Feb 29, 2024 · 39m · no-priors
No Priors Ep. 53 | With AMD CTO Mark Papermaster
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this episode of No Priors, hosts Sarah Guo and Elad Gil interview AMD CTO Mark Papermaster to explore AMD's architectural transformation, the engineering behind the Instinct MI300 accelerator, and the future of open-source AI software and hybrid edge-to-cloud computing.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. The hosts hold 18.1% of the talking time here. How this is scored →
The hosts as informed peer 4.7 Guest teaching 3.7 Guest disagreement 0.0 The hosts pushing back 0.1
speaking balance: gold is the hosts, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 15:02 Critique of uncompetitive market stagnation
Mark forcefully asserts that market environments lacking competition are fundamentally bad for everyone and lead directly to industry stagnation.
Hardest push from the hosts ▶ 21:25 Pushing past hype to identify true constraintsElad presses Mark directly on conflicting market narratives, demanding clarity on which supply constraints between packaging, fab capacity, and power are real.
Biggest teaching moment ▶ 25:35 Demystifying post-Moore's Law engineering realityMark breaks down how transistor shrinks no longer provide automatic power or cost reductions, requiring holistic heterogeneous packaging instead.
The host holds their own ▶ 34:03 Synthesizing latency bottlenecks across modern AI systemsSarah clearly demonstrates deep technical domain expertise by articulating how chained models and network latency are forcing compute architectures back onto local devices.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The hosts as informed peer | Guest teaching | Guest disagreement | The hosts pushing back | Why |
|---|---|---|---|---|---|---|
| Announcement: Conviction Embed Accelerator Applications Open | 1 | 1 | 0 | 0 | Segment begins with Sarah Guo's accelerator promotion and standard biographical interview setup. Mark shares an overview of his early semiconductor career across IBM and Apple. | |
| AMD's Market Portfolio and Transformation Journey | 4 | 3 | 0 | 0 | Elad demonstrates knowledge of neural architecture history across CNNs, RNNs, and transformers. Mark details AMD's CPU rebuild under Lisa Su and early heterogeneous compute strategy. | |
| Flagship MI300 Architecture and Target AI Workloads | 5 | 4 | 0 | 0 | Sarah inquires about ROCm and developer ecosystems, referencing portfolio company Lamini. Mark explains AMD's PyTorch founding role and Hugging Face integration. | |
| Open Source Philosophy and the ROCm Platform | 5 | 4 | 0 | 0 | Elad categorizes emerging AI cloud providers and questions whether their advantage persists after GPU supply eases. Mark articulates AMD's open-source philosophy avoiding walled gardens. | |
| Addressing Semiconductor Supply Constraints and Power Limits | 6 | 5 | 0 | 1 | Elad probes specifically into semiconductor bottlenecks, asking whether shortages stem from packaging, TSMC capacity, or datacenter power limits. Mark shows physical MI300 packaging and details multi-chiplet integration. | |
| Architectural Innovation and Holistic Design After Moore's Law | 5 | 5 | 0 | 0 | Sarah asks how AMD designs compute architectures following the slowdown of Moore's Law. Mark explains holistic system-level design across transistor nodes, chiplets, and software stacks. | |
| Supply Chain Resilience and Geographic Diversification | 5 | 4 | 0 | 0 | Sarah raises supply chain concentration and geopolitical vulnerabilities surrounding TSMC. Mark details geographic diversification initiatives across Arizona, Texas, Europe, and packaging facilities. | |
| Consumer Hardware Innovations and AI-Accelerated PCs | 5 | 4 | 0 | 0 | Elad benchmarks new AI edge devices including Rabbit, Humane, and Vision Pro. Mark highlights hardware prerequisites like photon-to-display latency to prevent motion sickness. | |
| Balancing Latency and Efficiency Across Cloud and Edge | 6 | 3 | 0 | 0 | Sarah frames the modern developer dilemma around model chaining latency vs centralized datacenter compute. Mark validates her point with edge-case autonomous driving architectures. |
Statements from this episode (19)
Papermaster: AMD chips power all Microsoft Xbox and Sony PlayStation consoles
“We're underneath all the Xbox, all the PlayStation, as well as many gaming devices that that, that you buy when you buy your add-in boards.”
