Explainer

NVIDIA DGX Spark Explained for Indian Buyers

In this article
  1. What the DGX Spark is
  2. Memory and what it allows
  3. How it performs
  4. How it sits against the alternatives
  5. Clustering and networking
  6. Software and fit
  7. NVIDIA's own figures
  8. Power and size
  9. Versions from other makers
  10. Buying in India
  11. What this means when buying
  12. Frequently asked questions
  13. Sources & credits

Short answer

The NVIDIA DGX Spark is a palm-sized desktop built on the GB10 Grace Blackwell chip with 128 GB of unified memory. NVIDIA says it runs models up to 200 billion parameters, but its 273 GB/s bandwidth keeps small-model speed well below an RTX 5090. It suits developers who prototype large models locally.

  • Capacity is the strength: 128 GB of unified memory.
  • Bandwidth of 273 GB/s limits generation speed.
  • Published tests show it leading a Ryzen AI Max+ 395 mini PC on prompt processing and matching it on generation.
  • Several makers sell GB10 versions; India supply runs through partners and resellers.

What the DGX Spark is#

NVIDIA's DGX Spark is a small desktop computer, 150 x 150 x 50.5 mm and 1.2 kg, aimed at AI developers. It runs NVIDIA DGX OS, a Linux system with NVIDIA's AI software preinstalled. The chip is the GB10 Grace Blackwell Superchip, which combines a 20-core Arm processor with a Blackwell GPU that has fifth-generation Tensor Cores.

The processor and graphics share one pool of memory, so the machine does not have the separate graphics memory limit of a desktop card. The same idea appears in Apple chips and in AMD's Ryzen AI Max+ platform, and the unified memory vs VRAM explainer covers how that choice affects speed.

Table
SpecDGX Spark
ChipGB10 Grace Blackwell Superchip
CPU20-core Arm (10 Cortex-X925 + 10 Cortex-A725)
GPUBlackwell, 5th-generation Tensor Cores, 4th-generation RT Cores
Memory128 GB coherent unified LPDDR5x (a 64 GB option also exists)
Memory interface256-bit, 273 GB/s
AI performanceUp to 1 PFLOP FP4 (theoretical, with sparsity)
StorageUp to 4 TB NVMe M.2, self-encrypting
NetworkingConnectX-7 at 200 Gbps, 10 GbE RJ-45, Wi-Fi 7, Bluetooth 5.4
Ports4x USB-C, 1x HDMI 2.1a
Power240 W power supply, GB10 TDP 140 W
Size and weight150 x 150 x 50.5 mm, 1.2 kg

Treat the 1 PFLOP figure carefully. NVIDIA footnotes it as theoretical FP4 using the sparsity feature, so it is not comparable with dense FP16 or FP8 figures from other makers. The what is TOPS explainer shows why the precision behind a headline number matters.

Memory and what it allows#

The headline feature is the 128 GB pool. NVIDIA says a single unit can run inference on models up to 200 billion parameters, and that two units joined by a direct cable can handle up to 400 billion. Lenovo and LMSYS quote 405 billion for two units. These counts assume FP4 weights, so they describe what loads, not how fast it runs.

Largest model size NVIDIA says DGX Spark can run, by number of unitsOne DGX Spark200 B parametersTwo DGX Spark units400 B parameters
Largest model size NVIDIA says DGX Spark can run, by number of units
ItemValue
One DGX Spark200 B parameters
Two DGX Spark units400 B parameters
Largest model size NVIDIA says DGX Spark can run, by number of units

The same pool is also the machine's main limit. LMSYS says the 273 GB/s of LPDDR5x bandwidth is the key bottleneck for inference, and Hardware Corner calls LPDDR5 a choke point for large models against high-end discrete GPUs. LMSYS places the GB10's AI capability roughly between an RTX 5070 and an RTX 5070 Ti.

For comparison, the Mac Studio with M5 Ultra is listed by Apple at 1.2 TB/s, and the MacBook Pro with M5 Max at 614 GB/s. Hardware Corner's test below has an Apple M3 Ultra ahead of the Spark on generation while the Spark leads on prompt processing, which is consistent with bandwidth mattering for generation.

How it performs#

No single test covers every rival, so the results below come from separate sources and are shown in separate blocks. Read each block on its own.

