Technology Analyst, "Silicon Co-Design". Read by 1.3k+ industry professionals. Breaking down the hardware reality of next-gen datacenters from first principles.
My in-depth AI hardware content is written for curious people and NOT written for the 90% of hype and buzzword driven people looking for a quick stock pick.
The internet is already saturated with trendy semiconductor industry content, especially in the software domain, that does not serve the engineers working in those industries well.
After personally attending major flagship semiconductor conferences, I find there is a massive difference between the way engineers talk about technical topics vs how they are portrayed on the surface.
I write in-depth technical content specifically for the system architects, engineering leaders, and engineers actually working on these systems to gain more system context. We dive into conventional architectures and real time constraints such as jitter, thermals, packaging, and the PDN from first principles and how they impact the performance of high-speed and high-power systems.
If you’re not interested in this level of depth and want to stay on the surface, this Substack is not for you.
However, if you’re curious to learn more, I invite you to check out a few of my deep dives on my Substack “Silicon Co -Design”.
https://t.co/atebRcUn6C
@zephyr_z9 Yeah this sounds like a case where NVIDIA is saying "We did our part correctly by designing a chip and defining gigantic PCB specs for others to hit; why can't others manufacture this gigantic PCB in time?".
@zuckerbinjain True, though in my opinion, most of the difficulty in chip design is not necessarily the technical concepts of VLSI, but getting up to speed with the bloated tool complexity of Cadence and Synopsys with a tapeout looming 😐
Yeah I do think Michael Burry is smart, but people put him on a pedestal due to ONE prediction about the 08 crisis.
People don't realize they are falling into the trap of "survivorship bias" and it causes them to give more weight to his opinions even when he is wrong half the time after the fact.
Yeah I feel like academia is really overly specialized and dependent on the topic you choose based on limited information.
It seems like people entering academia get locked into what ever topic they select and makes it harder for intellectual capital to shift focus to more meaningful challenges as they arise.
I myself have been to academic conferences like ISSCC with only an MS and seen the work of highly specialized PhDs. While its nice to see so much progress in so many specialty areas, its almost like each specialty don't really talk to each other and some specialties are treated more like darlings than others based on the commercial demand for products related to it.
There's this book I like called "Range: Why Generalists Thrive in a Specialized World" I really like and validates my experience.
@oprydai Yeah everyone is getting into software and its being rapidly commoditized because its fast and easy.
But hardware is slow and painful that not many people are willing to tolerate. Many people don't even know where to begin to learn hardware.
Optical communications is one of the hottest, yet least understood investment opportunities.
In my highest viewed Substack post, I provide a rigorous technical foundation of the optical communications space straight from expert sources of truth.
At the beginning, I introduce the conventional Intensity Modulated-Direct Detection optical approach with VCSELs used for short reach and highlight challenges with scaling conventional IM-DD to higher data rates.
After the paywall, I dive into more emerging trends for high speed and long reach, such as Co-packaged Optics, Silicon Photonics, External Modulation, and Coherent.
https://t.co/fqB35pqYWY
Interesting. I find that co-design of high speed SerDes and advanced packaging is really challenging to optimize simultaneously since the knowledge of both domains have been historically separated.
I feel there are many cross-domain constraints that each discipline should be aware of, such as edge density, sources of jitter in the SerDes, channel loss, and the serviceability of the CPO itself, for engineers to understand the broader system context and make efficient tradeoffs.
NVLink is Nvidia's crown jewel for CPO, but it faces a major physical constraint: the limited edge / beachfront density for high speed electrical interconnects.
As interconnect speeds scale past 200Gbps, this creates major real estate tradeoffs over which components are placed along the edge of the ASIC.
My latest deep dive discusses the co-design challenges of Marvell and Celestial AI's optical approach that perhaps attracted an investment from NVIDIA for use in their system.
https://t.co/Ee3U0y5VHk
Effective co-design requires engineers to step out of their comfort zones and embrace cross-domain knowledge to arrive at an optimized, system-level solution.
The following is a blurb from my upcoming post, "A Deep Dive into High-Speed HBM5 Interconnects: Signal Integrity and Co-Design Challenges" to be released next week.
Make sure you "follow" and "subscribe" so you'll be the first to know.
I think the concept of an AI pin will make a comeback. The concept was actually heavily criticized by influencers like Marques Brownlee when Humane released their AI pin as the worst product he has ever reviewed.
I think in the future, AI pins are going to function as easily accessible speech recognition for AI rather than typing in all of our thoughts on chatGPT.
While a lot of attention is being drawn to the current HBM shortage, I'm drawing up a post on the practical technical challenges in scaling HBM5 performance.
Here is a insight I made from the draft thus far:
Yeah, the SAT English test heavily favors a superficial familiarity with the obscurities of specific literary genres rather than critical thinking.
This is why I only got a 580 because I didn't really understand the point of studying genres not within my interests to impress gatekeepers, when most of the concepts tested are disconnected from business reality.
Marvell's Electro-Absorption Modulators and NVIDIA's DWDM can potentially be powerful together to form a complete scalable optical communication system.
Given the recent NVIDIA-Marvell partnership, it made me think about how both their co-packaged optics technologies can potentially combine with the advantages of both and how OMIB can be used to ease the shoreline density of ASICs.
In this post, I do a deep dive of the physics of EAMs along with a link to a deep dive of NVIDIAS DWDM.
https://t.co/IuuBwFK9oP