Papermaster: Computing requires combined scalar CPU and parallel GPU architectures
“And to me, it was clear that the industry needed That powerful combination of the serial, the scalar competing of these traditional CPU workloads and the massive parallelization that you get from a GPU.”
Papermaster: AMD has shipped integrated CPU-GPUs longer than any competitor
“So we've been shipping CPUs and GPUs combined for PC applications longer than anyone started shipping those in 20 11. With what we call APUs, accelerated processor units.”
Papermaster: AMD CPUs and GPUs power the world's largest supercomputers
“And so we focused first with you know, big government bids that ended up leading to supercomputer wins that we now have AMD CPU and AMD GPUs under the world's largest supercomputers.”
Papermaster: AMD's MI300 AI accelerator beats rivals on compute and power
“It's got a tremendous performance advantage, and we did that very purposely. We created very efficient engines for the math processing that you need for that training or inference processing, but we also brought the memory that you need to have more efficient …”
Papermaster: Most non-LLM AI inferencing currently runs on general-purpose CPUs
“The fact is the bulk of inferencing done today is done on general purpose CPUs, not the huge LLM inferencing, but, you know, just general inferencing for AI applications.”
Papermaster: Hugging Face tests new models equally on AMD and NVIDIA GPUs
“We've partnered with Clem and his team. They test as they release any of those language models they're testing on AMD with our Instinct GPUs equally as they're testing on NVIDIA.”
Papermaster: AMD is one of two qualified hardware platforms on PyTorch
“We're, One of two qualified offerings on on PyTorch, and so all of that testing is being done you know, routinely with the regression testing that's run literally every night on any software release.”
Papermaster: AMD builds CPU and GPU compilers on open-source LLVM
“Our compiler for our CPUs is LLVM. It's open source. The LLVM is underneath. Our compilers on our GPU, but more than just the compiler on the GPU, we've opened up the rock and stack.”
Papermaster: AMD rejects proprietary walled gardens in favor of open source
“Sarah, the point is, we're not about locking in someone with a proprietary wall garden software stack. What we want is we want to win with the best solution, and we want, or we're committed to open source and we're committed to giving our customers choice.”
Papermaster: AI compute supply constraints will disappear as AMD ramps production
“Well, that's definitely happening. I mean, the supply constraint will go away. We'll be a part of that. We're ramping up and shipping as we speak on our instant line, and it's going quite well.”
Papermaster: Electrical power availability will become the ultimate AI data center constraint
“This is, I think, ultimately going to be certainly a key constraint and you see you know, all the major operators looking for sources of power”
Papermaster: Slowing Moore's Law yields density gains but limits power efficiency
“And with Moore's law slowing, it means you still get those device improvements, but it Costs more. Your power's not coming down as much as it used to, and you are still getting that integration. You're still certainly being able to pack more devices, and, but …”
Papermaster: AMD is partnering with TSMC on its Arizona fab expansion
“So you see TSMC building fabs in, in Arizona. We're partnering with them.”
Papermaster: Semiconductor packaging requires geographic diversity beyond foundry capacity
“And so it goes beyond the foundry. It's the same thing with the packaging. So where do you, as you put those chips onto carriers and you need to interconnect it, you need that ecosystem to have geographic diversity as well.”
Papermaster: Local AI enablement will transform PCs into a new device category
“I mentioned the AI enablement in PCs. That's gonna, I think it's almost gonna make PCs a new category, because when you think of the kind of applications that you're going to be able to run with super high performance, but yeah, low power inferencing you can r…”
Guo: AI app developers prioritize latency and on-device compute execution
“I think in, like, in the new era of trying to create experiences and fighting, like, all these, like, new application companies are fighting latency. As a primary consideration, because you have the network, the models are slow, you're trying to chain models, …”
Papermaster: AI architectures should split workloads between cloud efficiency and edge latency
“Writing applications that where you don't have that latency, that, ah, you know, that, ah, dependency on, on a lag in computing, run it on the cloud. It's going to be the most, ah, it's going to be the most efficient because you're optimizing this massive data…”