Against a Ryzen AI Max+ 395 mini PC, llama.cpp discussion thread (community-reported). The same tester ran gpt-oss-120B at MXFP4 on a DGX Spark and a GMKtec EVO-X2, using the same llama.cpp build and the same settings. The Spark processed the prompt much faster, while generation was nearly equal.

gpt-oss 120B MXFP4 prompt processing: pp2048 prefill at depth 0 (tokens/s, higher is better)

gpt-oss 120B MXFP4 prompt processing: pp2048 prefill at depth 0NVIDIA DGX Spark (GB10): 1737.2 tokens/sNVIDIA DGX Spark (GB10)1737.2 tokens/sGMKtec EVO-X2 (Ryzen AI Max+ 395): 999.6 tokens/sGMKtec EVO-X2 (Ryzen AI Max+ 395)999.6 tokens/s

2 products. Highest: NVIDIA DGX Spark (GB10), 1737.17 tokens/s. Lowest: GMKtec EVO-X2 (Ryzen AI Max+ 395), 999.59 tokens/s.

gpt-oss 120B MXFP4 prompt processing: pp2048 prefill at depth 0
Producttokens/s
NVIDIA DGX Spark (GB10)1737.17
GMKtec EVO-X2 (Ryzen AI Max+ 395)999.59
Source: llama.cpp GitHub discussion 16578 (community-reported, tester eugr, Oct-Nov 2025) (opens in a new tab). llama-bench -fa 1 -ngl 99 -ub 2048 -p 2048 -n 32 -d 0,4096,8192,16384,32768; gpt-oss 120B MXFP4 GGUF; llama.cpp build 6816 for both; Spark CUDA default (no unified-memory flag); AMD HIP/ROCm. Single community tester; Spark later run on kernel 6.17.1 gave pp2048 1956.03 and tg32 60.57.

About this data: figures are copied from the named source, not measured by us, and the test system and settings are those the source states above. See all benchmark sources and credits.

gpt-oss 120B MXFP4 text generation: tg32 decode at depth 0 (tokens/s, higher is better)

gpt-oss 120B MXFP4 text generation: tg32 decode at depth 0GMKtec EVO-X2 (Ryzen AI Max+ 395): 47.5 tokens/sGMKtec EVO-X2 (Ryzen AI Max+ 395)47.5 tokens/sNVIDIA DGX Spark (GB10): 45.9 tokens/sNVIDIA DGX Spark (GB10)45.9 tokens/s

2 products. Highest: GMKtec EVO-X2 (Ryzen AI Max+ 395), 47.49 tokens/s. Lowest: NVIDIA DGX Spark (GB10), 45.87 tokens/s.

gpt-oss 120B MXFP4 text generation: tg32 decode at depth 0
Producttokens/s
GMKtec EVO-X2 (Ryzen AI Max+ 395)47.49
NVIDIA DGX Spark (GB10)45.87
Source: llama.cpp GitHub discussion 16578 (community-reported, tester eugr, Oct-Nov 2025) (opens in a new tab). llama-bench -fa 1 -ngl 99 -ub 2048 -p 2048 -n 32 -d 0,4096,8192,16384,32768; gpt-oss 120B MXFP4 GGUF; llama.cpp build 6816 for both; Spark CUDA default (no unified-memory flag); AMD HIP/ROCm.

About this data: figures are copied from the named source, not measured by us, and the test system and settings are those the source states above. See all benchmark sources and credits.

Prompt processing was 1,737.17 tokens per second on the Spark against 999.59 on the EVO-X2. Generation was 45.87 against 47.49. The gap on prompt processing widened with longer context: at a 32,768-token depth, the Spark managed 1,231.86 tokens per second against 348.59, and generation was 34.29 against 35.39. The same thread shows the Spark improving on a later software build, reaching 60.57 tokens per second for generation on a newer build and kernel, so launch-week figures understate later performance.

