- Future of Computing Newsletter
- Posts
- š¤š¦¾ What if We Canāt Just Build Bigger AI Data Centers Anymore?
š¤š¦¾ What if We Canāt Just Build Bigger AI Data Centers Anymore?
A Newsletter for Entrepreneurs, Investors, and Computing Geeks
Another week where the DeepSeek hype continued unabated. One key takeaway is how improved efficiency in utilizing compute resources can lead to better AI models. However, considering rebound effects, this efficiency may ultimately drive even greater overall compute consumption.
This raises an important question: What happens when we can no longer effectively scale single data centers? In this edition, weāll also cover the latest advancements in photonic interconnects, the viability of quantum computing, and the promise of neuromorphic computing.
Finally, weāll host our first Future of Quantum Computing Meetup together with {Tech: Berlin} at The Delta Campus in Berlin, proudly supported by Quantistry & Kipu Quantum. Will be featuring a keynote on "Quantum or AI - Why not both?" by Enrique Solano, Co-CEO and Cofounder at Kipu Quantum.
Spotlights
āļø ParityQC: Shaping the Future of Quantum Architecture (Future of Computing)
ParityQC, founded in early 2020 by Wolfgang Lechner and Magdalena Hauser, specializes in developing quantum computer architectures and an operating system called ParityOS. Their unique approach involves encoding quantum information using parity, which allows for more efficient qubit connectivity and scalability.
š¦¾ What happens when we canāt just build bigger AI data centers anymore? (The Register)
Due to power and resource constraints, the demand for computational power has led to challenges scaling single data centers. To address this, experts suggest distributing AI workloads across multiple interconnected data centers, effectively creating a virtual supercomputer. This approach leverages existing high-speed data centerr interconnects but faces challenges related to latency and bandwidth over long distances.
š¦¾ How analog in-memory computing could power the AI models of tomorrow (IBM Research)
Analog in-memory computing integrates memory and computation, reducing data transfer bottlenecks inherent in traditional architectures. A recent study demonstrated that mapping components of mixture of experts (MoE) models onto a 3D non-volatile memory structure in analog in-memory computing chips can outperform GPUs in both throughput and energy efficiency. This method is promising for deploying large transformer-based AI models in cloud and edge environments.
āļø The success and failure of quantum computing start-ups (Nature Electronics)
āThe success of quantum start-ups depends on intelligent business strategies, appropriate funding, adaptability, and the ability to balance scientific expertise with business acumen. Despite the risks, patient investment and resilience are crucial, as quantum computing has the potential to revolutionize computingāthough whether it will achieve this disruptive impact remains uncertain. The article draws parallels to AI's historical boom-and-bust cycles, suggesting that quantum computing may experience similar fluctuations before reaching maturity.
š¦¾ TSMC Founder Morris Chang (Acquired Podcast)
Ben and David from the Acquired Podcast had a chance to interview TSMC Founder Morris Chang in a rare English interview! Also, check out their TSMC (Remastered) podcast episode, covering the complete history & strategy of TSMC.
