS printout disagrees — do not invent a megabyte figure here.
Support status
ROCm 10.1.0 lists AMD Ryzen AI Max+ 395 (Radeon 8060S) as gfx1151 on the Ryzen APU table. That is official support on the cited matrix, not a tok/s card. llama.cpp Vulkan is the AMDai first token — same rule as other consumer Radeon pages. HIP / ROCm is second, and only after rocminfo lists gfx1151. RDNA 3.5 notes add kernel-version caveats for Ryzen AI Max. The known MES queue-removal hang lives on /rocm-problems — this page does not retell it.
Go deeper HIP after Vulkan tokens
HIP is the ROCm compile when rocminfo already lists gfx1151. Family packages are gfx-specific — a gfx1201 R9700 pin is the wrong binary. Stack Lab house profiles stay on the measured R9700. Radeon Stack Lab house profiles are the Ubuntu R9700 recipes. There is no Strix Halo package and no public download. Open in Stack Lab is not on this page because those profiles do not apply.
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 64 GB kit yields 44.8 GB; a 128 GB kit yields 89.6 GB. Left for the OS after that slice is 19.2 GB and 38.4 GB.
Qwen3.6-35B-A3B at 32K
| Quant | Kind | Weights | 64 GB kit | 128 GB kit | Speed |
|---|---|---|---|---|---|
| Q6_K | Estimate | 26.7 GB |
28.2 GB · Fits in VRAM (estimate) GPU slice 44.8 GB. Left for OS 35.8 GB. |
28.2 GB · Fits in VRAM (estimate) GPU slice 89.6 GB. Left for OS 99.8 GB. |
no cited number |
| Q5_K_M | Estimate | 24.8 GB |
26.2 GB · Fits in VRAM (estimate) GPU slice 44.8 GB. Left for OS 37.8 GB. |
26.2 GB · Fits in VRAM (estimate) GPU slice 89.6 GB. Left for OS 101.8 GB. |
no cited number |
| Q4_K_M | Estimate | 19.8 GB |
21.2 GB · Fits in VRAM (estimate) GPU slice 44.8 GB. Left for OS 42.8 GB. |
21.2 GB · Fits in VRAM (estimate) GPU slice 89.6 GB. Left for OS 106.8 GB. |
no cited number |
| UD-Q3_K_XL | Q3 estimate | 15.6 GB |
17.0 GB · Fits in VRAM (estimate) GPU slice 44.8 GB. Left for OS 47.0 GB. |
17.0 GB · Fits in VRAM (estimate) GPU slice 89.6 GB. Left for OS 111.0 GB. |
no cited number |
Qwen 3.8 Flash at 32K
| Quant | Kind | Weights | 64 GB kit | 128 GB kit | Speed |
|---|---|---|---|---|---|
| Q6_K | Estimate | 157.4 GB |
161.2 GB · Does not fit this GPU GPU slice 44.8 GB. Left for OS -97.217 GB. |
161.2 GB · Does not fit this GPU GPU slice 89.6 GB. Left for OS -33.217 GB. |
no cited number |
| Q5_K_M | Estimate | 147.4 GB |
151.1 GB · Does not fit this GPU GPU slice 44.8 GB. Left for OS -87.103 GB. |
151.1 GB · Does not fit this GPU GPU slice 89.6 GB. Left for OS -23.103 GB. |
no cited number |
| Q4_K_M | Estimate | 103.4 GB |
106.4 GB · Does not fit this GPU GPU slice 44.8 GB. Left for OS -42.390 GB. |
106.4 GB · Does not fit this GPU GPU slice 89.6 GB. Left for OS 21.6 GB. |
no cited number |
| UD-Q3_K_XL | Q3 estimate | 83.8 GB |
86.5 GB · Does not fit this GPU GPU slice 44.8 GB. Left for OS -22.539 GB. |
86.5 GB · Tight — little desktop headroom GPU slice 89.6 GB. Left for OS 41.5 GB. |
no cited number |
RAM needed at full context
64 GB kit
49.6 GB at 256K. Left after the model: 38.4 GB on 64 GB, 102.4 GB on 128 GB.
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. AMD lists max memory 128 GB — we do not invent a larger Strix Halo kit.
Go deeper Why 70% is an estimate
The unified-memory calculator already uses a 70% GPU slice for APU / iGPU
presets. That is a house estimate so the OS keeps a slice, not a Metal or
amdgpu reading. Raise or lower use on the
calculator share URL if your BIOS carve-out is different.
Speeds
No cited tok/s on this page. local desktop has no Strix Halo card, and this repo has no third-party Strix Halo 64 GB or 128 GB speed row we can quote. Empty stays empty. Do not reuse R9700 numbers.
Strix Halo vs R9700
Capacity versus bandwidth. Strix Halo can address a much larger shared pool (AMD max 128 GB). The R9700 is a 32 GB discrete card we actually measured, with a cited 640 GB/s figure. We do not invent an LPDDR5x GB/s number for the APU. Buy the APU when the file is bigger than 32 GB and you accept shared-memory speed. Buy the R9700 when you want the measured 35B in-VRAM path, or a second card. Two discrete cards are a different layout — see /multi-gpu. Against a Mac Studio, the shopper compare is /compare/radeon-vs-mac.
Check price · no live or invented prices
Ryzen AI Max+ 395 systems
Mini-PC and laptop search for the Strix Halo APU. No live or invented price. Confirm 64 GB vs 128 GB kit on the listing.
- Amazon: check price
- Newegg: check price
- B&H Photo: check price
- Micro Center: check price
- AMD Store: check price
Return terms vary by seller and item. Newegg's Hassle-Free page covers only Sold-and-Shipped-by-Newegg listings; marketplace items use the seller's policy. See the listing's Warranty & Returns tab.
As an Amazon Associate I earn from qualifying purchases. Some Newegg links use Rakuten Advertising and some B&H links use B&H’s affiliate program when those IDs are configured. Micro Center and AMD Store links are plain “also available at” links — no tracking. Buttons say “Check price” and never invent or live-quote a price. Full disclosure.
Sources
- AMD — Ryzen AI Max+ 395 product page
- AMD — ROCm 10.1.0 GPU specifications (Ryzen APUs; Dynamic + carveout)
- AMD — ROCm 10.1.0 compatibility matrix (Ryzen AI Max+ 395, gfx1151)
- AMD — RDNA 3.5 system optimization (Ryzen AI Max kernel notes)
- ROCm #5590 — MES failed to respond to msg=REMOVE_QUEUE
- ROCm #5724 — MES firmware hang / fault on Strix Halo
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. Strix Halo against the other published sheets:
Also: PRO W6800 · PRO W7700 · PRO W7800 · PRO W7900 · RX 6700 XT · RX 6800 XT · RX 6950 XT · RX 7600 · RX 7600 XT · RX 7700 XT · RX 7800 XT · RX 7900 GRE · RX 7900 XT · RX 7900 XTX · RX 9060 XT · RX 9070 · RX 9070 GRE