Against Apple, AMD and NVIDIA cards, Hardware Corner. Hardware Corner published the llama.cpp maintainer's launch-week Spark result alongside figures for three other systems. No software versions are given.

gpt-oss 120B MXFP4 text generation: tg32 decode (tokens/s, higher is better)

gpt-oss 120B MXFP4 text generation: tg32 decode3x RTX 3090: 124 tokens/s3x RTX 3090124 tokens/sApple M3 Ultra 256GB (Mac Studio): 70.8 tokens/sApple M3 Ultra 256GB (Mac Studio)70.8 tokens/sNVIDIA DGX Spark (GB10): 38.5 tokens/sNVIDIA DGX Spark (GB10)38.5 tokens/sAMD Strix Halo (Ryzen AI Max+ 395): 34.1 tokens/sAMD Strix Halo (Ryzen AI Max+ 395)34.1 tokens/s

4 products. Highest: 3x RTX 3090, 124.03 tokens/s. Lowest: AMD Strix Halo (Ryzen AI Max+ 395), 34.13 tokens/s.

gpt-oss 120B MXFP4 text generation: tg32 decode
Producttokens/s
3x RTX 3090124.03
Apple M3 Ultra 256GB (Mac Studio)70.79
NVIDIA DGX Spark (GB10)38.55
AMD Strix Halo (Ryzen AI Max+ 395)34.13
Source: Hardware Corner, First DGX Spark LLM benchmarks, 15 Oct 2025 (numbers from llama.cpp maintainer) (opens in a new tab). llama.cpp, gpt-oss 120B MXFP4, pp2048 and tg32; build not stated; launch-week build. Hardware Corner published the maintainer's Spark result alongside three other systems; no software versions given..

About this data: figures are copied from the named source, not measured by us, and the test system and settings are those the source states above. See all benchmark sources and credits.

Generation was 38.55 tokens per second on the Spark, against 34.13 on a Strix Halo machine, 70.79 on an Apple M3 Ultra with 256 GB, and 124.03 on three RTX 3090 cards. Prompt processing told a different story: 1,723.07 for the Spark, 339.87 for Strix Halo and 863.73 for the M3 Ultra.

Against desktop RTX cards, LMSYS. On gpt-oss-20b with Ollama, a model that fits on every card tested, the Spark generated 49.7 tokens per second against 205 for an RTX 5090 Founders Edition and 215 for an RTX PRO 6000 Blackwell.

GPT-OSS 20B MXFP4 (fits in every card tested): decode (tokens/s, higher is better)

GPT-OSS 20B MXFP4 (fits in every card tested): decodeNVIDIA RTX PRO 6000 Blackwell Workstation Edition 96GB: 215 tokens/sNVIDIA RTX PRO 6000 Blackwell Workstation Edition 9…215 tokens/sNVIDIA GeForce RTX 5090 Founders Edition 32GB GDDR7: 205 tokens/sNVIDIA GeForce RTX 5090 Founders Edition 32GB GDDR7205 tokens/sNVIDIA DGX Spark GB10 128GB LPDDR5x (273 GB/s): 49.7 tokens/sNVIDIA DGX Spark GB10 128GB LPDDR5x (273 GB/s)49.7 tokens/s

3 products. Highest: NVIDIA RTX PRO 6000 Blackwell Workstation Edition 96GB, 215 tokens/s. Lowest: NVIDIA DGX Spark GB10 128GB LPDDR5x (273 GB/s), 49.7 tokens/s.

GPT-OSS 20B MXFP4 (fits in every card tested): decode
Producttokens/s
NVIDIA RTX PRO 6000 Blackwell Workstation Edition 96GB215
NVIDIA GeForce RTX 5090 Founders Edition 32GB GDDR7205
NVIDIA DGX Spark GB10 128GB LPDDR5x (273 GB/s)49.7
Source: LMSYS Org, NVIDIA DGX Spark In-Depth Review (13 Oct 2025) (opens in a new tab). Ollama, batch 1; review says results may date as software improves. Numbers as stated in LMSYS text. Model fits in memory on all three..

About this data: figures are copied from the named source, not measured by us, and the test system and settings are those the source states above. See all benchmark sources and credits.

Prompt processing showed the same gap, at 2,053 against 8,519 and 10,108. When a model fits in a card's memory, the card wins on speed. The Spark's case rests on models that do not fit.

