Arm Linux and package compatibility need attention; OS and context share the memory.
Amazon confirms 128GB / 4TB. NVIDIA specifies LPDDR5X; the generic Amazon DDR5 table label is not used.
Amazon.com ASIN B0FWJ16CCH reviewed 24 Sep 2026 using a search copy from the preceding week. CPU, capacity and images follow this listing; chip specifications resolve inconsistent seller fields. Current seller, stock and checkout price can change.
DGX Spark 64GB vs 128GB: which configuration?
NVIDIA announced a 64 GB DGX Spark configuration on , with availability planned for and a stated starting price of $4,999. This page's catalog record and Amazon link remain the exact 128 GB / 4 TB configuration, ASIN B0FWJ16CCH. No matching 64 GB Amazon.com offer was verified for this review.
Official capacity figures and buying evidence checked 2026-10-06. Model-size limits are NVIDIA claims, not compatibility tests.
Configuration
Official memory
NVIDIA model-size claim
Buying evidence
New OEM configuration
64 GB LPDDR5X unified memory
Up to 100B parameters
Planned for October 23, starting at $4,999. No exact Amazon.com listing verified.
Catalog configuration
128 GB LPDDR5X unified memory
Up to 200B parameters
128 GB / 4 TB Amazon.com ASIN B0FWJ16CCH; current price and stock unverified.
Both official configurations use the NVIDIA GB10 Grace Blackwell Superchip, a 20-core Arm CPU, 273 GB/s rated memory bandwidth and 200 Gb/s ConnectX-7 networking. The extra memory is capacity, not evidence that a workload runs twice as fast. Runtime support, quantization, context length and available memory after system use still matter.
A second 128 GB system is a capacity and distributed-compute purchase, not an automatic quality upgrade. NVIDIA says ConnectX-7 can link up to four DGX Spark systems for larger models, faster inference and multi-agent workloads. The same official page advertises the following ceilings; they are platform claims, not Local AI Stats measurements or guarantees for Qwen3.8 Flash Next.
Official DGX Spark cluster capacity claims checked 2026-10-06. Distributed runtime support is required.
Systems
Aggregate unified memory
NVIDIA model-size claim
What the extra node can change
1 × 128 GB
128 GB
Up to 200B parameters
Baseline: validate the exact model, quantization, context, runtime and task quality here first.
2 × 128 GB
256 GB
Up to 400B parameters
More aggregate capacity and supported multi-node compute; not a promised 2× speedup.
4 × 128 GB
512 GB
Up to 700B parameters
Largest advertised desktop cluster; software, networking and workload scaling remain decisive.
If the current model already fits on one system and the blocker is answer quality, tool errors or long-horizon reliability, another node does not change the weights, quantization or agent setup by itself. Buy the second system only after the intended runtime supports the two-node path and a representative task shows that added capacity or throughput solves the actual bottleneck.
GIGABYTE announced its 64 GB AI TOP ATOM on , with availability planned for . Its official specifications list 64 GB systems with 1 TB or 2 TB storage and 128 GB systems with 1 TB, 2 TB or 4 TB. No exact 64 GB Amazon.com listing or US price was verified.
Official GIGABYTE configurations and exact Amazon.com evidence checked 2026-10-05. Availability is not live.
Configuration
Official storage options
Amazon.com evidence
Best fit
AI TOP ATOM 64 GB
1 TB or 2 TB
No exact listing, price or current stock verified.
Choose only after the exact model file, runtime, context and memory overhead fit inside the shared pool.
AI TOP ATOM 128 GB
1 TB, 2 TB, 4 TB
Exact 128 GB / 4 TB listing: ASIN B0FXHFBL54. Cached snapshot was unavailable for purchase; current state is unknown.
More headroom for larger weights, context, cache or concurrent work—not evidence of higher token speed.
Both use the NVIDIA GB10 Grace Blackwell Superchip, 20-core Arm CPU, 273 GB/s rated bandwidth, ConnectX-7 and 10GbE, and NVIDIA DGX OS / Ubuntu Linux. GIGABYTE's advertised capacity and compute figures are manufacturer specifications, not Local AI Stats benchmarks. Arm/Linux package support, quantization, context length and system-reserved memory still matter.
Amazon.com's title and canonical ASIN identify the exact GIGABYTE AI TOP ATOM with 128GB LPDDR5X, 4TB PCIe 5.0 NVMe SSD and NVIDIA DGX OS. The accessible Amazon page snapshot was crawled about three weeks before this review and displayed 'Currently unavailable.' This is not a current stock check; no price, seller or present availability is claimed.
The 128 GB unified-memory pool gives a 89.6 GB initial weight budget using our 30% (minimum 2 GB) reserve. The CPU, GPU and OS share this capacity; software allocation limits still apply.
129 catalog files pass this size check. The largest is Xor 1.1 at 70.214 GB (BF16). This is a catalog weight limit, not a measured maximum practical model size. Context length, runtime and architecture support can rule out a candidate.
Model snapshot: 14 October 2026. The examples below use MMLU-Pro; models without that score are not ranked in this table.
MMLU-Pro memory candidates · GB = 1 billion bytes. Published scores may use different settings or precision.
Local AI Stats has not tested this computer. These third-party results apply to their stated workloads.
Gemma 4 E4B · Q4_K · generation
59.42 tokens/s
Firsthand llama-bench tg128 result on DGX Spark 128GB, CUDA, 13 Apr 2026. 128 generated tokens; model reported as Gemma 4 E4B Q4_K (4.62 GiB). Commit and context depth were not published. One matching computer; not extrapolated to other GB10 systems.
StorageReview’s firsthand Linux vLLM serving test: 256 input / 256 output tokens, batch 64. Aggregate throughput across requests, not single-user chat speed. Quantization and exact software versions are not disclosed in the text; treat this as a reported comparison of two tested systems.
Same StorageReview serving suite: Linux vLLM, 256 input / 256 output tokens, batch 64. Aggregate throughput; quantization and exact software versions are not disclosed in the text. Results apply only to the two tested systems.
$4,699 USD reference. Historical reference: AMD footnote SHO-61 reported DGX Spark at $4,699 in May 2026. Reviewed 22 Sep 2026. This is a dated reference, not a current Amazon price; tax, shipping and bundle contents may differ.
Reference price per installed-memory GB: $36.711. This measures cost per GB, not speed or total running cost.
As an Amazon Associate I earn from qualifying purchases. Check the exact configuration, seller, marketplace and current price before buying.