Build Not yet confirmed elsewhere 1 publisher2 min readPublished Aug 25, 2026, 10:02 PM PDT

Mac Studio M5 Ultra vs DGX Spark: capacity says what fits, bandwidth says what you wait for

The bandwidth gap is 4.4x and the capacity gap is 4x, which is why these two boxes are not really competing. One decides whether a model fits; the other decides whether it is usable.

The Engineer · Build desk

How we use AI Send a correction

Photograph accompanying Mac Studio M5 Ultra vs DGX Spark: capacity says what fits, bandwidth says what you wait for — Photo: apple.com

What happened

NVIDIA's DGX Spark is shipping now at $4,699 with 128GB of coherent LPDDR5x at 273GB/s and the full CUDA stack.
That puts Apple roughly 4.4x ahead on memory bandwidth, 1.2TB/s against 273GB/s.
NVIDIA's own published figure for a Llama 3.3 70B QLoRA fine-tune on DGX Spark peaks at 5,079 training tokens per second.
Reviewers of the first DGX Spark reported a 100W cap and thermal throttling, with John Carmack naming the gap to the petaflop marketing.

Why it matters

decision Memory capacity and bandwidth settle the buy only when the job is serving tokens. For tuning work, the software stack settles it and the memory numbers become decoration.
cost What either box replaces is a metered cloud bill or a used-GPU build the author found unbuildable on consumer VRAM prices and electricity, so the $5,500 is bought against a recurring cost, not a...
contradiction A 240W datasheet envelope and a reported 100W ceiling, 42 percent of it, describe different machines; until that closes, vendor throughput numbers read as peaks rather than sustained rates.
capability For buyers who can wait on a price, half a terabyte of fast unified memory makes higher precision and long context on very large models a desk-side option rather than rented capacity.

Take the configuration both vendors point at and divide. A 200B-parameter model at 4-bit weights is roughly 100GB of memory ^[21], and decode produces one token at a time, with a full pass over the weights for each one ^[3]. That makes the memory bus a hard ceiling: 273GB/s across 100GB is about 2.7 tokens per second on DGX Spark ^[26], while 1.2TB/s across the same weights is about 12 ^[27]. Both figures sit before KV cache traffic and everything else a runtime does, so read them as the rate no benchmark will beat. NVIDIA's advertised support for up to 200 billion parameters ^[12] is a claim about what fits, and it holds. The quoted petaFLOP of FP4 ^[6] is compute for prefill and training, not for the token you are waiting on.

Price per gigabyte then crosses over. Spark is $4,699 for 128GB, or $36.71 per gigabyte ^[16]. The base Mac Studio is $5,499 for 96GB, or $57.28 ^[17], the worst ratio in the comparison. Apple's 256GB step costs $4,000 on top of base ^[9], so that machine lands at $9,499 and $37.10 per gigabyte ^[18], within about one percent of Spark ^[19] with the bandwidth advantage attached. The 512GB tier has no price and does not ship until late October ^[8], which is the part a purchase order cannot act on. Note also what the headline number actually is: 1.2TB/s is a 50 percent step over M3 Ultra's 800GB/s ^[22], so the ceiling moves 50 percent and the arithmetic above moves with it.

One caution on all of it. This comparison comes from a single dev.to analysis whose author states he has used neither machine and worked from Apple's newsroom, NVIDIA's datasheet and third-party DGX Spark reviews ^[13]. Spec sheets are precisely where sustained behaviour hides, which is what the Spark throttling reports were about.

So the jobs separate. If the work is generating tokens from a large model at a rate a person will sit through, the bus decides and there is one machine here. If the work is tuning, or anything that assumes CUDA, the spec sheet stops mattering: the inference servers people actually deploy were born on CUDA, and MLX remains Apple-only and slower to get day-one support for new open models ^[5]. Both boxes run off a standard wall outlet ^[14], so the constraint that used to settle this argument at home is gone, and the two memory numbers are what is left.

What to watch

Apple's price for the 512GB tier, which decides whether the dollars-per-gigabyte parity found at 256GB survives above it.
Whether shipping DGX Spark units hold the quoted 240W envelope under sustained inference, or repeat the first batch's throttling.
Independent decode benchmarks on M5 Ultra with 200B-class models, to see how far real tokens per second falls below the bandwidth ceiling.

Clarity's read

What the record supports and how the coverage leans. The claims behind it follow.

Reality

Evidence38
Adoption32
Hype gap+25
Incentives42
Confidence35

Why these scores

Claim ledger

Ranked by verification strength, evidence, and original report placement.

