DGX Spark specs: VRAM, memory bandwidth, TDP, power consumption and CUDA cores

DGX Spark VRAM, memory bandwidth, TDP and CUDA cores

There is no separate DGX Spark VRAM, because its Blackwell GPU and its 20-core Arm CPU share one pool of 128 GB of LPDDR5x unified memory, with a memory bandwidth of 273 GB/s. With the memory settings in NVIDIA’s vLLM playbooks, a model can use roughly 90 to 115 GB of that pool for weights and KV cache. The GB10 chip has a TDP of 140 W, the system runs from a 240 W external power supply, and NVIDIA declares a maximum power consumption of 233.2 W at the wall. The GPU has 6,144 CUDA cores with fifth-generation Tensor Cores, rated at up to 1 PFLOP at FP4 with sparsity.

The figures apply to the 128 GB system, which is also the only memory size of NVIDIA’s Founders Edition. NVIDIA lists a 64 GB configuration as well, which its product page marks “Coming Soon” and reserves for OEM partners (from 23 October 2026, according to NVIDIA’s blog), with the same chip and the same 273 GB/s; our comparison of the DGX Spark 64 GB and 128 GB covers what changes with half the memory.

DGX Spark specs: the full specification sheet

NVIDIA publishes the specifications on the product page, in the datasheet and in the user guide’s hardware overview, which differ in detail; the table names the source of each value.

NVIDIA DGX Spark product page (page updated 7 October 2026), datasheet 5651650 (September 2026) and user guide hardware overview (updated 10 September 2026), all read on 16 October 2026.

Three differences between the sources matter. The product page and the datasheet give the CUDA cores only as “Blackwell Generation”; the count of 6,144 comes from the user guide. The datasheet labels the 240 W row “Power Consumption”, while the product page calls the same figure “Power Supply”, which is the more accurate term, as the power section below shows. For storage, the user guide gives “1 TB or 4 TB”, the product page “Up to 4 TB”, and the datasheet 4 TB, the drive of the Founders Edition.

Why there is no VRAM figure, and how much of 128 GB a model can use

NVIDIA’s announcement of 18 March 2025 describes the design: the GB10 Superchip uses NVLink-C2C “to deliver a CPU+GPU-coherent memory model”. DGX OS, the desktop session, the processes on the Arm cores and the model all draw on the same 128 GB, so the size of a model is limited by that pool, minus what the rest of the system holds.

Two tools that normally answer the VRAM question give no reliable answer on this system. NVIDIA’s known-issues page (updated 10 September 2026) states that nvidia-smi “will display ‘Memory-Usage: Not Supported’ even though per-process GPU memory is listed”, and explains that integrated GPUs “do not have dedicated framebuffer memory”. The same page notes that the memory reported by cudaMemGetInfo “may be smaller than the actual allocatable memory”. On the 128 GB system, testers report that the operating system sees about 119 to 122 GiB.

We found no NVIDIA figure for the memory a model can use, and NVIDIA’s own vLLM settings differ by source. The single-Spark launch script that the vLLM playbook on build.nvidia.com links to (playbook updated 14 September 2026) sets --gpu-memory-utilization 0.7. The vLLM playbook on GitHub (updated 12 June 2026) uses 0.8 on one Spark, 0.4 in one agent recipe, and 0.8 or 0.9 on two; the TensorRT-LLM playbook gives the KV cache 0.80 to 0.9 of the memory left after the weights load. Applied to the whole pool, 0.7 gives about 90 GB, 0.8 gives 102.4 GB and 0.9 gives 115.2 GB. Plan with roughly 102 to 115 GB for weights and KV cache together, and with about 90 GB under NVIDIA’s single-Spark script. These figures are our arithmetic and change when NVIDIA changes a default. Which models fit in it, at which precision, is the subject of our article on what fits on a DGX Spark with 128 GB.

Memory bandwidth: where 273 GB/s comes from and what it limits

The user guide gives the memory as “16 channels (256 bit) LPDDR5X 8533”, together with a 4,266 MHz clock. LPDDR5X transfers data on both clock edges, so the interface runs at 8,533 million transfers per second. Multiplied by 256 bits, or 32 bytes per transfer, that gives 273 GB/s, the figure NVIDIA prints on the product page.

For a buyer, bandwidth and capacity answer different questions. The 128 GB decide which model loads, and the 273 GB/s how fast it generates text, because each new token for one user requires reading the active weights once. For a dense model, the ceiling in tokens per second is the bandwidth divided by the size of the weights. A dense 70B model in FP8 has about 70 GB of weights, so one user cannot exceed about 3.9 tokens per second. LMSYS measured Llama 3.1 70B in FP8 at 2.7 tokens per second with SGLang in October 2025. Our DGX Spark benchmarks collect the measured token rates from NVIDIA and independent testers.

DGX Spark VRAM compared with RTX PRO and H200 NVL cards

Capacity and bandwidth are the two DGX Spark figures that compare directly with a graphics card.

