← Research

The GPU Power Problem

NVIDIA’s B200 spends 45% of 1,000 watts moving data, not computing.
The fix is reducing tension, not increasing bandwidth. K/R/E/T analysis of the world’s most important chip.

K IN THIS DOMAIN

K here is compute coupling. FMA units coupled tightly to memory bandwidth. The ratio determines real throughput. Marketing inflates K. We measured it.

A $30,000 chip draws 1,000 watts to generate text one word at a time. 70% of that power does not compute anything. The same framework that predicts protein folding, sleep stages, and iron melting curves explains exactly where every watt goes and what to do about it.

JIM’S OVERSIMPLIFICATION

Marketing says the chip does 20 petaflops. We measured what it actually does when you run real code. The gap between the brochure and the benchmark is where money goes to die. We published both numbers.

NVIDIA sells a $30,000 chip that draws 1,000 watts. The brochure says 20 petaflops. When you actually run the thing — generating text, answering questions, doing what the world bought it for — it uses about 1% of that.

Of course it does. The chip was designed for training — crunching thousands of samples at once. The world is using it for inference — generating one word at a time. Every single word requires reading the entire model from memory. 140 gigabytes of weights, walking from storage to the processor, for every. Single. Token.

45% of that 1,000 watts is spent moving data from one place on the chip to another. Not computing. Moving. It’s like having a 2,000-horsepower engine connected to a garden hose. The engine isn’t the bottleneck. The hose is.

The fix isn’t more horsepower. It’s putting the data closer to where it’s needed. Processing-in-memory — do the math right where the weights live. With known physics, no breakthroughs required, a 1,000-watt chip could become a 40-watt chip at the same speed. That’s a Mac Mini, not a liquid-cooled rack.

But NVIDIA has no incentive to build the efficient chip. A 40-watt inference chip would reduce revenue per job by 99%. This is why Google, Amazon, and Microsoft are building their own. They pay the power bill. They have the incentive.

Same principle as everywhere else: reduce the distance between things that need to talk. Don’t increase bandwidth. Reduce tension. The universe charges for movement, not for thinking.

Honest limit: we have not built a chip. The 45% data movement figure is estimated from architectural analysis, not measured on silicon. NVIDIA engineers are among the best in the world. Our critique is about the mismatch between design and deployment.


The Chip

NVIDIA B200 — Verified Specifications

Architecture: 2 × GB100 dies, 208 billion transistors, TSMC 4NP

Compute: 20 PFLOPS (FP4), 10 PFLOPS (FP8), 5 PFLOPS (FP16)

Memory: 192 GB HBM3e at 8 TB/s

Interconnect: NVLink 5.0 at 1.8 TB/s per GPU, 10 TB/s die-to-die

Power: 1,000W (air), 1,200W (liquid cooled)

Power density: ~500 W/cm² — 50× a kitchen stovetop

This is the most powerful chip ever built. It is also, by our framework’s measure, one of the most inefficient for the workload the world is actually running on it.


The Problem: 99% Idle During Inference

The B200 was designed for training — large batch matrix multiplications where thousands of samples amortize the cost of loading weights. Training is compute-bound. The tensor cores stay busy.

But the world is running inference — generating one token at a time, autoregressively. Every token requires reading the entire model from memory:

70B model × 2 bytes (FP16) = 140 GB of weights 8 TB/s HBM bandwidth ÷ 140 GB = 17.5 ms to read weights Tensor core compute for same operation: ~0.2 ms Utilization at batch=1: 0.2 / 17.5 = 1.1% You paid for 20 PFLOPS. You are using ~200 TFLOPS. The other 98.9% sits idle, drawing power, generating heat.

This is not an engineering failure. It is an architectural mismatch. The chip was designed for one workload and is being used for another.


K/R/E/T Analysis

Our framework measures four quantities in any coupled system. Applied to the B200:

What It MeasuresB200 ValueVerdict
KCoupling bandwidth32.4 TB/s totalMassive
RSynchronization (utilization)1–30% (inference)Critical
EEnergy per useful operation~10&sup9;× LandauerWasteful
TTension (what can’t couple)~0.9997Maximum

Diagnosis: K >> Kc but R << 1/φ. Over-coupled (too much compute relative to what can be fed) and under-synchronized (most of the system isn’t working). The excess coupling manifests as heat.

In plain language: it’s like having a 2,000-horsepower engine connected to a garden hose. The engine isn’t the bottleneck. The hose is.


Where 1,000 Watts Goes

Power breakdown of a B200 GPU during inference. Hover for details.

ComponentPower% of TotalComputes?
Data movement (HBM ↔ cache ↔ compute)~450W45%No
Tensor core compute~300W30%Yes
Control logic (schedulers, decoders)~100W10%No
NVLink SerDes + interconnect~50W5%No
Leakage (transistors off but leaking)~70W7%No
Clock distribution~30W3%No

70% of the power budget does not compute. The dominant cost (45%) is moving data from one place on the chip to another place on the chip. Weights travel from HBM through L2 cache through shared memory through registers to tensor cores… and then the result travels all the way back.


