The Astonishing Computing Power of a Starlink V3 Satellite: Equivalent to 33 NVIDIA H200 GPUs

Critics of SpaceX’s space-based computing ambitions (StarMind) often question the ability of satellites to dissipate heat. They dismiss StarMind as nothing more than conceptual hype, as illustrated by the recent debate between Elon Musk and a Dutch professor on X (formerly Twitter).

This article offers a different perspective.

From the standpoint of communications low level compute, space-based computing is not an entirely new concept. Starlink satellites already possess substantial computing capabilities. The V3 satellites, in particular, which have already been deployed into orbit by Starship, boast computing power equivalent to an astonishing 33 NVIDIA H200 GPUs.

If thermal management were really such a problem, why haven’t Starlink satellites simply burned up?

In fact, space-based computing satellites can be viewed as the natural result of applying the principles of scaling to Starlink’s existing satellite architecture. No reason to regard this approach as “castle in the air”.

To understand why the computing power of a single Starlink V3 satellite could rival that of 33 Nvidia H200 GPUs, let’s first take a look at the computing power behind modern wireless communications.

1. The Hidden Computing Power Behind Wireless Communications

It’s not just LLM training and inference that require enormous computing resources. Our smartphones also rely on substantial computing power, often delivered through specialized accelerators.

Examples include the NPU (Neural Processing Unit), ISP (Image Signal Processor), and GPGPU (General-Purpose GPU), all of which regularly feature in smartphone performance benchmarks.

But the computing power required for the communication part of the cell phone is even more substantial.

Because users rarely interact directly with the cell phone communications subsystem, the baseband processor responsible for handling wireless communications tends to go unnoticed—except when we complain about poor signal reception or buffering while streaming videos.

The communications baseband, also known as the BPU (Baseband Processing Unit), is typically a heterogeneous multi-core processor. It contrasts with the APU (Application Processing Unit), usually based on ARM architecture, which runs the operating system, user interface, and applications.

The computing capabilities of a BPU are enormous. Measured by the number of operations performed per second, its processing capacity can even exceed that of all the other individual processors in a smartphone.

A significant proportion of the computations performed by baseband processors are handled by dedicated hardware accelerators, such as FFT engines and channel-decoding engines. This approach reduces power consumption, at the expense of some flexibility.

The mathematical operations involved in baseband processing include:

• FFT/IFFT (Fast Fourier Transform / Inverse Fast Fourier Transform): Large numbers of complex multiplications. Many algorithms can leverage FFT operations.
• Digital filtering: Multiply-accumulate operations, which can also be expressed as matrix operations.
• Synchronization: Sliding correlators, which can be treated as a form of digital filtering.
• Channel estimation: Large numbers of complex multiplications.
• MIMO processing: Extensive matrix operations.
• LDPC encoding and decoding: Large numbers of low-precision additions, numerical comparisons, summations, and scaling operations.
• Beamforming: Large numbers of high-speed complex multiplications.

As we can see, many of these operations ultimately boil down to multiplication, addition, and matrix computation, which are also used in AI/LLM training and inference domain.

2. NVIDIA Is Already Using GPUs to Power 5G Base Stations

Unlike smartphones, cellular base stations face relatively fewer power constraints (although telecom operators certainly feel the pain of their electricity bills).

As a result, they can rely on general-purpose processors, such as CPUs and GPGPUs, to perform communications processing, so called OpenRAN, Open-RAN or ORAN.

The Radio Access Network (RAN) for mobile communications has become an important business area for NVIDIA (also formerly an important part of IBM/Intel). Its solution is called Aerial CUDA-Accelerated RAN.

In NVIDIA’s official reference implementation, a GH200 GPU is used to implement the functionality of a 5G base station (RAN). A single GH200 can support up to 20 individual cells.

NVIDIA’s RAN solutions frequently appear in announcements and news involving telecom operators such as T-Mobile and Deutsche Telekom, as well as equipment manufacturers such as Nokia and Fujitsu.

The GH200 can be thought of as a heterogeneous SoC combining an H200 GPU with a multi-core ARM processor. This is conceptually similar to the combination of an application processor and a baseband processor in a smartphone.

The difference is that, in this configuration, the ARM application processor runs the Layer 2 and Layer 3 protocols of the 5G network stack, while the H200 GPU acts as the baseband processor.

According to NVIDIA’s official documentation, the reference 5G cell configuration uses the following parameters:

• Bandwidth: 100 MHz
• Modulation: 256QAM
• Subcarrier spacing: 30 kHz
• MIMO configuration: 4T4R
• Downlink: 4 spatial streams
• Uplink: 2 spatial streams
• Aggregate downlink throughput: 1.5 Gbps
• Uplink throughput: 210 Mbps
• Connected user equipment: 100 UEs
• TDD downlink/uplink configuration: 6:4:4

According to the Nvidia official report, A single GH200 can simultaneously run 20 such 5G cells, with all of them achieving their peak throughput.

