DGX Spark specs
A DGX Spark has 128 GB of unified memory, and that is also its VRAM: there is no separate graphics memory, and the GPU can use the same pool as the CPU. The rest of the core specs: 273 GB/s of memory bandwidth, a 20-core Arm CPU, up to 1 PFLOP of FP4 compute (with sparsity), a 1 TB or 4 TB SSD and a 240 W power supply, all on one NVIDIA GB10 chip. In practice, a 120-billion-parameter model at 4-bit (about 60 GB of weights) fits with room for its KV cache, and 273 GB/s caps how fast one conversation can generate tokens: a dense 70B model at 4-bit tops out near 7.4 tok/s.
Key specs
| Spec | Value | Source |
|---|---|---|
| Memory (VRAM) | 128 GB LPDDR5x unified memory. There is no separate VRAM: the GPU shares this pool with the CPU, so it is all GPU-addressable. About 121.6 GiB is usable for vLLM on GPUwerk nodes. | NVIDIA product page; NVIDIA known issues |
| Memory bandwidth | 273 GB/s (256-bit) | NVIDIA product page; NVIDIA hardware docs |
| SSD | 1 TB or 4 TB NVMe M.2, self-encrypting | NVIDIA hardware docs |
| CPU | 20 Arm cores: 10 Cortex-X925 and 10 Cortex-A725 | NVIDIA product page; NVIDIA hardware docs |
| AI compute | 1 PFLOP FP4, with sparsity | NVIDIA product page; NVIDIA hardware docs |
| Power supply | 240 W external; GB10 TDP 140 W | NVIDIA product page; NVIDIA hardware docs |
Each value was read on the linked NVIDIA page on 24 October 2026. The 121.6 GiB usable figure is the only one measured on GPUwerk systems. NVIDIA also sells a 64 GB version through partners; the rest of this page is about the 128 GB machine, which is the one GPUwerk rents.
To check a specific model and context length against the 128 GB, use the DGX Spark memory calculator.
Full spec table
| Spec | Value | Source | What it means for LLMs |
|---|---|---|---|
| Chip | GB10 Grace Blackwell superchip | NVIDIA product page | One chip for CPU and GPU. Blackwell GPU with 5th-generation Tensor Cores, so FP4 and FP8 formats run natively. |
| CPU | 20 Arm cores: 10 Cortex-X925 and 10 Cortex-A725 | NVIDIA product page, NVIDIA hardware docs | Handles the OS, tokenizing and data loading. Rarely the limit for serving. |
| CUDA cores | 6,144 | NVIDIA hardware docs | Context only. Decode speed on this box tracks memory bandwidth, not core count. |
| Memory | 128 GB LPDDR5x, coherent unified memory; about 121.6 GiB usable | NVIDIA product page; usable figure as vLLM reports it on GPUwerk nodes (vLLM guide) | Capacity: a 120B-parameter mixture-of-experts model fits at 4-bit (about 60 GB of weights) with room for the KV cache. |
| Memory bandwidth | 273 GB/s (256-bit, 4266 MHz) | NVIDIA hardware docs | The decode ceiling. Tokens per second is at most 273 GB/s divided by the bytes of active weights per token. |
| AI compute | Up to 1 PFLOP FP4 (with sparsity) | NVIDIA product page; sparsity per NVIDIA hardware docs | A peak figure that assumes sparsity. Compute matters for prefill, batch serving and fine-tuning; single-chat decode is bound by memory bandwidth. |
| Networking | ConnectX-7 NIC at 200 Gbps; two QSFP ports; 10 GbE RJ-45; Wi-Fi 7 | NVIDIA product page, NVIDIA hardware docs | Links two Sparks into 256 GB; NVIDIA rates that pair at models up to 400B parameters. It is not your internet uplink. |
| Storage | Up to 4 TB NVMe M.2 with self-encryption; NVIDIA's docs list 1 TB or 4 TB, partner machines come in 1, 2 or 4 TB | NVIDIA product page, NVIDIA hardware docs; partner options per the GB10 comparison | A 70B model at 4-bit is about 37 GB on disk, so 1 TB holds a couple of dozen checkpoints of that size. |
| Power supply and TDP | 240 W external supply; GB10 TDP 140 W | NVIDIA hardware docs (supply and TDP); NVIDIA product page | The 240 W is the supply rating, an upper bound for the whole system. NVIDIA publishes no figure for draw under load. |
| Idle power | About 37 W at launch; 25 W with a display and 22 W headless after the January 2026 update | Tom's Hardware, February 12, 2026, reviewer-measured | Measured on a Founders Edition, not by GPUwerk. A Dell Pro Max GB10 in the same test stayed at 35 to 37 W. |
| Dimensions and weight | 150 x 150 x 50.5 mm, 1.2 kg | NVIDIA product page | Founders Edition. Partner cases differ slightly, for example 1.48 kg for the ASUS Ascent GX10 (see the GB10 comparison). |
| Software | DGX OS 7.5.0 (Ubuntu-based); the July 2026 release is the newest entry | NVIDIA release notes | Ships with the NVIDIA driver and CUDA stack. The GPU is sm_121, and not every prebuilt wheel or kernel for other Blackwell chips carries over. |
Checked 2026-10-11 against the NVIDIA pages linked in the table. The usable-memory figure is the only one from GPUwerk systems. NVIDIA's product page says one Spark can "inference with AI models up to 200 billion parameters" and two linked Sparks up to 400B, without naming a precision; at 4-bit, 200B parameters is about 100 GB of weights, which leaves little room for a KV cache in 121.6 GiB. The idle figures come from the Tom's Hardware article of February 12, 2026; its body did not load for me, so I took them from a search summary and the site's power guide.
