Kris Murphy

StorageReview just calculated Pi to 314 trillion digits using 40x 60TB Micron Technology 6550 ION SSDs. Impressively, they did it with a single Dell Technologies PowerEdge 7725 server with dual 192-core AMD Epyc CPUs, 1.5TB of DDR5, and using 1/8th the power of the previous record. Outstanding work.

33

Paper 22/25 in our series of "25 papers for 25 years of warehouse-scale computing" is the Swift paper from SIGCOMM'20. 💡 Fun fact about today's paper: Swift is so simple, and its performance so good, that it completely changed people’s expectations of what a WSC network can do. In fact, initially it caused some confusion: twice, bugs were incorrectly filed for monitoring failures because highly-utilized links reported zero loss. So, what does Swift do? ⚡ TCP has dealt with internet congestion effectively by using packet loss to detect congestion. However, it doesn’t work well in data center networks which exhibit very high speeds combined with very low latencies and small-buffer switches. Swift showed that a very simple protocol based on accurate measurements of packet delays outperforms all others. Its simple, end-host-based approach achieves extremely low network queues, high link utilization, and fast convergence. Link to paper below: 👇 Swift: Delay is Simple and Effective for Congestion Control in the Datacenter, Gautam Kumar, Nandita Dukkipati, Keon Jang, Hassan Wassel, et al, SIGCOMM 2020 (2020), pp. 583–598  https://lnkd.in/grJMu2rn The style used for today's Gemini3 generated infographic for the paper: "Graffiti Wall: Brainstorming ideas spray-painted onto a brick wall background."

67

3 Comments

In Santa Clara, they’re not just building chips, they’re bending light to carry the future. OpenLight has closed an oversubscribed $34M Series A, moving from a Synopsys subsidiary into an independent, venture-backed company. The round was co-led by Xora Innovation and Capricorn Investment Group, with Mayfield, Juniper Networks (now part of HPE), Lam Capital, Legacy New Ventures, LLC, and K2 Access Fund stepping in. This is validation that integrated photonics is no longer stuck in R&D. The roots go back further than most headlines admit. Born in 2008 at UC Santa Barbara through Aurrion, acquired by Juniper in 2016, the tech shifted again in 2022 when Synopsys bought 75% for $90M. Now, with $124M in total backing, OpenLight steps out as a standalone company with momentum that feels inevitable. At the helm is Dr. Adam Carter, CEO, whose career at Oclaro, Foxconn Interconnect, and Cisco proves he knows how to turn #opticalinnovation into market traction. CFO James Dixon cut his teeth at Intel and led finance teams at Cisco’s data networking groups. VP of Operations Dr. Volkan Kaman carried silicon photonics from Aurrion to Juniper before bringing his engineering depth here. This team isn’t guessing at execution, they’ve done it before, and at scale. The technical edge is what separates OpenLight. While rivals glue external lasers onto chips, OpenLight integrates indium phosphide directly onto #silicon, embedding lasers, amplifiers, modulators, and #photodetectors into one die. That means fewer parts, simpler packaging, lower power, and higher speeds. Their 400G modulators run at just 0.6Vpp, with designs already on track for 1.6Tb/s and 3.2Tb/s PASICs. For AI #datacenters suffocating on bandwidth demand, this isn’t nice-to-have, it’s oxygen. Customers are growing just as fast. From 3 in 2023 to 17 in mid-2025, with 20+ designing through the OpenLight PDK, adoption is spreading. Revenue has grown tenfold since launch, with production deliveries targeted for late 2025 and royalty streams in 2026. With 360 patents and a portfolio set to top 450 next year, OpenLight has invested over $500M into a moat competitors can’t swim across. This Series A isn’t a test, it’s a scaling round. Expanding the PDK library, pushing 400G modulators and integrated lasers forward, broadening foundry options beyond Tower Semiconductor, and building out #customersupport. The global silicon photonics market is set to climb from $2.65B in 2025 to $9.65B in 2030 at 29.5% CAGR. OpenLight looks poised to set the pace. Takeaway: hardware doesn’t creep forward, it leaps. OpenLight just raised the capital to make that leap, and with Carter, Dixon, and Kaman leading, it looks less like a gamble and more like destiny. #Startups #StartupFunding #VentureCapital #SeriesA #OpticalTech #Photonics #SiliconPhotonics #Laser #LaserTech #Data #DataDriven #Technology #Innovation #TechEcosystem #StartupEcosystem If software engineering peace of mind is what you crave, Vention is your zen.

