Estimate · not a bench

Quant + Hardware Calculator

Three views, one engine: size a model, ask what a Radeon can hold, or read a decode-speed floor. Totals are estimates unless a tested badge shows a first-party R9700 number. Q3 is listed (estimate / community unless measured). Q1 / Q2 stay off. llama.cpp first. Those older 2.5-class and 8B names are not offered. Use with your AI assistant: npx -y amdai-fit-mcp is a stdio MCP server that lists Hugging Face GGUF files and estimates whether they fit a Radeon card. Docs ship with the npm package.

Open in Stack Lab

If the app is not installed, packages are not published yet. There is no public download URL. Read the Stack Lab install notes . If Stack Lab is already running on this machine, open the local app .

Use with your AI assistant: the public JSON API and the amdai-fit-mcp server documented there.

Estimate only — no tested recipe for this card × model.

Detect GPU from rocminfo / rocm-smi

Paste a marketing name or a gfx ID. gfx1201 is shared — the name wins when both appear.

Radeon
Model
Older models

Gemma, Llama 3, earlier Qwen 3 / 3.5, and other catalog rows. Still selectable. A share link that names one opens this list.

Quant, context, backend

Q3_K_M, Q3_K_S, UD-Q3_K_XL, and IQ3 are listed. Those rows are estimates unless a tested badge applies. Q1 / Q2 stay off, including draft heads. Our local desktop measurements stay Q4 and above.

How we test — measured cells cite run cards; estimates stay labeled.

Required VRAM · estimate

Fits in VRAM (estimate)

Per GPU vs the VRAM you entered. Totals in GiB (1024³). Desktop spare is extra.

Weights
—
KV cache
—
Overhead
—
Total
—

Per GPU · — GiB vs — GB

Usable pool · — GB

Overflow to system RAM · — GiB (—% of RAM)

Minimum system RAM needed

—

RAM needed at full context

—

Breakdown
Offloaded weights ·
—
Separable embed in RAM ·
—
KV in RAM at selected ·
—
KV in RAM at full context ·
—
Runtime / backend ·
—
OS / desktop baseline ·
—
Headroom ·
—

Best quant that fits fully in VRAM

—

Max context at that quant

—

Max context at the selected quant

—

VRAM vs context

Same estimate at each preset window. Heights are relative to the largest total in the sweep.

Matching Radeon options

Smallest published VRAM first. Fit uses this model, quant, context, and backend — plus desktop spare.

    Save this setup

    How this result is calculated

    weights_GiB ≈ total_params × 10⁹ × bpw / 8 / 1024³ A Hugging Face listing or a local GGUF file size replaces the formula when present. Active experts do not shrink the file.

    KV_GiB ≈ 2 × full_layers × heads × dim × context × bytes / 1024³ plus sliding-window layers at min(context, window), or MLA as layers × dim × tensors × context × bytes. Hybrid Qwen cards only grow token-length KV on full-attention layers.

    Backend overhead is a labeled floor plus a percent of weights+KV. llama.cpp Vulkan is the default path; HIP adds a small estimate. These are not R9700 measurements. Ollama multiplies KV by concurrent sequences; the other backends share one pool.

    A card fits when the total plus 5% spare lands in the usable pool (more spare on a split). Tight clears the label but not the spare. No overflows — the extra is system RAM offload.

    Speed floor: published bandwidth × 75% controller / (active params × bpw / 8). Offered only when the file stays in VRAM. First-party tok/s appear only on the tested badge.

    After the verdict

    Complete the build

    Size the rest of the box after you know whether this GPU and model fit. Buttons say check price. We do not invent a figure.

    Ask first

    What do you want first?

    Setup questions before the recommendation. They change the pick, not the VRAM math. Tokens per second are estimates unless a “Tested on our rig” number is showing. No cloud prices. Dual x8/x8 boards stay the default unless you explicitly pick Minimum budget.

    Priority

    Usage

    Model style

    Abliterated hub — what it is, how to vet files, the catalog. Educational only. Tested pick: being reworked.

    Board budget

    Might you add a second GPU later?

    Unless you picked Minimum budget, the default is yes: buy the dual x8/x8 board now so a later card is not stuck on a chipset leftover.

    Will you run larger models later?

    DDR5 kits are very expensive right now. Buy the capacity you will actually use. A 2-DIMM kit cannot be topped up later — you would replace it.

    Do you want the fastest cold loads?

    Motherboard

    Which board matches this build?

    Default is a dual x8/x8 board, including for one GPU, so a later card can sit on matching CPU lanes.

    AMD’s AM5 chipset table lists CPU graphics as 1×16 or 2×8. ROCm on Radeon multi-GPU wants identical width on the slots the GPUs use (x16+x16 or x8+x8), on CPU-connected slots — not a chipset leftover x4.