LMSYS also tested larger and batched work. Llama 3.1 70B at FP8 on SGLang gave 803 tokens per second for prompt processing and only 2.7 for generation. Batching helps the Spark: Llama 3.1 8B generation rose from 20.5 tokens per second at batch 1 to 368 at batch 32, and speculative decoding gave up to about twice the speed. LMSYS concludes the machine suits prototyping and small-model batching.

How it sits against the alternatives#

The Spark competes with three other designs for local models. The table shows what each maker or reviewer lists for memory, so the trade-offs are easy to see.

Table
OptionMemoryBandwidthWhere it fits
DGX Spark (GB10)128 GB unified273 GB/sLarge models, NVIDIA software stack
Ryzen AI Max+ 395 mini PCUp to 128 GB shared256 GB/sLarge models on x86, Windows or Linux
Mac Studio (M5 Ultra)Up to 512 GB unified1.2 TB/sLargest pools and the highest bandwidth (1.2 TB/s)
Desktop RTX 509032 GB GDDR7, 512-bitNot stated on NVIDIA India pageModels that fit, fastest speed

Tom's Hardware gives the Ryzen AI Max+ 395 a bandwidth of 256 GB/s. The Spark's edge over the AMD machine is NVIDIA's software and its stronger prompt processing in the tests above. The AMD machine is a normal x86 PC that runs Windows or Linux. The Mac Studio offers the biggest pool, and a desktop card wins whenever the model fits. The mini PC picks for local AI cover the AMD and Apple options in more detail.

Clustering and networking#

The ConnectX-7 network chip at 200 Gbps is unusual in a desk-sized machine. NVIDIA says two units connect directly with a cable and no switch, pooling 256 GB for models up to 400 billion parameters. Dell's India page lists two 200G QSFP ports on its Pro Max with GB10, and Lenovo lists two QSFP ConnectX-7 ports.

Tom's Hardware used a two-unit cluster in a later test of a very large model that would not fit on a single 128 GB machine. It found the cluster, using a different inference engine, much faster at prompt processing than a Mac running llama.cpp. The finding is a reminder that software choice can outweigh the hardware.

Software and fit#

The Spark ships with NVIDIA's own software environment. Because it is an Arm machine running Linux, tools built for NVIDIA GPUs and CUDA work, but programs for x86 Windows do not run natively. It is a development box, not a general-purpose PC. Those who want Windows applications or a familiar desktop are better served by other machines.

NVIDIA's own figures#

NVIDIA's technical blog publishes its own inference results for a single Spark, measured with 2,048 input tokens and 128 output tokens at batch 1. These are maker claims, and software versions are not stated. gpt-oss-120B at MXFP4 on llama.cpp reached 1,725.47 tokens per second for prompt processing and 55.37 for generation. gpt-oss-20B reached 3,670.42 and 82.74. Llama 3.1 8B at NVFP4 on TensorRT-LLM reached 10,256.9 and 38.65, and Qwen3 14B at NVFP4 reached 5,928.95 and 22.71.

Two points stand out. First, the rows use different engines and precisions, and gpt-oss-120B is a mixture-of-experts model that reads only part of its weights per token, so the figures do not fall in line with model size. Second, NVIDIA's 120B figure of 55.37 sits above the 38.55 and 45.87 that outside testers measured on launch-week builds, and close to the 60.57 the llama.cpp thread later recorded on a newer build. Software updates matter on this platform, so results from different dates are not directly comparable.

Power and size#

NVIDIA lists a 240 W power supply and a GB10 TDP of 140 W, in a 150 x 150 x 50.5 mm case. That is far below the 1000 W NVIDIA lists as the required system power for a desktop with an RTX 5090. The comparison is not like-for-like, since one is an adapter rating and the other a whole-system requirement, but it shows why the Spark fits on a shelf and needs no large power supply. ASUS lists one USB-C port with a 180 W power input on the Ascent GX10.

Versions from other makers#

Several companies sell GB10 systems with the same chip and memory.