64 GB model
| Spec | Value | Source |
|---|---|---|
| Memory | 64 GB LPDDR5X, sold only through OEM partners (Acer, ASUS, Dell, Gigabyte, HP and MSI), on sale from 23 October 2026, starting at $4,999. GPUwerk does not rent or host 64 GB machines. | NVIDIA product page (variant exists, OEM-only); partners, date and price from the NVIDIA blog, 2 October 2026 |
The Register describes the 64 GB model as having half the memory and storage of the 128 GB one, and reports the 128 GB price rising at the same time; the price page tracks what both cost now. NVIDIA's spec table gives one figure for both variants, a 256-bit interface at 273 GB/s, and does not list a separate bandwidth for the 64 GB model.
Same chip, other badges
ASUS, Dell, HP, Lenovo, MSI, Gigabyte and Acer all sell GB10 machines. They share the chip and, in the 128 GB version, the memory size and 273 GB/s. Storage, case, cooling and warranty vary. The GB10 machines compared page lines them up.
What the specs mean in tokens per second
Decode speed is roughly 273 GB/s divided by the active weights read per token. GPUwerk measured this on one GB10 node on 2026-09-02 with vLLM 0.27.1: 80.9 tok/s for Qwen3-Coder 30B-A3B (about 3B active parameters) and 6.0 tok/s for the dense Llama 3.3 70B at one request. Both are AWQ 4-bit, with 512 tokens in and 256 out. The mixture-of-experts model is fast because it reads little per token; the dense one reads about 37 GB for every token, which puts its ceiling near 7.4 tok/s. Logs and method are on the benchmarks page. For the datacenter comparison, see DGX Spark vs H100.
Electricity is included in the hourly rental rate. If you own a Spark, the power and electricity cost guide works through it at EU tariffs. What reviewers measured against these figures: memory bandwidth, FP4, FP8 and BF16 compute and CPU performance.
FAQ
How much memory does a DGX Spark have?
128 GB of LPDDR5x unified memory shared by the CPU and GPU, per NVIDIA's hardware documentation. vLLM on GPUwerk nodes reports about 121.6 GiB of it as usable. NVIDIA's product page also lists a 64 GB variant sold only through OEM partners.
How much VRAM does a DGX Spark have?
128 GB, because its unified memory is its VRAM. The GB10 has no dedicated graphics memory: NVIDIA's documentation says the GPU shares system memory with the CPU, so the GPU can use the whole 128 GB pool, less what the operating system and other processes hold. vLLM reports about 121.6 GiB usable on GPUwerk nodes. nvidia-smi shows "Memory-Usage: Not Supported" for this reason; NVIDIA says that is expected.
What is the DGX Spark's memory bandwidth?
273 GB/s, from a 256-bit LPDDR5x interface at 4266 MHz (NVIDIA hardware documentation). That figure caps single-stream decode speed: tokens per second is at most 273 GB/s divided by the bytes of active weights read per token.
How big is the DGX Spark's SSD?
NVIDIA's hardware documentation lists 1 TB or 4 TB of NVMe M.2 storage with self-encryption. Partner machines with the same GB10 chip also come with 1, 2 or 4 TB. GPUwerk rental nodes have 1 TB. A 70B model at 4-bit is about 37 GB on disk.
How much power does it draw?
NVIDIA lists a 240 W external power supply and a 140 W TDP for the GB10. Idle draw is lower: Tom's Hardware measured a Founders Edition at about 37 W at launch and 25 W (22 W with no display) after NVIDIA's January 2026 update; a Dell Pro Max GB10 in the same test stayed at 35 to 37 W. GPUwerk has not measured power itself.
Is the ASUS GX10 the same as the DGX Spark?
It uses the same GB10 chip with 128 GB of memory at 273 GB/s. Storage, case and warranty differ by vendor. NVIDIA sells its own Founders Edition; ASUS, Dell, HP, Lenovo, MSI, Gigabyte and Acer sell or have announced GB10 machines under their own names.
What CPU does it have?
A 20-core Arm CPU in the GB10: 10 Cortex-X925 and 10 Cortex-A725 cores (NVIDIA). It runs the OS, tokenizing and data loading; memory bandwidth, not the CPU, sets single-stream LLM speed.