    Vetted dual x8/x8 list — maker manuals only. No invented prices.

    1. ASRock X870 Taichi Creator

      Dual x8/x8 ✓ Our local desktop board

      X870 · PCIe 5.0 · x8/x8 · ATX · AM5 · 256 GB max · 4 DIMMs

      Maker spec page — widened spacing easily fits two 4-slot GPUs.

      Cite: ASRock X870 Taichi Creator specifications — Specifications → Slots: CPU 2 x PCIe 5.0 x16 Slots (PCIE1 and PCIE2), support x16 or x8/x8 modes; Memory: 4 x DDR5 DIMM Slots, Max. capacity of system memory: 256GB; Features — PCIe 5.0 x8/x8 Slots & Wide Spacing. As of 2026-10-28. Manual: ASRock X870 Taichi Creator user manual — Chapter 1.2 Specifications → Expansion Slot: CPU 2 x PCIe 5.0 x16 Slots (PCIE1 and PCIE2), support x16 or x8/x8 modes.

      Check price · no live or invented prices

      ASRock X870 Taichi Creator

      Local desktop board. Maker spec — two PCIe 5.0 x16 slots, x16 or x8/x8. Search only, not a SKU or a 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.

    2. ASUS ProArt X870E-CREATOR WIFI

      Dual x8/x8 ✓

      X870E · PCIe 5.0 · x8/x8 · ATX · AM5 · 256 GB max · 4 DIMMs

      Spec page — when both PCIEX16(G5)_1 and PCIEX16(G5)_2 are used they run at x8 each. PCIEX16(G5)_2 shares bandwidth with M.2_2 (x8/x4 if that M.2 is enabled).

      Cite: ASUS ProArt X870E-CREATOR WIFI tech spec — Memory: 4 x DIMM slots, max. 256GB, DDR5. Expansion Slots — AMD Ryzen 9000 & 7000: 2 x PCIe 5.0 x16 slots (supports x16 or x8/x8). Note: when you use both PCIEX16(G5)_1 and PCIEX16(G5)_2, they will run at x8 each. Form Factor: ATX 12 x 9.6 in. As of 2026-10-28. Manual: ASUS ProArt X870E-CREATOR WIFI user manual (E28680 V2) — Specifications → Expansion Slots: 2 x PCIe 5.0 x16 slots with Q-Release Slim (supports x16 or x8/x8). When you use both PCIEX16(G5)_1 and PCIEX16(G5)_2, they will run at x8 each.

      Check price · no live or invented prices

      ASUS ProArt X870E-CREATOR WIFI

      Maker spec — two PCIe 5.0 x16 slots, x16 or x8/x8. Search only, not a SKU or a 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.

    3. ASUS ProArt X670E-CREATOR WIFI

      Dual x8/x8 ✓

      X670E · PCIe 5.0 · x8/x8 · ATX · AM5 · 192 GB max · 4 DIMMs

      Spec page — two CPU PCIe 5.0 x16 slots support x16 or x8/x8. A third full-length slot is chipset x2, not a second CPU GPU slot.

      Cite: ASUS ProArt X670E-CREATOR WIFI tech spec — Memory: 4 x DIMM, Max. 192GB, DDR5. Expansion Slots — AMD Ryzen 9000 & 8000 & 7000: 2 x PCIe 5.0 x16 slots (support x16 or x8/x8 modes). AMD X670 Chipset: 1 x PCIe 4.0 x16 slot (supports x2 mode). As of 2026-10-28. Manual: ASUS ProArt X670E-CREATOR WIFI user manual (E21293 V2) — Specifications → Expansion Slots: 2 x PCIe 5.0 x16 slots (support x16 or x8/x8 modes). Recommended VGA configuration Dual VGA: PCIEX16_1 x8 + PCIEX16_2 x8.

      Check price · no live or invented prices

      ASUS ProArt X670E-CREATOR WIFI

      Maker spec — two PCIe 5.0 x16 slots, x16 or x8/x8. Search only, not a SKU or a 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.

    4. MSI MEG X870E ACE MAX

      Dual x8/x8 ✓

      X870E · PCIe 5.0 · x8/x8 · ATX · AM5 · 256 GB max · 4 DIMMs

      Spec page — PCI_E1 and PCI_E2 are from the CPU and support PCIe 5.0 x16/x0 or x8/x8 on Ryzen 9000/7000. PCI_E3 is a CPU x4 leftover that shares bandwidth with M.2_1 — not the second GPU slot.