Therefore, the aggregate downlink data rate supported by a single GH200 is:

20 × 1.5 Gbps = 30 Gbps

3. One Starlink V3 Satellite Equals 33 GH200 GPUs

Now let’s look at Starlink.

A single Starlink V3 satellite is expected to support an aggregate downlink data rate of 1 Tbps, distributed across approximately 2,000 beams (or cells).

Comparing this with the 30 Gbps supported by a single GH200:

1,000 Gbps ÷ 30 Gbps ≈ 33.3

In other words, from the perspective of aggregate downlink throughput, a single Starlink V3 satellite is equivalent to approximately 33 NVIDIA GH200 GPUs.

Why can we use aggregate data throughput as a proxy for estimating the computing power of a Starlink satellite?

The reason is that Starlink also employs OFDM waveforms and LDPC channel encoding and decoding, similar to those used in 5G systems.

Consequently, the underlying baseband processing operations should be broadly comparable.

In NVIDIA’s reference 5G configuration, the system uses relatively low-order MIMO rather than massive MIMO. Compared with Starlink’s SISO beam (cell) architecture, the computational complexity remains within a polynomial relationship.

As a result, the total computational workload should be approximately proportional to peak data throughput.

This makes aggregate throughput a reasonable basis for estimating the relative computing capabilities of the two systems.

4. The Power Consumption Comparison Also Makes Sense

Here’s another interesting observation.

A Starlink V3 satellite reportedly has a power budget of approximately 20 kW, while an NVIDIA H200 GPU consumes around 700 W.

Let’s do the math:

30 × 700 W = 21,000 W = 21 kW

In other words, the power consumption of a Starlink V3 satellite is also roughly comparable to that of 30 H200 GPUs.

Of course, some of the satellite’s power is consumed by transmitting RF signals, which carry energy away from the satellite and slightly reduce its thermal dissipation burden.

5. Anticipating the Counterarguments

I know some readers will disagree with this analysis. I’ve already anticipated your objections. (Though I also agree that this analysis is way too simple/rough/inaccurate.)

Here are a few points worth discussing.

Objection 1: Communications ASICs Are Not General-Purpose GPUs

The communications ASICs aboard Starlink satellites are inherently more power-efficient than GPGPUs. However, they are also much less flexible and cannot necessarily be used for LLM training and inference.

This objection certainly has merit.

We don’t know exactly how the baseband processing accelerators aboard Starlink satellites are designed, or where their architecture lies on the trade-off spectrum between specialization, power efficiency, and general-purpose programmability.

If their architecture already leans toward greater flexibility—for example, to facilitate firmware patches and future upgrades—then, in theory, running LLM training and inference workloads might be possible.

This would essentially be the reverse of what NVIDIA is doing with GPU-powered communications infrastructure.

NVIDIA maps communications workloads onto GPGPUs. In the reverse scenario, LLM workloads would be mapped onto the ASIC-based hardware of a communications satellite.

At its core, both approaches involve mapping mathematical operations onto heterogeneous hardware computing units.

Whether an efficient mapping strategy can actually be developed remains an open question.

Objection 2: Memory Capacity and Bandwidth

LLM training and inference require enormous memory capacity and bandwidth.

However, memory in communications processing systems is often distributed in relatively small blocks, embedded throughout different stages of the processing pipeline.

Consequently, the memory architecture of a communications processor may not be particularly well suited to LLM workloads.

On the other hand, today’s GPGPU memory architectures are not necessarily optimal for LLM workloads either.

It’s simply that CUDA’s enormous historical head start and ecosystem have made the AI industry largely dependent on GPGPU-based development.

Ultimately, this is still a problem of mapping the mathematical operations required by LLMs onto available hardware computing units.

When we also have to deal with the distributed memory access patterns and orchestration requirements of communications processors, the problem becomes considerably more complicated.

6. The Future of Space-Based Computing

Clearly, achieving an optimal space-based computing architecture will require many more rounds of engineering development and system-level optimization.

Are GPGPUs the optimal solution? For now, they may be a good choice—or perhaps even the only practical choice.

After all, getting something working is more important than achieving perfection.

Will future space-based computing systems continue to rely on GPGPUs? Only time will tell.

Perhaps a more fashionable way of describing the future would be:

An integrated high-efficiency hardware processing architecture combining communications, AI computing, and remote sensing.

Regardless, SpaceX’s ambitions in space-based computing are not simply a conceptual publicity stunt.

They can be understood as an engineering scaling law emerging from years of practical experience with space-based communications computing – Starlink satellites.

Every Starlink satellite is essentially a super-base station capable of supporting hundreds or even thousands of cells, and therefore contains an enormous amount of computing power.

SpaceX’s proposed space-based computing satellites are, in essence, simply another type of Starlink satellite—one that runs non-communications workloads instead of focusing exclusively on communications.

The computing power is already there. The next challenge is figuring out how to scale it and put it to work for AI/LLM.