TinkerLab
Jump to Config Artifact
01 // HOMELABNVIDIA GPUDocker / ComposeStreaming / API~15 min read & lab

High-VRAM Non-CUDA Workstations for 70B–122B Local LLMs: Intel Arc Pro B60 Dual 48GB ("Battlematrix" x8/x8 Bifurcation · up to 192GB) vs Tenstorrent QuietBox 2 (Blackhole 128GB) & Intel Crescent Island (160GB/480GB Xe3P)

How do you assemble 48GB to 192GB+ of dedicated accelerator VRAM under a desk without paying enterprise NVIDIA HBM prices? We benchmark two shipping non-CUDA workstation architectures and model Intel’s new inference-only accelerator: (1) Intel Arc Pro B60 Dual 48GB ("Project Battlematrix" — Maxsun / Sparkle dual-BMG-G21 dies on one dual-slot 400W PCB, scaling up to 4 cards / 8 GPUs / 192GB VRAM), including the critical PCIe 5.0 x8/x8 motherboard bifurcation requirement that silently hides 24GB if misconfigured; (2) Tenstorrent QuietBox 2 (liquid-cooled 2× Blackhole p300c = 4 Tensix ASICs · 128GB GDDR6 + 256GB DDR5 connected over internal 800GbE mesh); and (3) spec-derived sizing for Intel GPU D-Series "Crescent Island" (Xe3P · 160GB reference / 480GB ODM LPDDR5X @ 1.5 TB/s).

[KEY VERIFIED TAKEAWAY / TL;DR]In Tinker AI Lab's reproducible verification of "High-VRAM Non-CUDA Workstations for 70B–122B Local LLMs: Intel Arc Pro B60 Dual 48GB ("Battlematrix" x8/x8 Bifurcation · up to 192GB) vs Tenstorrent QuietBox 2 (Blackhole 128GB) & Intel Crescent Island (160GB/480GB Xe3P)" (tested on Measured on: 2× Maxsun Intel Arc Pro B60 Dual 48G Turbo (4× BMG-G21 dies = 96GB GDDR6 · PCIe 5.0 x8/x8 bifurcated on WRX90) & Tenstorrent QuietBox 2 (Ryzen 7 9700X + 2× Blackhole p300c liquid-cooled = 4 Blackhole ASICs · 128GB GDDR6 + 256GB DDR5) | Spec-estimate row: Intel GPU D-Series "Crescent Island" D270 (32 Xe3P cores · 160GB / 480GB LPDDR5X @ 1.5 TB/s · 350W) with Intel Battlematrix Stack: `intel/llm-scaler-vllm:latest` (`vLLM-XPU` 0.8.5+, OneAPI Level Zero, `ZE_AFFINITY_MASK`) & `llama.cpp` b4980+ (SYCL / Vulkan `-sm layer`), Tenstorrent Stack: `tt-metal` (`tt-metalium` v0.58+) & `TT-Inference-Server` (`vLLM` Tensix backend, `bfp8` / `bfp4` mesh execution), Models: Llama-3.3-70B-Instruct (FP8 / Q4_K_M), Qwen3.5-35B-A3B (FP8 / Q4_K_M), Qwen3.5-122B-A10B (Q4_K_M / FP8) & OpenAI gpt-oss-120b (MXFP4), verified 2026-11-11): How do you assemble 48GB to 192GB+ of dedicated accelerator VRAM under a desk without paying enterprise NVIDIA HBM prices? We benchmark two shipping non-CUDA workstation architectures and model Intel’s new inference-only accelerator: (1) Intel Arc Pro B60 Dual 48GB ("Project Battlematrix" — Maxsun / Sparkle dual-BMG-G21 dies on one dual-slot 400W PCB, scaling up to 4 cards / 8 GPUs / 192GB VRAM), including the critical PCIe 5.0 x8/x8 motherboard bifurcation requirement that silently hides 24GB if misconfigured; (2) Tenstorrent QuietBox 2 (liquid-cooled 2× Blackhole p300c = 4 Tensix ASICs · 128GB GDDR6 + 256GB DDR5 connected over internal 800GbE mesh); and (3) spec-derived sizing for Intel GPU D-Series "Crescent Island" (Xe3P · 160GB reference / 480GB ODM LPDDR5X @ 1.5 TB/s). Offline capability is classified as "100% Local Containerized Multi-Accelerator Serving" (Once Docker images (`intel/llm-scaler-vllm` or `tt-inference-server`) and local HuggingFace/GGUF weights are cached on NVMe, zero internet connectivity is required.), with key operational boundary: **Mandatory PCIe `x8/x8` Bifurcation on Arc Pro B60 Dual 48GB:** To hit an aggressive sub-$1,300 street price, partner Dual B60 48GB cards (Maxsun Turbo / Sparkle) **omit an onboard PLX PCIe switch chip**. Lanes `0–7` wire directly to GPU Die 0 (`24GB`) and lanes `8–15` wire to GPU Die 1 (`24GB`). If your motherboard BIOS cannot bifurcate the physical ×16 slot into **`x8/x8`**, Die 1 never powers up on the bus and `lspci` only sees a single 24GB GPU!
REPRODUCIBILITY MANIFEST // TESTED SPECS & LIMITS
Last Verified: 2026-11-11
Tested Hardware & OS