      Cite: MSI MEG X870E ACE MAX specification — Memory — 4x DDR5 UDIMM, Maximum Memory Capacity 256GB. Slot — PCI_E1 & PCI_E2 slots support up to PCIe 5.0 x16/x0 or x8/x8 (Ryzen 9000/7000). PCI_E1 Gen PCIe 5.0 up to x16 (From CPU); PCI_E2 Gen PCIe 5.0 up to x8 (From CPU). As of 2026-10-28. Manual: MSI MEG X870E ACE MAX English manual — Specifications → Expansion Slots: PCI_E1 & PCI_E2 (From CPU) support up to PCIe 5.0 x16/x0, x8/x8 (Ryzen 9000/7000). Form Factor: ATX 9.6 in x 12 in.

      Check price · no live or invented prices

      MSI MEG X870E ACE MAX

      Maker spec — PCI_E1/PCI_E2 from the CPU, PCIe 5.0 x16/x0 or x8/x8. Search only, not a SKU or a 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.

    5. ASRock X870E Taichi

      Dual x8/x8 ✓

      X870E · PCIe 5.0 · x8/x8 · E-ATX · AM5 · 256 GB max · 4 DIMMs

      Maker spec — PCIE1 and PCIE2 are CPU PCIe 5.0 x16 slots, x16 or x8/x8. Manual two-card CrossFire table is Gen5x8 + Gen5x8. E-ATX 12.0 in × 10.5 in — the case must take E-ATX.

      Cite: ASRock X870E Taichi specifications — Memory: Max. capacity of system memory: 256GB. Slots → CPU: 2 x PCIe 5.0 x16 Slots (PCIE1 and PCIE2), support x16 or x8/x8 modes. Form Factor: EATX 12.0-in x 10.5-in. As of 2026-10-28. Manual: ASRock X870E Taichi user manual — Chapter 1.2 Specifications → Expansion Slot: CPU 2 x PCIe 5.0 x16 Slots (PCIE1 and PCIE2), support x16 or x8/x8 modes. Expansion Slots (PCIe Slots): two graphics cards in CrossFire are Gen5x8 + Gen5x8.

      Check price · no live or invented prices

      ASRock X870E Taichi

      Maker spec — two PCIe 5.0 x16 slots, x16 or x8/x8. E-ATX. Search only, not a SKU or a 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.

    6. Gigabyte B850 AI TOP

      Dual x8/x8 ✓

      B850 · PCIe 5.0 · x8/x8 · ATX · AM5 · 256 GB max · 4 DIMMs

      Spec page — two PCIe 5.0 x16 slots run x16 or x8, x8. Manual — PCIEX16 and PCIEX8 are CPU slots; when PCIEX8 is populated, PCIEX16 drops to x8. PCIEX2 is chipset x2 and shares with M2C_SB, not a second GPU slot. Ryzen 8000 APUs make PCIEX8 unavailable.

      Cite: Gigabyte B850 AI TOP specifications — Memory — 4 x DDR5 DIMM sockets supporting up to 256 GB (64 GB single DIMM capacity). Full-length PCIe slots — 2 x PCIe 5.0 x16 (run x16 or x8, x8) with PCIe UD Slot X; 1 x PCIe 4.0 x2 Slot. As of 2026-10-28. Manual: Gigabyte B850 AI TOP user manual (Rev. 1004) — 2-2 Product Specifications → Expansion Slots: PCIEX16 and PCIEX8 integrated in the CPU; when PCIEX8 is populated, PCIEX16 operates at up to x8. Ryzen 9000/7000: PCIEX8 is PCIe 5.0 x8. Form Factor: ATX 30.5cm x 24.4cm.

      Check price · no live or invented prices

      Gigabyte B850 AI TOP

      Maker spec — two PCIe 5.0 x16 slots, x16 or x8/x8. Search only, not a SKU or a 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.

    7. ASUS ProArt B850-CREATOR WIFI NEO

      Dual x8/x8 ✓

      B850 · PCIe 5.0 · x8/x8 · ATX · AM5 · 256 GB max · 4 DIMMs

      Spec page — two PCIe 5.0 x16 slots support x16 or x8/x8; when both PCIEX16(G5)_1 and PCIE16(G5)_2 are used they run at x8 each. The chipset PCIEX16(G4) is x4 and shares with M.2_3 — not a second CPU GPU slot. Ryzen 8000 APUs: one x16 slot at x8/x4.