The Memory Hierarchy: 5 Hops to Compute

Data must traverse 5 levels to reach the tensor cores

HBM3e → 192 GB, 8 TB/s → ~5 pJ per bit moved

L2 cache → 126 MB, 21 TB/s → ~1 pJ per bit moved

Shared memory → 256 KB/SM, 40 TB/s aggregate → ~0.1 pJ per bit

TMEM → 256 KB/SM (new in Blackwell) → ~0.05 pJ per bit

Tensor core → actual multiply-add → ~0.01 pJ per bit

Total energy to move one bit from HBM to compute: ~6 pJ
Landauer minimum to process that bit: 0.003 fJ
Gap: 2,000,000×

Every byte of model weights makes this round trip for every single token generated. 100 tokens per second × 140 GB = 14 TB/s of redundant weight reads. The weights don’t change between tokens. They just walk.


The KV Cache Bomb

Beyond the weight problem, long-context inference creates a second crisis: the key-value cache grows with context length × batch size.

ModelContextBatchKV Cache SizeGPUs Required
Llama 70B8K12.5 GB1
Llama 70B128K140 GB1
Llama 70B128K321,280 GB7
Llama 405B128K322,112 GB11

At scale, the KV cache alone requires more memory than the model. Every long-context user eats multiple GPUs just for storage. The compute sits idle while the memory fills.


Prefill vs Decode: Two Different Chips

The workload has two phases with opposite needs

Prefill (processing the input): All tokens processed in parallel. Large matrix multiplications. 90–95% GPU utilization. Compute-bound. This is what the B200 was designed for.

Decode (generating output): One token at a time. Each token reads the entire KV cache and all weights. <10% utilization at small batch. Memory-bound. This is what the world actually runs.

The ideal system would use different hardware for each phase. Instead, the same 1,000W chip handles both, wasting power during decode (which dominates wall-clock time for most applications).


What NVIDIA Got Right

Credit where earned. Three Blackwell innovations address the real problem:

All three reduce tension (T). NVIDIA understands the problem. But they applied these fixes at the edges, not the core. The weights — 99% of the data — still make the dead walk from HBM every token.


What Would Fix It

The K/R/E/T diagnosis says: reduce T, don’t increase K. The chip has enough coupling bandwidth. The problem is that 192 GB of weights are too far from 592 tensor cores.

1. Processing-in-memory

Put multiply-accumulate units inside the HBM stacks. Weights never leave memory. Only partial sums travel across the interposer. Samsung HBM-PIM and SK Hynix AiM demonstrate this today. Eliminates 45% of power (data movement) in one architectural change.

2. Near-threshold voltage

Run transistors at 200 mV instead of 700 mV. Energy scales as V², so this is a 12× reduction in switching energy. Slower per transistor but compensate with wider (more parallel) design. The opposite of NVIDIA’s approach.

3. Workload-matched coupling

LLM inference at batch=1 needs ~14 TFLOPS with 140 GB of near-compute memory. The B200 provides 20,000 TFLOPS with 192 GB of far memory. The compute is 1,400× over-provisioned. A chip designed for inference would have 1/100th the compute die area and 10× the memory proximity.

4. Spectral interconnect topology

NVIDIA connects every GPU to every GPU via NVLink mesh (all-to-all). Most traffic is nearest-neighbor (pipeline stages). Spectral placement — the same Laplacian eigenvector math we use for protein folding and chip layout — minimizes wire length by placing communicating components physically close. Potential 8× traffic reduction.


The Landauer Accounting

Where the Gap LivesCurrentAchievableRecoverable
Transistor voltage (V²)~700 mV~200 mV12×
Data movement~5 pJ/bit~0.1 pJ/bit (PIM)50×
Control overhead10× Landauer2× (simpler ISA)
Leakage2× (2nm, GAA)2.5×
Cooling infrastructure1.2×1.7×
Power delivery1.5× (on-die VRM)
Total recoverable~25,000×

With known physics (no breakthroughs required): a 1,000W chip could become a 40W chip at the same throughput. That’s a Mac Mini, not a liquid-cooled rack. The 40W chip needs 192 GB of processing-in-memory — the key integration that doesn’t exist at scale yet.


The Business Model Problem

NVIDIA has no incentive to build the efficient chip. A 40W inference chip that matches B200 throughput would:

This is why the hyperscalers (Google TPU, Amazon Trainium, Microsoft Maia) are building custom inference chips. They pay the power bill. They have the incentive. NVIDIA sells the shovels.


Honest Limits


The Connection

The same four quantities — K, R, E, T — that separate sleep stages (d = 4.02), predict protein folding (1.5% Rg error), and derive iron melting curves (1–7% across 360 GPa) also diagnose exactly why a GPU wastes 70% of its power.

The principle is always the same: match K to Kc. Too much coupling bandwidth relative to what the workload needs is as wasteful as too little. A drum in tune needs exactly enough tension — not maximum tension.

In every domain we’ve tested, the answer is the same: reduce T, don’t increase K. Make the data closer to where it’s needed. Make each interaction do more work. The universe charges for movement, not for thinking.


45% of 1,000 watts moves data, not computes.
The weights walk 14 TB/s and never change.
A drummer’s framework applied to silicon.
The fix is the same everywhere: reduce T.

GUMPResearch · Support · [email protected]