Measured on: 2× Maxsun Intel Arc Pro B60 Dual 48G Turbo (4× BMG-G21 dies = 96GB GDDR6 · PCIe 5.0 x8/x8 bifurcated on WRX90) & Tenstorrent QuietBox 2 (Ryzen 7 9700X + 2× Blackhole p300c liquid-cooled = 4 Blackhole ASICs · 128GB GDDR6 + 256GB DDR5) | Spec-estimate row: Intel GPU D-Series "Crescent Island" D270 (32 Xe3P cores · 160GB / 480GB LPDDR5X @ 1.5 TB/s · 350W)

Pinned Stack & Model Versions
Intel Battlematrix Stack: `intel/llm-scaler-vllm:latest` (`vLLM-XPU` 0.8.5+, OneAPI Level Zero, `ZE_AFFINITY_MASK`) & `llama.cpp` b4980+ (SYCL / Vulkan `-sm layer`)Tenstorrent Stack: `tt-metal` (`tt-metalium` v0.58+) & `TT-Inference-Server` (`vLLM` Tensix backend, `bfp8` / `bfp4` mesh execution)Models: Llama-3.3-70B-Instruct (FP8 / Q4_K_M), Qwen3.5-35B-A3B (FP8 / Q4_K_M), Qwen3.5-122B-A10B (Q4_K_M / FP8) & OpenAI gpt-oss-120b (MXFP4)
Offline & Network Boundary
100% Local Containerized Multi-Accelerator Serving

Once Docker images (`intel/llm-scaler-vllm` or `tt-inference-server`) and local HuggingFace/GGUF weights are cached on NVMe, zero internet connectivity is required.

Known Limits & Failure Modes
  • **Mandatory PCIe `x8/x8` Bifurcation on Arc Pro B60 Dual 48GB:** To hit an aggressive sub-$1,300 street price, partner Dual B60 48GB cards (Maxsun Turbo / Sparkle) **omit an onboard PLX PCIe switch chip**. Lanes `0–7` wire directly to GPU Die 0 (`24GB`) and lanes `8–15` wire to GPU Die 1 (`24GB`). If your motherboard BIOS cannot bifurcate the physical ×16 slot into **`x8/x8`**, Die 1 never powers up on the bus and `lspci` only sees a single 24GB GPU!
  • **Above 4G Decoding + >=32GB Resizable BAR Per Die:** Each BMG-G21 die exposes a 24GB VRAM aperture (`Region 2` BAR2 = `32GB` power-of-two window). Installing two Dual B60 cards (`4 dies = 96GB`) or four Dual B60 cards (`8 dies = 192GB`) requires **128GB–256GB of contiguous 64-bit MMIO address space**—consumer boards with buggy UEFI MMIO tables may hang at POST unless `Above 4G Decoding` and `MMIO High Base` (`>= 512G` on WRX90/Xeon) are enabled.
  • **Intel "Crescent Island" (GPU D-Series) rows are SPEC-DERIVED ESTIMATES:** Crescent Island (`Xe3P` with 16-deep systolic XMX and 160GB/480GB LPDDR5X @ `1.5 TB/s`) strips out all display/ray-tracing silicon for pure datacenter/server inference; figures in Table 1 for Crescent Island are modeled from published `1.5 TB/s` bandwidth (`× 0.65` efficiency) rather than lab-measured.
Hit an edge-case failure?Open Troubleshooter →

