Just noticed that DGX Spark is now available on STPL India @4.61L
One can order on official Nvidia partner in India.
Unlike graphic cards,i’m not sure how demanding it is and how old this news is (availability in India), it can be a useful h/w for people looking to host llm/vllms locally and even fine-tune to an extent on its unified 128GB memory
Yup got listed today itself. Its mainly for devs, not for training or proper inference according to me as the memory bandwidth is too low for running anything at meaningful speeds alongwith that enormous price tag.
That’s unfortunate. I remember this being pretty big news at one point.
a good performance comparison with 5090
Also a quick comparison with flagship consumer grade cards of past 3 generations ( pardon ChatGPT for any discrepancy)
The table below prioritizes manufacturer specifications and official documentation where available; when a field is not explicitly published in those sources, it is marked accordingly.
| Attribute | DGX Spark (GB10 Grace Blackwell system) | RTX 3090 | RTX 4090 | RTX 5090 |
|---|---|---|---|---|
| Product type | Turnkey SFF AI system | Discrete GPU | Discrete GPU | Discrete GPU |
| Architecture | Grace Blackwell (integrated GPU+CPU) | Ampere | Ada Lovelace | Blackwell |
| CUDA cores | 6,144 | 10,496 | 16,384 | 21,760 |
| RT cores | 4th gen (count not listed in user guide) | 82 (2nd gen) | 128 (3rd gen) | 170 (4th gen) |
| Tensor cores | 5th gen (count not listed in user guide) | 328 (3rd gen) | 512 (4th gen) | 680 (5th gen) |
| AI TOPS (marketing metric) | Up to 1,000 TOPS (and up to 1 PFLOP FP4 w/ sparsity) | Not listed on NVIDIA 3090 page | Not listed on NVIDIA 4090 page | 3,352 AI TOPS (listed) |
| Memory capacity | 128GB unified LPDDR5X | 24GB GDDR6X | 24GB GDDR6X | 32GB GDDR7 |
| Memory bus / interface | 256-bit unified (LPDDR5X) | 384-bit | 384-bit | 512-bit |
| Memory bandwidth | 273 GB/s | ~936 GB/s (commonly cited; appears in multiple benchmark/spec compilations) | ~1008 GB/s | 1,792 GB/s |
| PCIe | Not a PCIe card | PCIe 4.0 x16 | PCIe 4.0 x16 | PCIe 5.0 x16 |
| Total graphics power | System-level; GPU is integrated | 350W (TGP) | 450W (TGP) | 575W (TGP) |
| NVLink / SLI | N/A | NVLink supported on RTX 3090-class consumer generation; limited to 2-way | No NVLink connector on RTX 4090 | No NVLink connector indicated; focus is PCIe Gen5 and DLSS 4 |
| Display outputs | 1× HDMI 2.1a | HDMI 2.1 + 3× DP 1.4a | HDMI 2.1a + 3× DP 1.4a | HDMI 2.1b + 3× DP 2.1b |
| Video encode/decode | 1× NVENC + 1× NVDEC | Listed as 1 encoder / 1 decoder generation info | Listed as 2 encoders / 1 decoder generation info | 3× 9th-gen encoders; 2× 6th-gen decoders |
| Form factor / dimensions | 150×150×50.5 mm; 1.2 kg | 313×138 mm (FE listed) | 304×137 mm (FE listed) | 304×137 mm; 2-slot |
| CPU | 20-core Arm (10 Cortex-X925 + 10 Cortex-A725) | Host CPU dependent | Host CPU dependent | Host CPU dependent |
If RTX 5090 is a Ferrari the DGX Spark is an SUV
RTX 5090 can run smaller models blazing fast while DGX Spark can run large models slowly.