      Cite: ASUS ProArt B850-CREATOR WIFI NEO tech spec — Memory: 4 x DIMM slots, max. 256GB, DDR5. Expansion Slots — AMD Ryzen 9000 & 7000: 2 x PCIe 5.0 x16 slots (supports x16 or x8/x8 modes). Note: when you use both PCIEX16(G5)_1 and PCIE16(G5)_2, they will run at x8 each. Form Factor: ATX 12 x 9.6 in. As of 2026-10-28. Manual: ASUS ProArt B850-CREATOR WIFI NEO user manual (E27577) — English user’s manual (E27577, March 2026) → Expansion Slots: 2 x PCIe 5.0 x16 slots (supports x16 or x8/x8 modes). When you use both PCIEX16(G5)_1 and PCIE16(G5)_2, they will run at x8 each.

      Check price · no live or invented prices

      ASUS ProArt B850-CREATOR WIFI NEO

      Maker spec — two PCIe 5.0 x16 slots, x16 or x8/x8. Search only, not a SKU or a 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.

    System RAM

    Which DDR5 kit matches this build?

    Recommend 96 GB (2x48 DDR5) for most builds. With two R9700s it runs Qwen 3.8 Flash Q4 at full 256K context (estimate: about 67 GB needed). Two DIMMs, so it stays at AMD’s official 2-DIMM speed.

    DDR5 kits are very expensive right now. Buy the capacity you will actually use. A 2-DIMM kit cannot be topped up later — you would replace it.

    We still recommend 96 GB (2x48 DDR5) for most builds, so Qwen 3.8 Flash stays an option. 32 GB (2x16) is enough only if you will run the 35B on one card and never want Flash.

    AMD’s official Ryzen 9000 memory speed is DDR5-5600 with 2 DIMMs and DDR5-3600 with 4 DIMMs (product pages). Running 4 DIMMs at 6000 is an overclock. Four DIMMs usually run at a lower DDR5 speed than two. That is a bandwidth trade, not a capacity trade.

    You cannot top up a 2-DIMM kit later. Memory makers say do not mix kits, and 192 GB / 256 GB only exist as matched 4x48 or 4x64 sets — moving up means replacing the kit you bought.

    Buttons say Check price. Any price shown carries the date we checked it — never a live or invented price.

    35B only · one card

    G.SKILL Flare X5 32 GB (2x16) DDR5-6000 CL30

    32 GB · 2x16 DDR5 · F5-6000J3038F16GX2-FX5

    You cannot top up a 2-DIMM kit later. Memory makers say do not mix kits, and 192 GB / 256 GB only exist as matched 4x48 or 4x64 sets — moving up means replacing the kit you bought.

    Cite: G.SKILL F5-6000J3038F16GX2-FX5 specification — 32GB (16GBx2); Tested Latency 30-38-38-96; Tested Voltage 1.35V; AMD EXPO; Do not mix memory kits. As of 2026-10-29.

    Check price · no live or invented prices

    G.SKILL Flare X5 32 GB (2x16) DDR5-6000 CL30

    Named kit F5-6000J3038F16GX2-FX5. Search only — no live or invented price. A 2x16 kit cannot be topped up later.

    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.

    Recommended · most builds

    G.SKILL Flare X5 96 GB (2x48) DDR5-5600 CL40

    96 GB · 2x48 DDR5 · F5-5600J4040D48GX2-FX5

    96 GB (2x48 GB) DDR5, the right size for most builds. Our pick: G.SKILL Flare X5 96 GB DDR5-5600 (F5-5600J4040D48GX2-FX5), AMD EXPO + Intel XMP. On G.SKILL’s tested list for the ASRock X870E Taichi, Gigabyte B850 AI TOP, and GIGABYTE X870E AORUS MASTER X3D. Two sticks keep the full 5600 speed on Ryzen 9000; four sticks drop it. ASRock X870 Taichi Creator, ASUS ProArt X870E-CREATOR WIFI, and ASUS ProArt B850-CREATOR WIFI NEO are untested on the 96 GB QVLs we read — not proof they are incompatible. Want the fastest tested kit? The 6000 CL28 version is about $920 more for at most ~7% more memory bandwidth, so only worth it if you already run models out of system RAM all day. Prices change; check the listing.

    $1,479.99 on Oct 30, 2026 — Newegg listing sold by Newegg (list price on the search page, not a cart check). Prices change; check the current price before you buy.

    Cite: G.SKILL F5-5600J4040D48GX2-FX5 specification — 96GB (48GBx2); Dual Channel Kit; AMD EXPO + Intel XMP; SPD Speed (Default) 5600 MT/s; Tested Latency 40-40-40-89; Do not mix memory kits. As of 2026-10-30.

    Check price · no live or invented prices

    G.SKILL Flare X5 96 GB (2x48) DDR5-5600 CL40

    Named kit F5-5600J4040D48GX2-FX5 (AMD EXPO + Intel XMP). Recommended size for most builds (2x48). Newegg list price $1,479.99 on 2026-10-30, sold by Newegg (dated snapshot, not a live quote). Amazon searches the FX5 part number.