1. Three Non-CUDA Paths to 48GB–480GB Accelerator Memory: Battlematrix Dual-Die GDDR6 vs Tensix Mesh vs Xe3P LPDDR5X

For teams serving `70B` dense models (`Llama-3.3-70B`), `120B` MoE models (`Qwen3.5-122B-A10B`, `gpt-oss-120b`), or high-concurrency multi-agent `vLLM` queues without spending $8,500–$30,000 per NVIDIA workstation/datacenter GPU, three distinct hardware philosophies have matured in 2026: 1. **Intel Arc Pro B60 Dual 48GB ("Project Battlematrix"):** Packs **two independent Xe2 `BMG-G21` GPUs** (`160 XMX engines` + `24GB 192-bit GDDR6 @ 456 GB/s` each = **`48GB` total, `912 GB/s` aggregate**) onto a single dual-slot, 400W blower PCB powered by one 16-pin cable. On a workstation board with PCIe bifurcation (such as Threadripper PRO WRX90 or Xeon W790), **two cards deliver `96GB` (`1,824 GB/s` aggregate)** for ~$2,500, and **four cards deliver `192GB` across 8 XPU dies** inside a single tower! 2. **Tenstorrent QuietBox 2 (Blackhole `128GB` Liquid-Cooled Appliance):** Instead of routing multi-GPU `AllReduce` traffic through a host CPU’s PCIe root complex, QuietBox 2 pairs two dual-ASIC **Blackhole `p300c` liquid-cooled cards (`4 Blackhole Tensix ASICs total · 128GB GDDR6 @ ~2,048 GB/s aggregate + 256GB host DDR5`)** linked directly over Tenstorrent’s native **800GbE chip-to-chip Ethernet/NoC fabric**. Tensor Parallelism across all 4 ASICs runs with sub-microsecond fabric latency inside a whisper-quiet deskside tower (`~$9,999` turnkey). 3. **Intel GPU D-Series "Crescent Island" (`Xe3P · 160GB / 480GB LPDDR5X`):** Intel’s answer to long-context agentic inference replaces scarce HBM with a wide **LPDDR5X subsystem delivering `1.5 TB/s`** and **160GB (Intel D270 reference) up to 480GB (ODM custom)** on a single **350W air-cooled PCIe 5.0 ×16 card**, backed by 32 **Xe3P** cores featuring 16-deep systolic XMX pipelines with hardware **`MXFP4`, `NVFP4`, and `FP8`** support.
[WARNING]Motherboard Compatibility Warning Before Buying Arc Pro B60 Dual 48GB: Check your motherboard manual’s PCIe bifurcation table BEFORE ordering! Mainstream Intel B860/Z890 and AMD B650/B850 boards often only bifurcate Slot 1 as `x8/x8` (supporting 1 Dual card = 48GB) while wiring Slot 2 as `x4` via chipset (which will ONLY initialize 1 of the 2 GPUs on a second Dual card!). For 2–4× Dual B60 cards (96GB–192GB), use WRX90, TRX50, Xeon W790, or EPYC boards with full `x8/x8` bifurcation per slot.

2. Benchmark & Sizing Matrix: Arc Pro B60 Dual (48GB / 96GB) vs Tenstorrent QuietBox 2 (128GB) vs Crescent Island (160GB)

We benchmarked **1× Arc Pro B60 Dual (`48GB`, `TP=2`)** and **2× Arc Pro B60 Dual (`96GB`, `TP=4`)** using Intel’s official `intel/llm-scaler-vllm` (`vLLM-XPU`) container and `llama.cpp` SYCL (`-sm layer`), compared directly against the **Tenstorrent QuietBox 2 (`128GB` Blackhole 4-ASIC mesh)** running `TT-Inference-Server`, plus a **spec-derived estimate** row for **Intel Crescent Island D270 (`160GB LPDDR5X @ 1.5 TB/s`)**:
Table 1: High-VRAM Non-CUDA Accelerator Comparison (Context = 16,384; Crescent Island row is SPEC-ESTIMATE)
Accelerator Setup & Bus TopologyTotal VRAM & Aggregate BandwidthQwen3.5-35B-A3B MoE (4-bit / FP8)Llama-3.3-70B Dense (Q4 / FP8)Qwen3.5-122B-A10B / gpt-oss-120b
**1× Intel Arc Pro B60 Dual (`2 dies · PCIe 5.0 x8/x8 · 400W`)****48 GB GDDR6** (`912 GB/s` = 2×456)**98.4 tok/s** (`vLLM-XPU TP=2`)**14.2 tok/s** (`INT4/Q4_K_M`, tight 42.5GB)OOM on 1 card (needs CPU `--n-cpu-moe` offload)
**2× Intel Arc Pro B60 Dual (`4 dies · Battlematrix · 800W`)****96 GB GDDR6** (`1,824 GB/s` = 4×456)**146.0 tok/s** (`FP8 TP=4`, C=1) / **620+ tok/s** (C=16)**23.8 tok/s** (`FP8`, 72GB across 4 dies)**64.5 tok/s** (`gpt-oss-120b` MXFP4) / **58.2 tok/s** (`122B-A10B` Q4)
**Tenstorrent QuietBox 2 (`2× p300c = 4 Blackhole ASICs · Liquid`)****128 GB GDDR6** (`~2,048 GB/s` + 800GbE mesh)**162.5 tok/s** (`bfp8` 4-chip mesh)**31.6 tok/s** (`bfp8` 1D/2D Tensix mesh, low TTFT)**78.4 tok/s** (`122B-A10B bfp8` — 100% in 128GB GDDR6)
**1× Intel "Crescent Island" D270 (`32 Xe3P · 350W`, SPEC-ESTIMATE)****160 GB LPDDR5X** (`1,500 GB/s` single die; ODM up to 480GB)*~185–210 tok/s est.* (`FP8/MXFP4`, zero TP overhead)*~24–27 tok/s est.* (`FP8`, 72GB + 88GB free for KV!)*~95–115 tok/s est.* (`gpt-oss-120b` / `122B-A10B` native MXFP4/FP8)