15

1 Comment

The new #cuQuantum SDK v25.11 is here, introducing cuPauliProp and cuStabilizer to supercharge large-scale #quantum simulations on NVIDIA GPUs.⚡️ Delivering up to 1,000x speedups for large-scale quantum circuit validation and #errorcorrection workloads, these libraries enable developers to simulate 127-qubit "utility" circuits and massive QEC datasets with unprecedented efficiency. ✨ 🔗 Dive into the benchmarks and code samples here:

13

🛡️ The Self-Healing Link: How PCIe Guarantees Zero Data Loss PCIe isn't just fast - it’s obsessed with integrity. While we talk about GT/s, the Data Link Layer is running a high-stakes game of "Sequence Math" to ensure not a single bit is lost. If you've ever wondered how a packet survives a noisy Gen5 trace, here is the logic that happens in every controller, every microsecond. 🔢 The TLP "Gatekeeper": NRS Logic When a TLP arrives, the receiver compares the Sequence Number (Seq Num) to the Next Receive Sequence (NRS). There are only three possible outcomes: 1. Seq Num = NRS (The Perfect Packet) * Action: Accepted and sent to the Transaction Layer. Logic: NRS increments, the Nak flag clears, and an Ack DLLP is scheduled. 2. Seq Num < NRS (The Ghost Duplicate) * Action: TLP is discarded. Logic: This happens if a transmitter replays on its own. We don't increment NRS; we just send an Ack for the last good packet (NRS-1) to stay in sync. 3. Seq Num > NRS (The Missing Link) * Action: TLP is discarded. Logic: Something was lost at the Physical Layer. We set the NAK_SCHEDULED flag and fire off a Nak DLLP immediately. No waiting for timers here—we need a replay ASAP. 💾 The Safety Net: The Retry Buffer Transmitters don't just "send and forget." Every TLP is stored in a Retry Buffer until an Ack is received. The "Heavyweight" TLP: A single TLP can take up to 4,138 Bytes of buffer space (Seq Num + Prefixes + Header + 4KB Payload + ECRC/LCRC). The Rule: You don't clear the buffer until the Ack arrives. If a Nak hits or the timer expires, the "Replay" begins. ⏱️ Replay Timer: The Final Arbiter The Replay Timer is the "heartbeat" of link reliability. Starts: On the last symbol of a TLP transmission. Restarts: Every time an Ack is received (if the buffer isn't empty) or when a re-transmission happens. Safety Valve: It holds reset during Link Re-training (LTSSM Recovery) because you can't acknowledge what you can't see! 💡 The Takeaway: Reliability in PCIe isn't magic; it’s a strict combination of Sequence Tracking, Buffer Management, and Precision Timing. Have you ever had to debug a Replay Timer timeout? Was it a signal integrity issue or a credit overflow? 👇 #PCIe #HardwareEngineering #SignalIntegrity #DataCenter #HighSpeedDesign #SiliconEngineering #Debugging

25

2 Comments

Cal Poly is launching their new AI Factory powered by NVIDIA DGX B200 systems. The new AI Factory will give students, faculty, and regional partners access to the same infrastructure accelerating breakthroughs across industry and research worldwide. Discover how NVIDIA and Cal Poly are partnering to advance applied AI education and innovation: https://bit.ly/4oLtsYW

1

Hot take: the memory shortage didn’t cause AI infrastructure problems, it removed the mask. When DRAM, HBM, and NVMe are all constrained, weak architecture gets exposed fast. Idle GPUs, long inference times, and “normal” 30–50% utilization shouldn’t be acceptable anymore. This blog connects the dots 👇 http://spr.ly/6047hpgLh

5

GenAI is becoming a reality on billions of Arm-based devices with Arm KleidiAI, SDOT, and ExecuTorch 0.7. These frameworks enhance AI capabilities for both new and older devices, enabling faster model startups, lower latency, and easy integration. Exciting times ahead as Arm pushes the limits of AI! https://okt.to/o9sdVL

27

Master the next generation of GPU programming with our new deep dive into NVIDIA #CUDA Tile. Learn how to implement high-performance matrix multiplication by shifting from thread-level to block-level programming. Read the full blog.

43

After a decade deep in the AMD/Xilinx ecosystem with Model Composer, I’ve integrated Intel Altera’s DSP Builder into my daily workflow. It’s not a shift—it's an expansion. Both tools are powerhouses, but they have fundamentally different "personalities." One feels like total manual control; the other feels like high-level automation. I’ve summarized my take on the two below, but I’m curious what the community thinks: - Which tool is your "go-to" for specific math-heavy applications? - Which environment do you find more efficient for building "AI Light" models? #FPGA #IntelAltera #AMD #Xilinx #DSPBuilder #ModelComposer #AI #EdgeAI #DigitalSignalProcessing

8

I’m not surprised to see NVIDIA move ahead of the market — and start to separate GPUs into compute-bound and memory-bound SKUs. Rubin CPX: announced as a prefill / compute-bound accelerator 128GB GDDR, 2TB/s bandwidth Lower PFLOPs than Rubin GPU (30 vs 50 on mx-fp4) Ok, so this means Prefill could be faster only if CPX is considerably cheaper and clustered in larger numbers. I haven’t simulated CPX yet, but on paper this looks like a cost-optimization play for prefill workloads, not a game-changer. Unless there’s a hidden “long-context” accelerator baked in, this is incremental. The real issue right now: Decode! Prefill has never been the bottleneck at scale. Decode is where multi-turn research chats stall, and where agentic systems groan for tokens under autoregressive generation. 12 TB/s per GPU and with NVL144 is a good step — but is inefficient for autoregressive models. Until we see decode-optimized GPUs (likely requiring new memory technologies), Rubin CPX won’t move the needle on user-perceived latency or throughput in multi-turn inference. Early intuition: Rubin CPX is good, but the market is really waiting on Nvidia (or someone else) to ship a decode-centric accelerator. That’s where the next leap in sheer performance comes from.

