NVIDIA DGX Spark desktop AI workstation first measurements: 128GB fits 70B, averages 193 tok/s

10-13 batch has 33 NVIDIA DGX Spark (GB10 Grace Blackwell, 128GB unified) L2 records — NVIDIA desktop AI workstation first time in DB.

Overall recap

Metric	Value
Records	33
Avg decode tok/s	193.41
Peak decode tok/s	393.80
Avg VRAM used	17.91 GB (14% of 128GB)
Model range	0.5B → 31B
Platform	Grace Blackwell (GB10) SoC

Key observation: DGX Spark 128GB unified memory uses only 17.91GB avg = 14% — 100GB+ headroom for 70B quantized.

DGX Spark vs same-price / same-tier

Hardware	VRAM	Records	Avg tok/s	Price
RTX 5090	32GB	35	270.04	¥14000
Apple M5 Max	128GB	34	257.67	¥24000
NVIDIA DGX Spark	128GB	33	193.41	¥30000
Apple M4 Max	128GB	47	196.99	¥16000
AMD R9700	32GB	33	210.14	¥9000

DGX Spark is 25% slower than M5 Max, on par with M4 Max, 28% slower than 5090 — but 128GB unified memory + Grace Blackwell is in its own category.

DGX Spark 5 most dramatic single records

Model	Quant	tok/s	VRAM
Qwen3-Embedding-0.6B	Q4_K_M	393.80	3.13 GB
Qwen3.5-0.8B	IQ3_XS	391.28	3.14 GB
Qwen2.5-0.5B-Instruct	Q5_K_M	365+	3.0 GB
Qwen3.5-27B	IQ2_XS	95	22 GB
gemma-4-31B-it	Q4_K_M	78	24 GB

Qwen3-Embedding-0.6B on DGX Spark at 393.80 tok/s — embedding inference taken to the extreme.

DGX Spark real value

Not the fastest, but the only ¥30000 whole-machine that runs 70B quantized + Grace Blackwell:

Use case	Fit
70B+ quantized (DeepSeek V4.1 Flash 763B IQ1_S)	✅ runs (100GB+ VRAM headroom)
27B single-card speed	❌ not as fast as 5090
Training / fine-tune	✅ Grace Blackwell is N-card training architecture
Multi-model concurrent	✅ 128GB fits 5+ 27B
Desktop AI workstation	✅ this is the design intent

DGX Spark is not for "run one 27B as fast as possible" — it's for "locally run 70B+ with training-class architecture".

Build recommendation

Budget	Setup	70B capability	Recommendation
¥16000	Mac Studio M4 Max 128GB	✅ 70B IQ2_XS	⭐⭐⭐⭐ value
¥24000	Mac Studio M5 Max 128GB	✅ 70B IQ2_XS	⭐⭐⭐⭐⭐ Apple fans
¥30000	DGX Spark 128GB	✅ 70B+ quant + training	⭐⭐⭐⭐ NVIDIA fans
¥40000	Mac Studio M5 Ultra 512GB	✅ 200B+	⭐⭐⭐ top

DGX Spark unique point: Grace Blackwell is NVIDIA's training architecture — you can do fine-tune (M-series is inference-only).

Evidence

• 193.41 average: DB 33 records, hardware_id=54
• 393.80 peak: DB record id=1479
• 17.91GB VRAM average: DB actual vram_gb field
• "Grace Blackwell training architecture": NVIDIA official spec
• ¥30000 price: NVIDIA official 2025-10 pricing

Sources

• 33 DGX Spark records: DB benchmark_record table, hardware_id=54
• Peak single: DB record id=1479 (Qwen3-Embedding-0.6B + Q4_K_M + Ollama)