3. Pre-Flight PCIe Bifurcation & ReBAR Audit + Launching `intel/llm-scaler-vllm` Across Multi-Die B60 Dual Cards

Before pulling model weights on an **Arc Pro B60 Dual 48GB** build, always run a three-point hardware audit: (1) count enumerated Intel display controllers in `lspci` (**must equal `2 × physical cards`**), (2) verify every die negotiated **`Speed 32GT/s, Width x8`** (`PCIe 5.0 ×8`) and **`Region 2: Memory ... [size=32G]`** (Resizable BAR active), and (3) allocate **`--shm-size=32g`** when launching `intel/llm-scaler-vllm` so Intel OneCCL multi-XPU shared-memory ring buffers never crash mid-prefill.
[TIP]Tenstorrent QuietBox 2 Quick-Check: On QuietBox 2, run `tt-smi` before launching `TT-Inference-Server` to verify all 4 Blackhole ASICs (`0, 1, 2, 3`) report healthy firmware and `800G` Ethernet mesh training links `UP`; pass `MESH_DEVICE=T3K` / `BH_QB2` as documented in your `tt-inference-server` release.
Verify Arc Pro B60 Dual x8/x8 bifurcation, 32GB ReBAR & launch Battlematrix vLLM-XPU
# 1. Verify BOTH BMG-G21 dies per Arc Pro B60 Dual 48GB card are visible (expect 2 dies for 1 card, 4 for 2 cards)
lspci -nn | grep -iE "(VGA|Display|3D).*8086"

# 2. Confirm PCIe Gen5 x8 (32GT/s, Width x8) & 32GB Resizable BAR (Region 2) on every die
sudo lspci -vv -d 8086: | grep -E "Region 2: Memory|LnkSta:"

# 3. Launch Intel Battlematrix vLLM-XPU container across 2 dies (48GB) or 4 dies (96GB)
docker run --rm -it \
  --device /dev/dri --group-add video --group-add render \
  --shm-size=32g --ipc=host -p 8000:8000 \
  -v ~/.cache/huggingface:/root/.cache/huggingface \
  -e ZE_AFFINITY_MASK=0,1 \
  -e SYCL_CACHE_PERSISTENT=1 \
  -e SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1 \
  intel/llm-scaler-vllm:latest \
  --model Qwen/Qwen3.5-35B-A3B-Instruct \
  --tensor-parallel-size 2 --quantization fp8 --max-model-len 65536
FIELD Q&A // 常见实测疑问与边界解答

Frequently Asked Questions & Edge Cases

Q: Why does my brand-new Intel Arc Pro B60 Dual 48GB card only show 24GB VRAM in Linux and Windows?

Your PCIe ×16 slot is running in un-bifurcated `x16` mode! Because the Arc Pro B60 Dual connects GPU 0 to PCIe lanes `0–7` and GPU 1 to PCIe lanes `8–15` **without an onboard PCIe switch chip**, the host root complex treats lanes `0–15` as a single link and stops scanning after lane 7 fits GPU 0 (`x8`). Reboot into BIOS, navigate to **Advanced → PCI Subsystem / Onboard Devices → PCIEX16 Bifurcation**, and change it from **`Auto` / `x16`** to **`x8/x8` (`2×8`)**. Both 24GB GPUs (`48GB` total) will immediately appear on next boot.

Q: For a ~$10,000 deskside 128GB–192GB local AI budget, should I build a 4× Arc Pro B60 Dual (192GB) tower, buy a Tenstorrent QuietBox 2 (128GB), or buy a Mac Studio M5 Ultra (256GB/512GB)?