76

4 Comments

AI for drug discovery is a hot topic right now. Behind the theme is a trove of systems optimization problems only starting to be looked into. Cuequivariance v0.6.0 is now out with accelerated attention with pair bias, optional bias support for triangular multiplicative update, and JAX support for triangle kernels amongst others. Release notes: https://lnkd.in/gq8-mXKa Thanks to my amazing team that makes this possible.

51

1 Comment

NEW: I wrote a book on chiplets. Below is the title of the book. "Chiplet-Based System on a Chip (SoC) Architecture." It's for engineers, architects, managers, and students who want a single place  to understand how chiplets are changing the way we design chips. ​ ​ The book covers: ​ o Chiplets-based SOC Architecture development o Partitioning and interconnect trade-offs​ o Packaging challenges (2.5D, 3D, 3.5D)​ o Test, validation, and debug in multi-die systems​ o Security, and Software o Cost, yield, and supply chain implications​ o Case studies from AMD, Intel, NVIDIA, Apple, and more​ ​ It's now available on Amazon in hardcover, paperback, and/or Kindle. Links in the comments.

183

11 Comments

What if better MoE design could unlock more accuracy per FLOP and per Byte? Our new paper, LatentMoE, shows how a principled approach to sparsity can improve accuracy at iso-inference cost. LatentMoE is adopted in NVIDIA’s Nemotron-3 Super and Ultra models. More coming soon. Paper: https://lnkd.in/gGRyC9Xh Blog: https://lnkd.in/gPtPe97U 

409

1 Comment

🤖 AI is quietly reshaping our future while most people are still oblivious. Idaho National Laboratory (INL) has partnered with NVIDIA to advance "PROMETHEUS," their groundbreaking new nuclear reactor project. After a 14-month overhaul, they’ve successfully brought the reactor to its first criticality, combining traditional energy with cutting-edge AI technology. This isn’t just about energy—it's a glimpse into how AI is powering the next wave of innovation in complex infrastructure and energy security. In my view, this signals a massive leap in AI integration beyond software—into real-world, physical systems that impact our daily lives and security. When big labs like INL use AI to optimize and control nuclear processes, it means smarter, safer, and more efficient energy solutions are on their way. For entrepreneurs and anyone interested in tech, this is a clear sign that mastering AI isn’t optional anymore; it’s the key to staying ahead in a rapidly changing economy. 1 ✅ AI is now directly involved in high-stakes energy projects that shape our future. 2 ✅ Big players like NVIDIA and INL show how AI is transforming industries once thought to be immune to automation. 3 ✅ Being aware of these developments helps you spot new opportunities to innovate, invest, or build your own AI-powered solutions. If AI can now control complex systems like nuclear reactors, what will you do to leverage AI for your own success? Will you ignore this wave or ride it to new heights? Try learning about AI applications in your industry today. Build or test small AI tools to increase your productivity or security. Evaluate how AI can optimize your business operations or personal projects. Protect your digital assets by staying informed on the latest AI security practices. Audit your current skills and identify areas where AI literacy can give you an edge. ✅ I understand how AI is being used in real-world critical infrastructure. ✅ I can see at least one way to use AI to my advantage. ✅ I am willing to test one AI or cybersecurity tool this week. #artificialintelligence #cybersecurity

1

The most impactful and important SDK release Trainiun Neuron team has worked on since inception: Native PyTorch support for Trainium, open source kernel (NKI) compiler and pre-optimized NKI kernel library, Neuron Explorer (IMO, the fastest and most intuitive ML profiler in industry), optimized inference implementation for major MoE and video models (GPT-OSS, Qwen, Pixtral….), complete vLLM v1 support !! All with full support for Trainium3. Using Trainium, whether for research, training , or large scale inference just became more intuitive and simple. https://lnkd.in/gjfhjv4j

626

10 Comments

🚀 Impressive innovation from Pliops! Making AI inference not just faster but dramatically more efficient — 2X–4X more users at a fraction of the cost per token. Seamless integration with NVIDIA, AMD, and Dell infrastructure is a real game-changer. If you're building scalable LLM inference platforms, this is worth a serious look. #AI #LLMinference #vLLM #SGLang #GenAI #Infrastructure #GPUaaS #CloudComputing #Nvidia #AMD #Dell #Pliops

10

1 Comment

TACC’s exploration of FP64 emulation on Horizon marks an interesting pivot in High Performance Computing. With hardware vendors aggressively optimizing for low-precision AI workloads (FP8/FP16), the density of native FP64 logic is relatively diminishing. By utilizing the Ozaki scheme to emulate double-precision on Blackwell’s tensor cores, TACC expects to outperform native FP64 execution by 2-3X. This effectively turns 'AI hardware' into high-fidelity scientific instruments, proving that software complexity can bridge the gap between GenAI infrastructure and rigorous physical simulation. https://lnkd.in/gb3-g238

51

4 Comments