    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.

    Fastest tested: G.SKILL Trident Z5 Neo RGB F5-6000J2836F48GX2-TZ5NR (Newegg sold the white -TZ5NRW). On G.SKILL’s list for the same three boards plus MSI MEG X870E GODLIKE. About $920 more for at most ~7% more theoretical bandwidth (89.6 → 96.0 GB/s).

    $2,399.99 on Oct 30, 2026 — Newegg listing sold by Newegg for the white -TZ5NRW (list price on the search page, not a cart check). Prices change; check the current price before you buy.

    Fastest tested · Check price at Newegg

    RAM sizes are estimates · speeds measured on our rig

    Is more RAM worth it?

    For most people, no. 96 GB (2x48 DDR5) is the right size for most builds: with two R9700s it runs Qwen 3.8 Flash Q4 at full 256K context (estimate: about 67 GB needed). Spending more on RAM lets you load a larger, smarter model, but more of that model has to sit in system RAM instead of on the graphics cards, so it answers a lot slower. It is usually not worth it unless you really want the maximum intelligence you can run at home and are okay with it being slow.

    RAM kit Largest model it fits (two R9700s, 256K) RAM needed (estimate) What you get
    96 GB Qwen 3.8 Flash UD-Q4_K_XL ~67 GB est. Recommended for most builds. Runs Flash Q4 at full 256K context with room to spare.
    128 GB Qwen 3.8 Flash UD-Q5_K_XL / UD-Q6_K_XL ~111–121 GB est. Same model, higher-precision file. A bit smarter, and more of it sits in system RAM.
    192 GB GLM-5.3-Flash UD-Q4_K_XL ~147 GB est. A larger model. Smarter on hard problems, and much more of it sits in system RAM.
    256 GB GLM-5.3-Flash UD-Q5_K_XL / UD-Q6_K_XL ~184–232 GB est. The largest files we size. Maximum intelligence on this hardware, and the slowest output.

    Why it gets slower: every token reads the model weights. The R9700’s own memory is 640 GB/s (AMD spec). Dual-channel DDR5 on a Ryzen 9000 desktop reads at roughly 55–64 GB/s, so weights that live in system RAM are read about 10x or more slower (spec-sheet estimate, not a benchmark). The bigger the file, the more of it lands in system RAM.

    Measured on our rig: Qwen3.6-35B-A3B UD-Q5_K_XL on one R9700, fully in VRAM (llama.cpp, 4K) decodes about 102 tok/s (102.39). Qwen 3.8 Flash UD-Q6_K_XL on two R9700s with experts in DDR5-5200 (Strata, 8K) decodes about 33 tok/s (32.8) — roughly a third of the speed for a bigger, smarter model. We have not measured GLM-5.3-Flash; expect it to be slower still. Run cards: 35B one card, Flash Q6 offloaded.

    Only ever running the 35B on one card? 32 GB (2x16) is enough for that, because the model stays in VRAM. Skip it if you might want Qwen 3.8 Flash later — a 2x16 kit cannot be topped up.

    RAM needs are estimates for two R9700s at 262144 context with f16 KV (file size from Hugging Face, expert split from the GGUF headers, 10 GB for the OS). Not measured host RAM. The calculator adds 16 GB headroom, so it can round one size higher.

    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.

    CPU

    Which Ryzen matches this build?

    The model lives in GPU memory, so the CPU mostly feeds the cards. The 9600X is enough. Any Ryzen 9000 desktop chip has 24 usable PCIe 5.0 lanes for x8/x8 on the right board.

    Every Ryzen 9000 desktop chip lists PCIe 5.0 with 28 total / 24 usable lanes and 256 GB max (AMD product pages), so x8/x8 works on a board whose maker manual documents that split.

    One Ryzen 9000 chiplet’s memory read is about 64 GB/s at the default fabric clock, below DDR5-6000’s 96 GB/s theoretical. Those are spec-sheet caps, not measurements on our cards.

    Buttons say Check price. We never invent a CPU price.

    Pick

    AMD Ryzen 5 9600X

    6 cores · one chiplet · 65 W TDP · PCIe 5.0, 28 total / 24 usable

    The model lives in GPU memory, so the CPU mostly feeds the cards. The 9600X is enough. Any Ryzen 9000 desktop chip has 24 usable PCIe 5.0 lanes for x8/x8 on the right board.

    Cite: AMD Ryzen 5 9600X product page — Native PCIe Lanes (Total/Usable) 28, 24; Max. Memory 256 GB; Max Memory Speed 2x1R DDR5-5600, 4x1R DDR5-3600; Default TDP 65W. As of 2026-10-29.