NVIDIA product pages for DGX Spark, RTX PRO 5000, RTX PRO 6000 Blackwell, RTX PRO 6000 Blackwell Server Edition and H200 (H200 NVL figures marked preliminary by NVIDIA), read on 16 October 2026; the RTX PRO 5000 page leaves the 72 GB bandwidth cell empty, and 1,344 GB/s for both sizes is from the NVIDIA datasheet our bandwidth comparison cites; the Max-Q edition has the same memory and bandwidth as the Workstation Edition; the ratios are our arithmetic, each bandwidth divided by 273 GB/s.

One Spark holds more than any single RTX PRO card in the table, and each of those cards has five to seven times its bandwidth, which raises the token generation ceiling by the same factor on a model that fits in the card’s memory. The card’s memory is its own, while the Spark’s 128 GB also hold the operating system. The H200 NVL has a little more capacity than the Spark and 17.6 times the bandwidth, and is built to serve many concurrent users in a server. Our GPU memory bandwidth comparison covers every card we supply, with bus widths and per-pin data rates.

We supply the DGX Spark and every card in this table on one EU contract and invoice. Tell us the model, its precision and how many people use it at once, and we reply within one business day with a configuration and a quote.

DGX Spark TDP and power consumption

NVIDIA’s product page gives the GB10 chip a TDP of 140 W and defines it as the “Thermal Design Power of the GB10 chip, including CPU and GPU”. The user guide adds that “100W is available for other system components (ConnectX-7, Wi-Fi, SSD, USB-C, etc.)”. Together they make up the 240 W of the external power supply, which the Quick Start Guide rates at 48 V and 5 A. In the thread “DGX Spark Power Clarification” on NVIDIA’s developer forum, a post from an NVIDIA account of 31 October 2025 states that “DGX Spark’s peak total system power is 240W”.

What the system draws at the wall is lower. The Quick Start Guide lists energy values under Commission Regulation (EU) No 617/2013 on computers, measured to IEC 62623 and EN 62623 at 230 V and 50 Hz.

NVIDIA DGX Spark product page and user guide hardware overview (USB-C power port); DGX Spark Quick Start Guide (power rating and EU section under Commission Regulation (EU) No 617/2013), read on 16 October 2026.

The DGX Spark max power consumption that NVIDIA declares is 233.2 W, and 240 W is the rating of the supply that has to cover it. The same forum post notes that the wattage nvidia-smi displays “measures only GPU power”, so it does not show the draw of the whole system. NVIDIA’s Quick Start Guide states “Use only the included 240W power adapter with DGX Spark”, which matters when a replacement adapter is ordered.

CUDA cores, Tensor Cores and the 1 PFLOP figure

The GPU in GB10 has 6,144 CUDA cores, fifth-generation Tensor Cores with FP4 support and fourth-generation RT Cores, according to the user guide. NVIDIA’s headline is up to 1 PFLOP at FP4, which the product page prints without its basis, and the user guide also gives up to 1,000 TOPS for inference. The datasheet footnote to that row reads “Theoretical FP4 TOPS using the sparsity feature”, and the user guide states “up to 1 PFLOP (petaFLOP) at FP4 precision with sparsity”. We found no dense FP4 figure for GB10 from NVIDIA. Tensor Core throughput sets the speed of prompt processing, fine-tuning and image generation, which are compute-bound; it does not raise the token generation ceiling that the 273 GB/s set.

Storage, networking and linking two or four Sparks

The Founders Edition has a 4 TB self-encrypting NVMe M.2 drive. For the office network there is one 10 GbE RJ-45 port with Wi-Fi 7 and Bluetooth 5.4. The ConnectX-7 adapter, at 200 Gbps, has two QSFP ports on the back for linking Sparks. NVIDIA’s clustering guide states that it “supports up to three DGX Spark systems connected directly through cables, and up to four systems when using a switch”, and lists three approved direct-attach cables, the Amphenol NJAAKK-N911, its 0.5 m version NJAAKK0006 and the Luxshare LMTQF022-SD-R. NVIDIA’s product page gives up to 400 billion parameters for two linked 128 GB systems and up to 700 billion for four; our article on linking two DGX Spark systems covers the link speed and the measured results.

For a team, four Sparks draw at most 932.8 W at the wall by NVIDIA’s declaration, 4 × 233.2 W, and 152 W at idle. With the switch added, that stays well within one 16 A circuit at 230 V, which carries about 3.7 kW. Two Sparks connect with one approved cable. Four need a switch, and NVIDIA’s guide asks for cables “known to provide at least 200 Gb/s”.

We supply two DGX Spark systems for two developers or four for a team, with manufacturer warranty. Describe the team, the models and how the Sparks should be linked, and we reply within one business day with a configuration and a quote.

What we supply

We supply the NVIDIA DGX Spark Founders Edition, with 128 GB of unified memory and a 4 TB self-encrypting drive, across the EU on one contract and invoice, with manufacturer warranty. We also supply the cards it is compared with here, the RTX PRO 5000 with 72 GB, the RTX PRO 6000 in all three editions and the H200 NVL, on their own or in AI servers built to order. From the model, its precision, the context length and the number of concurrent users, we check whether one Spark, two or four, or a GPU server fits before you order. See the DGX Spark page and our professional GPU catalogue.

FAQ