[1]
Apple announced a new Mac Studio with M5 Ultra offering up to 512GB of unified memory at 1.2TB/s of memory bandwidth, starting at $5,499 with 96GB.
ReportedSupported2 sources— create a free account to open them View cited source
[2]
Mac Studio M5 Ultra starts at $5,499 with 96GB of unified memory.
ReportedSupported2 sources— create a free account to open them View cited source
[3]
In the decode phase of LLM inference the model generates one token at a time and every token requires a full pass over the weights; once the model is loaded, inference is almost entirely a memory-bandwidth problem.
ReportedSupported2 sources— create a free account to open them View cited source
[4]
NVIDIA's published numbers show a Llama 3.3 70B QLoRA fine-tune hitting a peak of 5,079 tokens per second of training throughput on DGX Spark.
ReportedSupported2 sources— create a free account to open them View cited source
[5]
Serious inference servers such as vLLM and TensorRT-LLM were born on CUDA, while Apple's MLX ecosystem is Apple-only, smaller, and slower to receive day-one support for new open models.
ReportedSupported2 sources— create a free account to open them View cited source
[6]
NVIDIA quotes up to 1 petaFLOP of FP4 for DGX Spark.
ReportedSupported2 sources— create a free account to open them View cited source
[7]
512GB fits vastly more than 128GB, or fits big models at higher precision with long context; Apple's pitch is that enormous LLMs run entirely on device.
ReportedSupported2 sources— create a free account to open them View cited source
[8]
The 512GB Mac Studio configuration ships late October and Apple has not priced it.
ReportedSupportedView cited source
[9]
The 256GB memory upgrade on Mac Studio M5 Ultra costs $4,000 on top of the base configuration.
ReportedSupportedView cited source
[10]
NVIDIA DGX Spark has 128GB of coherent LPDDR5x at 273GB/s, costs $4,699, is shipping now, and runs the full CUDA stack.
ReportedSupportedView cited source
[11]
Apple's machine has roughly 4.4x the memory bandwidth of DGX Spark: 1.2TB/s versus 273GB/s.
ReportedSupportedView cited source
[12]
128GB fits a 200B-parameter model at 4-bit quantization, which is what NVIDIA advertises for DGX Spark: support for up to 200 billion parameter models.
ReportedSupportedView cited source
[13]
The author of the comparison discloses he has used neither machine and worked from Apple's newsroom, NVIDIA's datasheet and independent DGX Spark reviews.
ReportedSupportedView cited source
[14]
Both machines are desk-side boxes that run off a standard wall outlet, and neither needs a circuit upgrade.
ReportedSupportedView cited source
[15]
The author priced out running a 200B-parameter model at home without cloud tokens and found the used-GPU route dead-ending, because a card with enough VRAM either does not exist at consumer prices or breaks the budget on electricity.
ReportedSupportedView cited source
[16]
DGX Spark costs $36.71 per gigabyte of memory.
DerivedSupported2 sources— create a free account to open them
[17]
The base Mac Studio M5 Ultra costs $57.28 per gigabyte of memory.
DerivedSupported2 sources— create a free account to open them
[18]
A 256GB Mac Studio M5 Ultra costs $9,499, or $37.10 per gigabyte.
DerivedSupported2 sources— create a free account to open them
[19]
At 256GB the Mac's price per gigabyte is within about one percent of DGX Spark's.
DerivedSupported2 sources— create a free account to open them
[20]
The Mac Studio's top memory configuration holds 4x the capacity of DGX Spark.
DerivedSupported2 sources— create a free account to open them
[21]
A 200B-parameter model at 4-bit weights is about 100GB of weights in memory.
DerivedSupported1 source— create a free account to open them
[22]
M5 Ultra at 1.2TB/s raises the bandwidth ceiling by 50 percent over M3 Ultra's 800GB/s.
ReportedContested2 sources— create a free account to open them View cited source
[23]
DGX Spark is quoted at roughly 240W maximum power draw.
ReportedContested2 sources— create a free account to open them View cited source
[24]
The original DGX Spark launched to complaints that it capped at 100W and thermally throttled under sustained load, with John Carmack publicly calling out the gap between the petaflop marketing and shipping performance.
ReportedContested2 sources— create a free account to open them View cited source
[25]
The reported 100W cap on early DGX Spark units is about 42 percent of the quoted 240W maximum.
DerivedContested2 sources— create a free account to open them
[26]
DGX Spark's theoretical decode ceiling on a 4-bit 200B model is about 2.7 tokens per second.
DerivedInsufficient2 sources— create a free account to open them
[27]
Mac Studio M5 Ultra's theoretical decode ceiling on the same 4-bit 200B model is about 12 tokens per second.
DerivedInsufficient2 sources— create a free account to open them

Sources

1 independent publisher whose own reporting we read for this story.

dev.to
1 article · August 25, 2026
Mac Studio M5 Ultra vs NVIDIA DGX Spark: The $5,500 Local AI Bet

Ask Clarity

Have a question this piece left open? The answer comes from its reporting and claim ledger — nothing else.

