Dwarkesh Patel Podcast
Episode overview

Reiner Pope – Chip design from the bottom up

Dwarkesh Patel Podcast · 22 May 2026 · 5 Egleze moments
Reiner Pope – Chip design from the bottom up ▶

What happened in this episode?

On this episode of the Dwarkesh Podcast, host Dwarkesh Patel was joined by Reiner Pope, CEO of AI chip startup Maddx, for a technical deep-dive into how AI accelerators work at the transistor level. Pope, whose company Patel has invested in, walked through the fundamental building blocks of chip design from logic gates up to full production chips, explaining why low-precision arithmetic provides quadratic rather than linear speedups and why data movement costs typically dwarf actual computation costs. A central revelation was that in traditional CPU and GPU architectures, roughly 87% of circuit area and energy goes to moving data between register files and logic units rather than performing calculations, which motivated the introduction of systolic arrays and tensor cores. Pope explained that NVIDIA recently revised their performance specifications to better reflect the quadratic scaling with bit precision, now advertising 3x speedup for FP4 versus FP8 in their B300 generation chips, though he noted this still understates the theoretical 4x advantage. The conversation covered why FPGAs provide deterministic latency for applications like high-frequency trading despite being 10x less efficient than ASICs, how clock cycles work and what determines them, and the fundamental architectural differences between GPUs and TPUs. Pope offered a novel framing that GPUs are essentially arrays of tiny TPUs, with the key trade-off being that GPUs enable higher data movement bandwidth between vector and matrix units through massive parallelism, while TPUs use fewer large units that better amortize register file costs. Toward the end, Pope disclosed that Maddx is developing split-table systolic arrays designed to capture advantages of both GPU and TPU architectures while avoiding their respective bottlenecks.

What are the key points?

Which topics are covered?

How is this page sourced?

Programme
Dwarkesh Patel Podcast
Published
22 May 2026
Original episode
Open the original source ↗

Egleze verifies the speaker, source and timestamp for each linked moment. Unless a moment is explicitly marked independently verified, statements remain attributed claims by the named speaker rather than established facts.

Watch original episode More from Dwarkesh Patel Podcast

5 moments from this episode

Source-linked · editorially selected
01
AI & Tech

NVIDIA Changed Performance Specs After Acknowledging Quadratic Bit Width Scaling

at 15:00 · Attributed speaker claim
NVIDIA made a change. Historically, up until B100 or B200, every time you have the bit precision, you double the flop count. That ratio is actually slightly wrong. It should be an even bigger, you should get an even bigger speedup than than you might otherwise think. NVIDIA's product specs have sort of started acknowledging that in B300 and beyond where the FP4 is 3 times faster than the FP8. Though it should be 4x.

Pope revealed that NVIDIA recently revised their published performance specifications to reflect the quadratic area scaling with bit precision. Previously, NVIDIA claimed 2x speedup when halving precision, but now advertises 3x for FP4 vs FP8 starting with B300 chips. Pope noted even this understates the theoretical advantage, which should be closer to 4x based on circuit area fundamentals.

02
AI & Tech

Data Movement Costs Dominate Logic Operations by Factor of 7 to 1

at 24:22 · Attributed speaker claim
So this circuit I've described here, almost all of the cost, like 7/8 of the cost is in the reading and writing the register file. And only a tiny fraction of the cost is in the logic unit itself.

Pope demonstrated through circuit analysis that in traditional CPU and GPU cores, 87% of the transistor area and energy goes to moving data between registers and compute units, with only 13% performing actual calculations. This fundamental inefficiency was the primary motivation for introducing specialized tensor cores and systolic arrays in modern AI accelerators.

03
AI & Tech

FPGA Energy Efficiency Not as Superior to ASICs as Commonly Assumed

at 1:15:08 · Attributed speaker claim
If you run a chip much slower and you only clock it once every 1,000 clock cycles or something, you will have 1,000 times fewer transitions. It'll be about 1,000 times less energy consumption, but not a substantial advantage in energy efficiency.

Reiner Pope, CEO of Maddx, challenged the common assumption that FPGAs achieve superior energy efficiency through slower clock cycles. While FPGAs do consume less total energy when clocked slowly, Pope explained that this doesn't translate to efficiency gains because chips consume energy primarily when toggling bits, not while idle. This technical insight contradicts conventional wisdom in hardware design circles.

04
AI & Tech

GPUs Are Arrays of Tiny TPUs with Different Trade-offs

at 1:17:16 · Attributed speaker claim
At a very high-level point of view, the GPU has a lot of tiny, tiny TPUs tiled across the whole chip. You can sort of think of scaling this thing down into a really tiny unit with a smaller matrix unit, smaller vector unit. And that is sort of what an SM is.

Pope provided a novel architectural comparison showing that GPU streaming multiprocessors are essentially miniature TPUs. The key difference is that GPUs use many small units enabling higher bandwidth between vector and matrix operations through parallel wiring, while TPUs use fewer large units that amortize register file costs better but create data movement bottlenecks.

05
AI & Tech

Maddx Developing Split-table Systolic Arrays to Bridge GPU-TPU Architecture Gap

at 1:20:05 · Attributed speaker claim
We've talked publicly about something which we call a susplittable systolic array, which is sort of, in some sense, you can think of as like big systolic arrays that can be small systolic arrays as well.

Pope disclosed that his company Maddx is developing what they call split-table systolic arrays, designed to combine the flexibility of GPU-style small compute units with the efficiency of TPU-style large arrays. This architectural innovation aims to capture advantages of both approaches while avoiding their respective trade-offs around data movement and area efficiency.