    Check price · no live or invented prices

    AMD Ryzen 5 9600X

    6 cores, 24 usable PCIe 5.0 lanes. Enough when the model fits in VRAM. Search only — no live or invented 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.

    Skip Ryzen 8000G for dual GPU

    Never use a Ryzen 8000G APU for dual GPU. The 8700G and 8600G have PCIe 4.0 with 8 graphics lanes — the second CPU x16 slot goes dead. The 8500G and 8300G have 4 graphics lanes, so even one card runs at x4.

    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.

    Power supply

    Which PSU matches this build?

    Sized for two R9700s. AMD lists 300 W TBP per card (product page). Two cards plus a 9950X can reach about 1,100 W. Two native 12V-2x6 cables, one per card. You said a second GPU might come later, so this skips the 850 W step and buys the PSU once.

    The R9700 takes a 12V-2x6 connector (AMD product page). Buy one native 12V-2x6 cable per card. Many 1000 W units ship only one.

    Cite: AMD Radeon AI PRO R9700 product page — Total Board Power (TBP) 300W; Minimum PSU Recommendation 750W; External Power Connectors 12V-2x6. As of 2026-10-29.

    Buttons say Check price. We never invent a PSU price.

    1200 W · 2 native 12V-2x6

    Thermaltake Toughpower GT 1200W

    Two native 12V-2x6 cables (each up to 600 W). Value pick for two R9700s.

    Cite: Thermaltake Toughpower GT 1200W — PCIe (12+4Pin) 600W × 2 cables; ATX 3.1; 80 PLUS Gold; 1200 W. As of 2026-10-29.

    Check price · no live or invented prices

    Thermaltake Toughpower GT 1200W

    1200 W ATX 3.1 with two native 12V-2x6 cables. Search only — no live or invented 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.

    1200 W · 2 native 12V-2x6

    be quiet! Straight Power 12 1200W

    Two native 12V-2x6 cables. Combined 12V-2x6 limit is 800 W: fine at 300 W TBP per card, tight if both cards spike together.

    Cite: be quiet! Straight Power 12 1200W — 12V-2x6 cables 2; Max. combined power 12V-2x6 800 W; ATX 3.1; 80 PLUS Platinum. As of 2026-10-29.

    Check price · no live or invented prices

    be quiet! Straight Power 12 1200W

    1200 W ATX 3.1 with two native 12V-2x6. Combined 12V-2x6 limit 800 W. Search only — no live or invented 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.

    1300 W · 2 native 12V-2x6

    Thermaltake Toughpower TF3 1300W

    Two native 12V-2x6 cables included. Headroom pick for two R9700s.

    Cite: Thermaltake Toughpower TF3 1300W — Dual native 12V-2x6 connectors; two PCIe 12+4 pin cables; 80 PLUS Titanium; 1300 W. As of 2026-10-29.

    Check price · no live or invented prices

    Thermaltake Toughpower TF3 1300W

    1300 W ATX 3.1 with two native 12V-2x6 cables. Search only — no live or invented 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.

    1500 W · 2 native 12V-2x6

    Corsair HX1500i (2025)

    Two 12V-2x6 GPU cables. Rated 1500 W at 115–240 V but 1200 W at 100–115 V, so on a weak 120 V outlet plan on 1200 W.

    Cite: Corsair HX1500i (2025) product page — Dual 12V-2x6 cables; TOTAL POWER 1200W (100V - 115V) 1500W (115V - 240V); ATX 3.1. As of 2026-10-29.

    Check price · no live or invented prices

    Corsair HX1500i (2025)

    Two 12V-2x6 cables. 1200 W at 100–115 V, 1500 W at 115–240 V. Search only — no live or invented 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.

    Storage

    Which NVMe holds the models?

    Your lanes need about 1587 GiB (every Q4–Q6 quant of Qwen3.6-35B, Qwen3.8-Flash-Next and GLM-5.3-Flash, one extra file of download headroom, 150 GiB for the OS). A 2 TB drive is 1,863 GiB usable. Gen4 TLC with DRAM is enough: published tests show Gen5 loads big GGUFs only about 9% faster on the default mmap path (15.2 s vs 13.8 s). Spend the difference on capacity or RAM.

    Run models from NVMe, never from a hard drive. On dual-GPU boards use an M.2 slot that does not share GPU lanes.

    llama.cpp’s default mmap load is usually page-fault/CPU bound, not drive-bound. Direct I/O loads at disk speed. A third-party mmap test loaded a 40.5 GB Q4_K_M file in 15.2 s on a Gen4 990 PRO versus 13.8 s on a Gen5 T705 (about 9%).