Each wins a different workload: **(1) 4× Arc Pro B60 Dual (`192GB` Battlematrix, `~1,600W` peak)** excels at **high-concurrency multi-user `vLLM` batching (`C=16–64`)** and fits `120B–235B` models across 8 XMX-accelerated GPUs at the lowest $/GB of discrete VRAM, at the cost of blower fan noise and high power draw. **(2) Tenstorrent QuietBox 2 (`128GB` liquid-cooled)** gives you turnkey whisper-quiet liquid cooling, zero host-PCIe `AllReduce` bottlenecks via its internal 800GbE Tensix mesh, and a 100% open-source RISC-V hardware/software stack (`tt-metal`) for models up to `122B`. **(3) Mac Studio M5 Ultra (`256GB / 512GB · 1.2 TB/s · <220W`)** remains unbeatable if you want to run massive `235B–671B` MoE models silently at `<220W` for 1–4 concurrent local users.

REPRODUCIBLE ARTIFACT // 配套可参数化附件

Intel Arc Pro B60 Dual 48GB ("Battlematrix") PCIe x8/x8 Bifurcation Audit & vLLM-XPU Launcher

Verifies that motherboard PCIe x8/x8 bifurcation exposes both BMG-G21 dies per Arc Pro B60 Dual 48GB card, checks 32GB Resizable BAR apertures and /dev/dri render nodes, and launches intel/llm-scaler-vllm with Tensor Parallelism.

launch-b60-dual-battlematrix-vllm.sh ↗
Edit variables above to update the configuration live
#!/usr/bin/env bash
# Tinker AI — Intel Arc Pro B60 Dual 48GB ("Battlematrix") x8/x8 Audit & vLLM-XPU Launcher (Guide #45)
set -euo pipefail

MODEL="Qwen/Qwen3.5-35B-A3B-Instruct"
TP="2"
MASK="0,1"
PORT="8000"

echo "=== [1/4] Auditing PCIe x8/x8 Bifurcation & Enumerated Intel BMG-G21 GPUs ==="
GPU_COUNT=$(lspci -nn | grep -icE "(VGA|Display|3D).*8086:(e20b|e20c|e211|e212)" || true)
lspci -nn | grep -iE "(VGA|Display|3D).*8086" || true
echo "-> Detected Intel Arc Xe2 GPU dies: $GPU_COUNT (Expect 2 per Arc Pro B60 Dual 48GB card!)"
if [ "$GPU_COUNT" -lt "$TP" ]; then
  echo "[WARN] Enumerated GPU dies ($GPU_COUNT) < requested TP_SIZE ($TP)!"
  echo "       Enter motherboard BIOS -> PCIe Slot Configuration -> Set slot from [Auto/x16] to [x8/x8] (or [x8/x8/x8/x8])."
fi

echo "=== [2/4] Verifying Above 4G Decoding & >=32GB Resizable BAR (BAR2) Aperture ==="
sudo lspci -vv -d 8086: | grep -E "Region 2: Memory|LnkSta:" | head -n 8 || true
ls -la /dev/dri/renderD*

echo "=== [3/4] Checking Host IOMMU P2P Passthrough (iommu=pt recommended on WRX90 / Xeon) ==="
grep -o "iommu=[^ ]*" /proc/cmdline || echo "Note: Add iommu=pt to GRUB_CMDLINE_LINUX_DEFAULT if multi-die AllReduce stalls."

echo "=== [4/4] Launching intel/llm-scaler-vllm (TP=$TP, ZE_AFFINITY_MASK=$MASK, Port=$PORT) ==="
docker run --rm -it \
  --device /dev/dri \
  --group-add video --group-add render \
  --shm-size=32g \
  --ipc=host \
  -p "$PORT:8000" \
  -v "$HOME/.cache/huggingface:/root/.cache/huggingface" \
  -e ZE_AFFINITY_MASK="$MASK" \
  -e SYCL_CACHE_PERSISTENT=1 \
  -e SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1 \
  intel/llm-scaler-vllm:latest \
  --model "$MODEL" \
  --tensor-parallel-size "$TP" \
  --quantization fp8 \
  --max-model-len 65536 \
  --gpu-memory-utilization 0.90 \
  --trust-remote-code

More Reproducible Engineering Guides

View All 48 Guides →
Local AI / HomeLab

Apple Silicon Mac Local Model Benchmark: Fixed Tasks, Quantization, Context & Speed

Verified 2026-10-23 →
Local AI / HomeLab

NVIDIA PC Model Loading & VRAM Diagnostics: Spotting PCIe Shared-Memory Spillover

Verified 2026-10-23 →
Local AI / HomeLab

Ollama + Open WebUI Operations: Installation, Persistence, Backup, Upgrade & Safe Restore

Verified 2026-10-23 →