Four DGX Sparks, 2.1x faster

Note: This is a public session with Bruce Boston and Tim Ellis. Numbers below are what they read on screen or said out loud. The shipping product is turbolapper - AFM (Apple Intelligence on-device; data stays on your Mac). See requirements and Technology.

Four boxes, 2.1 times one

The October 1, 2026 session reads a Hugging Face community handbook on the NVIDIA DGX Spark. On the chart in front of them, four Sparks are 2.1 times faster than a single Spark. Tim calls that a large loss against a clean multiply. He also notes a milder loss at 3.7 on the same chart, and says one for one is the best return. They are not multiplying. Losses show up as you add boxes.

Summary

• Chart: four DGX Sparks at 2.1x one Spark, as read from the handbook.
• Why the group is cautious: memory bandwidth. Tim's estimate puts Apple hardware at least 2x, and he thinks closer to 12 or 15x, the Spark. He says he is forgetting the exact multiplier.
• Power, as they read it: idle around 15 watts, about 100 watts when maxed, and something like 1,200 watts for four at full tilt.
• You get: a reason to do the payoff math before a fourth box, plus Bruce's check on an inflation-and-index claim.
• Next: Watch the session · Get the app

How the cables actually work

Each Spark has two QSFP ports. One box just runs. Two boxes connect with a QSFP cable, which is how Tim's pair is set up. Three can sit in a ring so each talks to the others directly. Four is where he would consider a switch. A ring of four still works, but two boxes that are not neighbors talk through the ones beside them, and he does not call that ideal.

The handbook's Spark Ring software covers switchless pairs and four-spark rings. They describe it as alpha software, pinned to a version. Tim would not be surprised if that software falls apart past four in a ring.

Tim owns two Sparks. The pair cost him $10,000. He says stock is disappearing and street prices have moved: more than $5,000 each, often six or seven thousand, and he has heard of $8,000.

Memory speed is the argument

Bruce does not have Sparks. The group is not planning to build around them. The reason they give is memory speed.

Tim's read of the handbook: Apple hardware memory bandwidth is very fast. He says the multiplier is at least 2x the Spark, and he thinks it is closer to 12 or 15x. He also says he is forgetting the actual number, so treat 12 or 15 as his estimate, not a lab result from this session.

Coming from Apple, people hear "unified memory" and assume the memory is fast. On a Spark, he says, unified means the GPU can reach the memory. It does not mean the memory is fast. The GPU itself does not hold much. If inference has to read the whole model on every token, that read goes through the slow path.

Quantization is the workaround they walk through. A 4-bit quant is about a quarter the size of FP16, so you pull about a quarter of the data per token. Their example is a model that needs 160 GB: the 4-bit version is far smaller, and bandwidth matters less. NVIDIA has a 4-bit format the GPU reads natively. They point at quality charts, including KL divergence, and do not read a single loss number off them.

Local versus cloud, two different reasons

The handbook's cost section, as they tell it, lands on the same headline Bruce has used for Macs: you are better off running locally than renting the same models in the cloud. The reason is different.

Apple hardware is fast because of memory bandwidth, and it draws more power. Sparks are slower, and they draw a small amount of power. In the cloud you are often sharing the machine. The handbook's case for the Spark is that the box is yours.

Bruce's earlier payoff math was for Mac hardware. One Apple configuration took so long to break even that it ran off the chart. Tim's take on four Sparks: it may be a long while, if ever, before that cluster is cheaper than cloud. On the chart they are reading, cloud is the cost-effective cut. Local is the speed you do not share, and you still have to pay for the whole box.

The power notes they read: Sparks idle around 15 watts and run about 100 watts when maxed. Four at full tilt is something like 1,200 watts in the telling, which is why Tim says the total draw must be more than 100 watts a box, and why four at full tilt wants two household circuits. Three at full tilt, he says, fit a normal residential line.

One more software note from Tim: a random AI project of the day is more likely to run on NVIDIA than on Apple right now. His guess is that in about three years they land on the same day.

The inflation claim Bruce checks

Before the Spark handbook, they read a note from VB Barzov (the page they show is bar.com/index.php, money matters), a friend of a friend who had died a couple of weeks earlier. The bet Tim pulls out: if there is inflation, invest in the stock market, and the rest of the bubble talk is a side percent.

The illustration is Zimbabwe in 2007. Hyperinflation. A purchase that doubles by the afternoon. 80% unemployment. The stock market up 12,000% over 12 months, beating inflation and other stock markets in real terms, adjusting for the currency.

Bruce, the economist on the session, does not let the line sit as a rule. His objection is the messiness: a government policy that prints the money also knocks out a stable price for inputs, and that is one reason unemployment follows. The index can rise in the currency that is falling apart. That is not the same thing as a tidy instruction to buy the index whenever inflation shows up.

FAQ

Are four DGX Sparks four times faster?

On the chart they read, four Sparks were 2.1 times faster than one. Tim called that a large loss against a clean multiply, and he noted a milder loss at 3.7 on the same chart.

Why is the group not buying more Sparks?

Memory bandwidth. Tim's estimate is that Apple hardware is at least 2x, and he thinks closer to 12 or 15x, the Spark. He says he is forgetting the exact multiplier. Unified memory on a Spark means the GPU can reach the memory, not that the memory is fast.

What inflation claim does Bruce check?

A note from VB Barzov: if there is inflation, invest in the stock market. The illustration is Zimbabwe in 2007, with 80% unemployment and a stock market up 12,000% over 12 months. Bruce's objection is that printing the money also breaks input prices, so the index line is messier than the chart.

Watch and go further

• Watch: Don't buy a fourth DGX Spark until you watch this
• Related: One week of analyst work · Four DGX Sparks · FRED and MCP · Community
• Product: Get the app · Technology · Pricing