    File sizes: Hugging Face tree API, summed shard bytes (unsloth repos), checked 2026-10-24 and 2026-10-28. Buttons say Check price. We never invent an SSD price.

    Pick

    Samsung 990 PRO 2TB

    MZ-V9P2T0BW · PCIe Gen4 · 7,450 MB/s sequential read · 1,200 TBW · 2 GB LPDDR4 · TLC

    Cite: Samsung 990 PRO NVMe SSD datasheet Rev. 2.0 — 2 TB: Sequential Read 7,450 MB/s; TBW 1,200 TB; 2 GB LPDDR4; TLC. As of 2026-10-28.

    Check price · no live or invented prices

    Samsung 990 PRO 2TB

    Gen4 TLC with 2 GB LPDDR4. Datasheet sequential read 7,450 MB/s, 1,200 TBW. Search only — no live or invented 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.

    Case

    Which case fits the cards?

    Your build needs at least 7 expansion slots and room for a 271 mm card. One card: any of the three maker-checked ATX cases fit a 271 mm R9700. Blower R9700s push heat out the back; case fans cool the air the cards breathe. Point the case fans at the GPUs. A published 4-card build paused on a CPU-following curve when the card area hit 68 C; full-speed case fans cooled that area from 67 C to 57 C (GPU junction only 1–4 C).

    Blower R9700s push heat out the back; case fans cool the air the cards breathe. Point the case fans at the GPUs. A published 4-card build paused on a CPU-following curve when the card area hit 68 C; full-speed case fans cooled that area from 67 C to 57 C (GPU junction only 1–4 C).

    Cite: Published 4x R9700 build guide — card area 67 C vs 57 C. As of 2026-10-28.

    Buttons say Check price. We never invent a case price.

    Slot spacing for this board is not verified here: check its manual layout; an 8+ slot case is the safe choice. That applies to the ASUS ProArt X870E / X670E / B850 boards and the ASRock X870E Taichi.

    Pick

    Fractal Design Torrent

    7 expansion slots · GPU 461 mm · PSU 230 mm · cooler 188 mm · 360 radiator front or bottom

    Airflow pick: the most included intake, bottom fans blow straight at the cards. Fans included: 2x 180 mm front + 3x 140 mm bottom.

    Cite: Fractal Design Torrent — 7 expansion slots; E-ATX; GPU 461 mm; PSU 230 mm; 2x 180 mm front + 3x 140 mm bottom included; 360 radiator front or bottom; CPU cooler 188 mm. As of 2026-10-28.

    Check price · no live or invented prices

    Fractal Design Torrent

    7 expansion slots; E-ATX; GPU 461 mm; PSU 230 mm; 2x 180 mm front + 3x 140 mm bottom included. Search only — no live or invented 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.

    Also fits

    Fractal Design Meshify 2

    7 expansion slots · GPU 470 mm · PSU 250 mm · cooler 185 mm · 360 radiator top or front

    All-rounder: E-ATX room and storage; add a third front fan for two cards. Fans included: 2x 140 mm front + 1x 140 mm rear.

    Cite: Fractal Design Meshify 2 — 7 expansion slots; E-ATX; GPU 470 mm; PSU 250 mm; 2x 140 mm front + 1x 140 mm rear included; 360 radiator top or front; CPU cooler 185 mm. As of 2026-10-28.

    Check price · no live or invented prices

    Fractal Design Meshify 2

    7 expansion slots; E-ATX; GPU 470 mm; PSU 250 mm; 2x 140 mm front + 1x 140 mm rear included. Search only — no live or invented 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.

    Room to grow

    Phanteks Enthoo Pro 2 Server Edition

    11 expansion slots · GPU 503 mm · PSU 240 mm · cooler 195 mm · 360 radiator top or front or side

    Room to grow: 11 slots and a side fan bracket for a third card or a riser; buy fans separately. Fans included: None listed as included on the product page.

    Cite: Phanteks Enthoo Pro 2 Server Edition — 11 expansion slots; E-ATX / SSI-EEB; GPU 503 mm; PSU 240 mm; fans not listed as included; 360 radiator top, front or side; CPU cooler 195 mm. As of 2026-10-28.

    Check price · no live or invented prices

    Phanteks Enthoo Pro 2 Server Edition

    11 expansion slots; E-ATX / SSI-EEB; GPU 503 mm; PSU 240 mm; fans not listed as included. Search only — no live or invented 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.

    Blower R9700s push heat out the back; case fans cool the air the cards breathe. Point the case fans at the GPUs. A published 4-card build paused on a CPU-following curve when the card area hit 68 C; full-speed case fans cooled that area from 67 C to 57 C (GPU junction only 1–4 C).

    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.