Table
Maker and modelMaker figuresIndia note
ASUS Ascent GX10GB10, 64 GB or 128 GB LPDDR5x, ConnectX-7, 10 GbEA launch in India was reported by My Mobile India
Lenovo ThinkStation PGX128 GB LPDDR5x at 273 GB/s, 1 TB or 4 TB, 240 WDigit.in reports it is sold in India
Dell Pro Max with GB10128 GB or 64 GB LPDDR5x, 1 TB to 4 TB, two 200G QSFPDell India's page directs buyers to Dell or a partner
Gigabyte AI TOP ATOM128 GB LPDDR5x at 273 GB/s, 4 TB M.2Digit.in reports it is sold in India
MSI EdgeXpertGB10 platformANT PC, an Indian integrator, says it delivered the first unit

TechRadar has reviewed Acer's Veriton GN100, and HP sells the ZGX Nano G1n. Check India availability with the seller. The platforms share the same chip, so differences come down to storage options, ports, cooling and warranty. Check each maker's page for the exact configuration and support terms in India before you choose.

Buying in India#

NVIDIA's page lists no price and directs buyers to its marketplace. Digit.in reports that the DGX Spark and several partner versions are sold in India, and BIS Infotech names PNY, RP tech India and TD SYNNEX among India partners. Stock and configuration vary, so confirm the exact memory and storage with the seller and ask how warranty service works locally. Prices change, so check current prices before buying.

What this means when buying#

  • AI developers who want to run 70B to 200B-class models locally and test them with NVIDIA's stack.
  • Teams that already work with NVIDIA's software and want a local box that runs the same tools.
  • Researchers who value two-unit clustering for very large models.

It is a weaker fit for buyers who want the fastest answers from small models, where a discrete RTX card leads. See best GPU for local LLM for that route, best mini PC for local AI for compact alternatives, and the AI workstations and PCs hub for the full range. The AI laptop vs desktop explainer covers the portable side. Before ordering any GB10 machine, confirm four things with the seller: the exact memory and storage configuration, whether it is a single unit or a two-unit bundle, the warranty terms in India, and who handles repairs. The platform is identical across makers, so these practical points are where versions differ.

Frequently asked questions#

What is the NVIDIA DGX Spark?

It is a small desktop AI computer from NVIDIA built around the GB10 Grace Blackwell chip with 128 GB of unified memory. It runs NVIDIA DGX OS and targets developers who run and prototype large models locally.

How much memory does the DGX Spark have?

NVIDIA lists 128 GB of coherent unified LPDDR5x on a 256-bit interface at 273 GB/s, with a 64 GB option also shown. The large pool lets it hold models that would not fit on a normal graphics card.

How large a model can it run?

NVIDIA says a single unit can run inference on models up to 200 billion parameters, and two units joined by a cable up to 400 billion. These figures describe what loads, not speed.

Is the DGX Spark faster than an RTX 5090?

No, not when a model fits in the card. LMSYS measured 49.7 tokens per second for generation on the Spark against 205 on an RTX 5090 for gpt-oss-20b.

How does it compare with a Ryzen AI Max+ 395 mini PC?

In one same-tester llama.cpp thread, the Spark led prompt processing on gpt-oss-120B and was roughly level on generation. See the performance section above for the figures.

Is the DGX Spark sold in India?

Digit.in reports that it and several partner versions are sold in India through partners and resellers. NVIDIA's page lists no price, so confirm details with the seller.

Can it run Windows software?

It is an Arm machine running NVIDIA DGX OS, a Linux system. Programs written for x86 Windows do not run natively. Most AI tools have Linux builds, so check that your software supports Arm Linux before buying.

Is 1 PFLOP a fair speed measure?

It is NVIDIA's theoretical figure at FP4 using sparsity. It is not comparable with dense FP16 or FP8 numbers, and measured speeds in the sources above are far lower.

Sources & credits#

We have not tested these products ourselves. Specs come from the makers; benchmark figures come from the sources we name.

Facts and figures in this article come from these sources. Articles are compiled from the makers' published specifications and the benchmarks named above, and checked by the NextAIGrid Editorial Team.

Disclaimer

This article is for information only. Specifications and availability change, so check details on the maker's page before buying. Found an error? Tell us.

How was this explainer put together?

It draws on maker specification pages and published third-party benchmarks, named in the article. We have not tested these products ourselves. Specs come from the makers; benchmark figures come from the sources we name. Specifications and sources were last checked on 17 November 2026. Read how we research and our editorial policy.

Spotted an error? Tell us. Confirmed fixes are listed on the corrections page.

What questions do readers ask next?

Back to top