GPU and Server Configuration for Enterprise Local Drawing AI — 20-User Reference Cost

この文書の適用条件

そのまま実行できるコマンド

# Runnable model sizes by GPU VRAM (quantization: FP16 / INT4 approximate)

VRAM 24GB (RTX 5090 single)
  FP16:   7B–13B (no quantization needed)
  INT4:   14B–30B (loadable with quantization)
  Use:    Dev PoC, personal experiments, simple drawing recognition

VRAM 48GB (RTX 6000 Ada single)
  FP16:   13B–30B (no quantization needed)
  INT4:   30B–70B (loadable with quantization)
  Use:    Extended dev PoC, simple CAD drawing recognition

VRAM 96GB (RTX PRO 6000 Blackwell single)
  FP16:   30B–50B (no quantization needed)
  INT4:   50B–110B (loadable with quantization)
  Use:    Enterprise small-scale (1–5 users), Qwen3-VL-32B single

VRAM 192GB (B200 single)
  FP16:   70B–100B (no quantization needed)
  INT4:   100B–400B+ (loadable with quantization)
  Use:    Large-scale, multiple VLM concurrent, fine-tuning

Important: Model size is not half of VRAM requirement.
Example: Qwen3-VL-32B requires ~64GB at FP16 (32B parameters × 2bytes).
Quantization (INT4) reduces this to ~16GB, but with a precision/speed tradeoff.

# GPU count calculation for 20 enterprise users (example)

【Conditions】
- Model: Qwen3-VL-32B (FP16: 64GB VRAM)
- Concurrent users: 20
- Max concurrent requests: 5
- VRAM per request: ~16GB (INT4 quantized)
- Response time target: under 30 seconds
- GPU VRAM per unit: 96GB (RTX PRO 6000 Blackwell)

【Calculation】
1) Concurrent requests per GPU
   = 96GB / 16GB = 6 parallel

2) GPUs needed for 20 concurrent users
   = ceil(5 / 6) = 1 ... theoretically 1 GPU suffices

3) In practice, add margin for:
   - User growth
   - Failover capacity
   - Model update switching time
   → Recommended: 2 GPUs (replica A / replica B)

【Key point】
"20 users = 2 GPUs" is not universal.
Actual concurrent requests, drawing page count, resolution, and model size change the result.
PoC measurement is the only reliable method.

結果の読み方

こういう状況で使います

• Want to build local drawing AI but unsure of GPU selection criteria
• Cannot decide between RTX 5090 and RTX PRO 6000
• Want to know if Mac M4 or NVIDIA is better for drawing AI
• Do not know how many GPUs are needed for 20-user drawing AI
• Cannot determine if B200 is really necessary
• Want to know approximate cost for enterprise drawing AI setup
• Unclear on the role split between VLM and CAD Engine

考えられる原因（可能性の高い順）

1. 01

   GPU selection is made "by price"

   Choosing the most powerful GPU is not the right approach. Required VRAM depends on model size, quantization rate, and concurrent requests — meaning the most effective GPU varies by these factors.

2. 02

   The incorrect formula "users = GPU count"

   Standard GPU sizing is based on VRAM capacity and model size, not user count. Even in a 20-user environment, one GPU may suffice depending on per-request workload.

3. 03

   VLM and CAD Engine role split is not understood

   The VLM judges "what needs to change," while the CAD Engine generates "precise coordinates and shapes." Expecting the VLM to do both leads to coordinate precision failures.

4. 04

   B200 is being recommended beyond necessity

   B200 is the most powerful GPU, but it is overkill for typical 20-user enterprise workloads. RTX PRO 6000 × 2 is sufficient in most cases and has better cost efficiency.

確認手順

1. 1

   Analyze current drawing workload volume and frequency

   参照のみ

   Analyze daily and monthly drawing processing volume, average pages per drawing, resolution, and format to estimate required VRAM and processing capacity.

2. 2

   Select target model size

   参照のみ

   Reference: 4B–8B VLM for small-scale verification, 8B–32B for dev PoC, Qwen3-VL-32B or larger for production. Target model size determines required VRAM.

3. 3

   Measure actual VRAM consumption in a PoC environment

   低

   Send real concurrent requests with the selected model and measure VRAM usage, response time, and capacity. Measured values — not estimates — are the basis for final decisions.

対応方法

バージョン・環境による違い

これで解決しない場合に確認すること

• Analyze actual drawing workload size distribution

  Analyze average page count, resolution, and format per drawing to derive reference values for required VRAM capacity.

• Measure peak concurrent request volume

  Measure concurrent request peaks in the current (or similar) system to use as input for GPU configuration calculations.

• Verify quantization precision impact

  Confirm whether output precision with INT4 quantization meets business requirements in a PoC. FP16 may be needed for precision-critical use cases.

この文書の根拠と限界

よくある質問

この文書がカバーする質問

• What is the configuration for enterprise drawing AI server?
• Which is better for drawing AI: RTX 5090 or RTX PRO 6000?
• How many GPUs needed for 20-user AI server?
• Can I build drawing AI with Mac M4?
• Is B200 really necessary?

リスク表示の意味

• 参照のみデータと設定を変更しません。
• 低影響は限定的ですが、権限と負荷の確認が必要です。
• 中性能・ロック・コストに影響する可能性があります。
• 高障害・データ損失・復旧作業が発生する可能性があります。
• 専門家レビュー必須本番適用前に別途レビューが必須です。

GIIPの対応範囲

The above is general sizing guidance not tied to specific GPU products. Regarding what GIIP covers: GIIP handles the full operations layer including GPU server design and selection, model deployment and optimization, AI Gateway-based request distribution, GPU utilization and model state monitoring, incident response, model version updates, cost optimization, and capacity expansion decisions. GPU procurement and installation are the customer's responsibility, but GIIP can support the surrounding software layer and operational design.

関連サービス

• GIIP AIインフラ構築支援
• GPUサーバー設計・導入支援
• AIモデル導入・最適化