    CPU cooler

    Which cooler, since Ryzen 9000X ships none?

    The 9600X ships with no cooler (AMD lists Thermal Solution (PIB): Not included on the 9600X, 9900X and 9950X product pages.) At 65 W a dual-tower air cooler is plenty; AMD recommends a premium air cooler.

    AMD lists Thermal Solution (PIB): Not included on the 9600X, 9900X and 9950X product pages. MoE expert-offload decode is limited by memory and PCIe traffic, not CPU compute, so AMD’s liquid recommendation is about peak boost, not decode speed.

    Buttons say Check price. We never invent a cooler price.

    Tall DIMM heatspreaders can hit the front fan; check RAM height against the cooler.

    Pick

    Thermalright Peerless Assassin 120 SE

    155 mm dual-tower air

    The 9600X ships with no cooler (AMD lists Thermal Solution (PIB): Not included on the 9600X, 9900X and 9950X product pages.) At 65 W a dual-tower air cooler is plenty; AMD recommends a premium air cooler.

    Cite: Thermalright Peerless Assassin 120 SE — Height 155 mm; dual-tower 120 mm; 6 heatpipes. As of 2026-10-28.

    Check price · no live or invented prices

    Thermalright Peerless Assassin 120 SE

    Dual-tower 120 mm air cooler, 155 mm tall, 6 heatpipes. Search only — no live or invented 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.

    Check price · no live prices

    Cards in this VRAM neighborhood

    Estimate only. We show the selected card and any listed GPU whose published VRAM is at least the per-GPU total. Buttons open a search — never a stale price. As an Amazon Associate I earn from qualifying purchases.

    Buy the box

    Your parts list

    Based on: 1 GPU(s), stay at this size.

    1. GPU

      AMD Radeon AI PRO R9700 32GB x1

      32 GB per card; the site’s measured cards.

      Newegg: check price

    2. Motherboard

      ASRock X870 Taichi Creator

      Default is a dual x8/x8 board, including for one GPU, so a later card can sit on matching CPU lanes.

      Newegg: check price Maker spec

    3. CPU

      AMD Ryzen 5 9600X

      The model lives in GPU memory, so the CPU mostly feeds the cards. The 9600X is enough. Any Ryzen 9000 desktop chip has 24 usable PCIe 5.0 lanes for x8/x8 on the right board.

      Newegg: check price Maker spec

    4. RAM

      G.SKILL Flare X5 96 GB (2x48) DDR5-5600 CL40

      We still recommend 96 GB (2x48 DDR5) for most builds, so Qwen 3.8 Flash stays an option. 32 GB (2x16) is enough only if you will run the 35B on one card and never want Flash. You cannot top up a 2-DIMM kit later. Memory makers say do not mix kits, and 192 GB / 256 GB only exist as matched 4x48 or 4x64 sets — moving up means replacing the kit you bought.

      Newegg: check price Maker spec

    5. PSU

      Thermaltake Toughpower GT 1200W

      Sized for two R9700s. AMD lists 300 W TBP per card (product page). Two cards plus a 9950X can reach about 1,100 W. Two native 12V-2x6 cables, one per card. You said a second GPU might come later, so this skips the 850 W step and buys the PSU once.

      Newegg: check price Maker spec

    6. Storage

      Samsung 990 PRO 2TB (MZ-V9P2T0BW)

      Your lanes need about 1587 GiB (every Q4–Q6 quant of Qwen3.6-35B, Qwen3.8-Flash-Next and GLM-5.3-Flash, one extra file of download headroom, 150 GiB for the OS). A 2 TB drive is 1,863 GiB usable. Gen4 TLC with DRAM is enough: published tests show Gen5 loads big GGUFs only about 9% faster on the default mmap path (15.2 s vs 13.8 s). Spend the difference on capacity or RAM.

      Newegg: check price Maker spec

    7. Case

      Fractal Design Torrent

      One card: any of the three maker-checked ATX cases fit a 271 mm R9700. Blower R9700s push heat out the back; case fans cool the air the cards breathe. Point the case fans at the GPUs. A published 4-card build paused on a CPU-following curve when the card area hit 68 C; full-speed case fans cooled that area from 67 C to 57 C (GPU junction only 1–4 C).

      Newegg: check price Maker spec

    8. Cooler

      Thermalright Peerless Assassin 120 SE

      The 9600X ships with no cooler (AMD lists Thermal Solution (PIB): Not included on the 9600X, 9900X and 9950X product pages.) At 65 W a dual-tower air cooler is plenty; AMD recommends a premium air cooler.

      Newegg: check price Maker spec

    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.

    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.

    esc

    Type to filter. Enter opens the first match. Full-text search lives on /search.