Headlines
ā”ļø Wave Photonics announces the release of the most expansive photonics PDK in the world for wavelengths from 493nm - 1550nm (Wave Photonics)
š§ Neuromorphic system using capacitor synapses (Nature Scientific Reports)
š§ With Nature article and $4 million grant, Schuman advances community-level neuromorphic computing (Eureka Alerts)
š§ Neuromorphic Semiconductor Chip Learns and Corrects Itself: a KAIST research team has developed a memristor-based integrated system (Tech Briefs)
š§ Rapid learning with phase-change memory-based in-memory computing through learning-to-learn (Nature Communications)
š¤ Aleph Alpha raised $500m, then ditched the AI race. What comes next? After pivoting away from LLMs, the German AI darling unveils a new tool to help B2B customers (Sifted)
š¤ TĆ¼lu 3 405BāThe first application of fully open post-training recipes to the largest open-weight models to surpass the performance of DeepSeek V3 (Allen AI)
š¦¾ ASML posts record figures: The Dutch company closes 2024 with an annual turnover of a good 28 billion euros (heise online)
š¦¾ ASML to launch High-NA lithography machines to boost chip manufacturing (Innovation Origins)
š¦¾ Japanās Semiconductor Photoresist Monopoly (Asianometry)
š¦¾ China makes inroads in DRAM chips in challenge to Samsung and Micron (KrAsia)
š¦¾ EdgeConneX is looking to invest ā¬600 million in data centers in Heusenstamm on the outskirts of Frankfurt, Germany (DCD)
š¦¾ Data Centers Consume 3% of Energy in Europe: Understand Geographic Hotspots and How AI Is Reshaping Demand (Power Magazine)
š¦¾ New Data Center Developments: This curated selection will help you stay on top of the latest data center development news (Data Center Knowledge)
š¦¾ The Road To Super Chips: Challenges in achieving orders of magnitude performance improvements in processors (Semiconductor Engineering)
š¦¾ The Era of Ubiquitous AI Computing by MediaTek and Counterpoint Research (Mediatek)
āļø Quantum Sourceās Scalable Photon-Atom Technology Enables Practical Quantum Computing (Quantum Insider)
āļø ParityQC Introduces Parity Twine: Quantum Algorithm Synthesis Achieves World-Record Efficiency (Quantum Computing Report)
āļø The quantum computing reality check (Infoworld)
āļø Useful quantum computing is inevitableāand increasingly imminent (MIT Technology Review)
āļø European Commission Funds ā¬3M Quantum Chip Project to Address Scalability Challenges (Quantum Computing Report)
Funding News
ā”ļø Undisclosed ā NLM Photonics: bring energy-efficient, high-performance electro-optic modulation technology to AI, data centers, quantum computing, and more (InsideHPC)
ā”ļø $21M Series A ā iPronics: Optical Networking Engine in AI data centers, enabling fast, scalable, and high-bandwidth communication for energy-efficient AI (Photonics)
āļø ā¬100M Series B ā Alice & Bob: creating the first universal, fault-tolerant, and useful quantum computer (Tech EU)
ā check out our interview with Alice & Bob: Shaping the Future of Fault-Tolerant Quantum Computing from fall 2023
Deep Dive: AI Compute Under Hardware Constraints
Another week has passed, and the excitement around DeepSeek shows no signs of slowing down: Following the release of their reasoning model, R1, they have now introduced a set of image generation models.
As Sean Goedecke put it: āThe Chinese AI lab1 DeepSeek recently released their new reasoning model R1, which is supposedly (a) better than the current best reasoning models (OpenAIās o1- series), and (b) was trained on a GPU cluster a fraction the size of any of the big western AI labs. Unlike the big western AI labs, theyāve released a paper explaining what they did.ā
While there is suspicion that Nvidia H100 GPUs are being smuggled into China via Singapore, circumventing U.S. export controls, DeepSeekās technical paper suggests that many of their innovations stem from working within hardware constraints. Their breakthroughs make the most sense when optimizing for H800 GPUsāa cut-down version of the H100 with reduced memory and bandwidthārather than relying on top-tier hardware.
DeepSeek has demonstrated genuine innovation, opting not to use Nvidiaās CUDA but instead programming GPUs directly with āPTX, a low-level instruction set for Nvidia GPUs that is basically like assembly languageā (Stratechery). This shows that U.S. export controls have had an unintended effect: they have driven DeepSeek to develop novel software-based optimizations to work around the memory limitations of the H800s.
Their work highlights just how much more performance can be squeezed from GPUs with low-level engineering. In contrast, Western AI companies havenāt needed to take this routeāraising more capital to secure additional Nvidia GPUs has been the easier and more common approach. DeepSeekās success suggests that software-side efficiency gains could become an increasingly important frontier in AI development, especially as hardware and energy constraints materialize.
š” Love these insights? Forward this newsletter to a friend or two who can subscribe here