Clarity analysis

How the coverage leans and what the record actually supports. Read the methodology

Perspective Coverage

How much of the coverage speaks to builders, operators and investors. It measures emphasis, not whether the story is right.

Builder

For a builder the story reduces to two numbers and one trade: 512GB at 1.2TB/s versus 128GB at 273GB/s. Capacity decides whether a 4-bit 200B model (about 100GB of weights) fits at all; bandwidth decides how many tokens per second you get out of it once it does. The article's own arithmetic puts theoretical decode ceilings around 2.7 tok/s for DGX Spark and 12 tok/s for M5 Ultra, but those are ceilings, not measurements, and the cited third-party DGX numbers (38.6 tok/s generation on GPT-OSS 120B, 5.4 tok/s on Llama 3.3 70B) land far below naive bandwidth math. The counterweight is toolchain: vLLM and TensorRT-LLM grew up on CUDA, NVIDIA has a published QLoRA fine-tune path at 5,079 tok/s training throughput, and MLX remains Apple-only and slower to day-one support new open models. A builder buying for inference of very large quantized models leans Apple; a builder buying for fine-tuning and stack breadth leans DGX Spark.

Reality

Evidence, adoption, hype gap, incentives and confidence, each read independently from the current record on the scale shown beneath it. How these are measured.

Evidence38: 0100
Vendor specs plus third-party DGX benchmarks, no hands-on and no primary links
The cluster contains exactly one source, a developer blog post whose author discloses he has used neither machine. Hard numbers on the DGX Spark side are traceable to NVIDIA's datasheet and to named independent reviews (Hardware Corner, Level1Techs), which is genuine third-party grounding. The Apple side rests entirely on relayed newsroom specs plus bandwidth arithmetic, with no reviewer units, no measured decode figures, and an internal inconsistency in the M3 Ultra baseline (800GB/s vs 819GB/s). The reasoning about bandwidth-bound decode is sound and self-consistent, but nothing in the cluster independently verifies the M5 Ultra configuration or price.
Adoption32: 0100
One box shipping and benchmarked, the other announced with its key config unpriced
DGX Spark is described as shipping now at a known price with independently benchmarked performance and post-launch field reports, which is real deployment evidence. The Mac Studio M5 Ultra is at announcement stage: the 96GB base has a price, the 256GB step has an upgrade cost, the 512GB configuration that carries the whole capacity argument is unpriced and dated late October, and reviewers have no units. There are no user counts, unit volumes, or organizational deployments in the cluster for either machine.
Hype gap+25: −100 understated0 aligned+100 overstated
Debunks vendor headlines, then leans on its own unmeasured ceilings
The article is itself corrective on vendor hype: it discounts NVIDIA's 'up to 1 petaFLOP' FP4 headline against measured generation rates of 38.6 tok/s and 5.4 tok/s, and flags that Apple's '4.3x faster AI performance' compares against the two-generation-old M3 Ultra and covers prompt processing rather than decode. That pushes the gap toward zero. It stays positive because the comparison's decisive numbers, a 4.4x bandwidth advantage translating 'almost directly' into tokens per second and a ~12 tok/s Mac ceiling, are unmeasured arithmetic on hardware nobody has reviewed, priced against a 512GB configuration Apple has not priced, and because the cited DGX measurements show real decode falling well short of naive bandwidth math.
Incentives42: 0100
Vendor marketing materials as primary inputs, offset by an explicit no-hands-on disclosure
Two of the three input streams are vendor-controlled (Apple newsroom, NVIDIA datasheet), and both vendors have a direct commercial interest in the desk-side local-AI framing the article adopts. Mitigating factors: the author discloses up front that he has used neither machine, names his sources, states his buying lens, and actively discounts both vendors' headline claims using third-party reviews. The cluster discloses no vendor relationship, sponsorship, or affiliate arrangement in either direction, so this score reflects only the sourcing incentives that are visible.
Confidence35: 0100
Single publisher, one-sided measurement, decisive SKU unpriced
Confidence is limited by structure rather than by internal quality. There is one source and one publisher, so no cross-publication corroboration is possible; measured performance exists for only one of the two machines; the Apple configuration that anchors the capacity argument is unpriced and unshipped; and the source contradicts itself on the M3 Ultra bandwidth baseline. The spec-level facts about DGX Spark and the qualitative software-ecosystem contrast are the parts of this story that can be held with reasonable confidence.

Topics and entities

Follow any of these and your For You feed starts watching them — no settings page required.

Topics

AI Workstation HardwareFollow
Sustained Performance vs Peak SpecsFollow
Memory-Bandwidth-Bound DecodeFollow
Cost Per Gigabyte of AI MemoryFollow
CUDA Ecosystem Lock-InFollow
Quantization and Model FootprintFollow
Local LLM Inference Follow