Vikram Sekar

High Bandwidth Flash (HBF). A NAND-based memory tier promising 8–16x the density of HBM.

If it works, GPUs could ship with terabytes of memory instead of hundreds of GB. First samples land 2026.

The real question: does HBF complement HBM—or disrupt it?

High Bandwidth Flash: NAND’s Bid for AI Memory Can HBF be a new memory tier for AI workloads? Where its limits will show, and what are use-cases for a high capacity memory thats slower than HBM or GDDR.

HBM is hitting a wall.

Costs are high. Capacity is capped. DRAM scaling is over. Yet AI models keep getting bigger.

Sandisk + SK Hynix think they have the answer.

Sometimes you find art in the craziest of places!

Read more:

open.substack.com/pub/viksnews...

A Beginner’s Guide to Interconnects in AI Datacenters Understanding performance and trade-offs shaping interconnect choices in AI datacenters.

This week’s deep dive:

how GPUs, racks, and entire datacenters are wired together — and why copper vs. optics is the next big frontier.

www.viksnewsletter.com/p/a-beginner...

Interconnects are the hidden backbone of AI datacenters.

A single rack carries ~2 km of copper. Multiply that across hundreds of racks, and the business opportunity becomes clear.

The choice of interconnect defines performance, power, and cost.

Why Documentation is the Missing Link Between AI and Chip Design Chip design has always depended on unwritten know-how. To bridge tribal knowledge and machine intelligence, engineering culture needs new ways to capture and share the messiness of real design work.

viksnewsletter.substack.com/p/why-docume...

Chip design knowledge is tribal.
AI promises to replace engineering grunt work—but machines need documentation to learn.

Problem? Documentation is the one thing engineers hate doing.

The real challenge: how do we make documentation machine-readable and engineer-friendly?

Read more: 👇🏽

Why is HBM so Hard to Manufacture? An in-depth look into the incredible engineering behind the memory used in GPUs for AI accelerators.

Only three companies make it: SK Hynix, Samsung, and Micron.

And a Chinese competitor: CXMT.

This week, I break down why HBM is so hard to manufacture — from fundamentals to cutting-edge bonding. 👇🏽
www.viksnewsletter.com/p/why-is-hbm...

The memory wall is the real bottleneck in AI hardware.

GPUs today can deliver TFLOPS, but they sit idle waiting for data.

What bridges the gap is High Bandwidth Memory (HBM) — the only memory tech that can keep up with modern AI training.

HBM is so critical that it makes up 50–60% of GPU cost.

Only a few companies dominate the EDA tool market- hard to disrupt the space

What Chip Design Can Learn from Supersonic Jets Lessons from Boom Supersonic’s engineering culture and what they mean for the future of chip design.

www.viksnewsletter.com/p/what-chip-...

A jet broke the sound barrier with 50 engineers and 1/10th the usual budget.

Chip design still needs armies of people and billions. Why?

Outdated workflows. Clunky tools. Rigid hiring.

Boom Supersonic built its own software platform and leapt ahead.

Who will build the chip industry’s “mkChip”? 👇

Literally the best picture of High Bandwidth Memory (HBM) on Chip-on-Wafer-on-Substrate (CoWoS) packaging I have ever seen.

The Manufacturing Challenge Behind Nvidia’s Plan to Mount AI Chips Directly on PCB Platforms Is substrate-like PCB technology sufficiently ready to replace ABF substrates that are ubiquitous in AI chips today?

Read more here:

open.substack.com/pub/viksnews...

That’s why Nvidia’s exploration of CoWoP is interesting. By mounting interposers directly to PCB platforms, CoWoP could bypass ABF entirely.

But the question remains: can PCB makers hit the density, yield, and reliability that AI accelerators demand?

Nvidia’s AI chips lean on TSMC’s CoWoS packaging — where GPUs + HBM sit on an interposer mounted to an ABF-based substrate.

ABF is the industry standard but also a major bottleneck — and the U.S. has no domestic source.

How to Escape the Corporate Rut for Midcareer Chip Engineers Simple ideas to keep your paycheck, grow skills, and enjoy the work again.

How you benefit:
→ You build a rare skill mix, restore meaning, and open doors
→ All without risky job-hopping

Read for the complete blueprint:👇🏽
www.viksnewsletter.com/p/how-to-esc...

How to learn effectively?
→ Pick one adjacent skill
→ Use what companies offer: tools, experts, tuition support → Run a 6-week sprint: 3 hours/week, one mentor chat/week, one demo
→ Request a 10–20% trial assignment to apply the skill
→ Ship proof: a design note, a small block, or a script

The cycle is common:
→ Early excitement, then sameness, then burnout risk.
→ Promotions rarely solve the core issue

Apply a three-step plan:
→ Identify your career phase
→ Identify your personal engineer archetype
→ Run focused learning sprints

Chip engineers feel in a rut after almost a decade on the job.

Here's an actionable plan to get out of it. 👇🏽

First, why engineers feel stuck:
→ You do a tiny task in a huge flow
→ When impact is unclear, motivation drops and you feel replaceable

A Primer on Transformer Architecture: Model Parameter Calculations, Estimating GPT-5 Architecture A detailed look at transformers, the role of attention, and decoder-only architecture; Learning to count model parameters based on architecture and making guesses on GPT-5 internal construction.

www.viksnewsletter.com/p/a-primer-o...

For paid subscribers:
→ Parameter counting using Google’s BERT-Base
→ Detailed look at GPT2/GPT3
→ An educated guess at GPT4
→ Speculating on GPT5
→ References

Downloadable Google Sheet you can use to make your own parameter estimates and speculate on GPT-5!

See comment 👇🏽

For free subscribers:
→ Birth of the Transformer
→ Tokens, Embeddings and Vocabulary
→ QKV, Context Window, and Batch Sizes
→ Attention Calculation Methods
→ Decoder Transformer Architecture

GPT-5 is here. 🥳 With sloppy graphs, hype and disappointment.

We know it has trillions of parameters, but where does it come from?

Today's deep dive post covers everything you need to know.

You will even make your own GPT-5 estimates at the end of the post.

7 Unwritten Rules for High-Impact Engineering Careers What they don’t teach you about thriving in technical organizations, but will define your success.

5. Your visibility is your own responsibility
6. Cross boundaries without stepping on toes
7. Spot CYA but don’t get sucked in

Read more here:
www.viksnewsletter.com/p/7-unwritte...

1. Your value is the specific problems you can solve
2. The best projects are rarely assigned — you have to find them
3. Know who can say yes (and when they will)
4. Survive high-stakes design reviews

7-rules for high impact engineering careers:

That no-one ever tells you...

No way to tell

Yes true- but over 90,000 people took the survey. Data set is large enough.

The Decline of MATLAB: What Engineers Should Learn Instead A frank look at why Python and Rust is the future of the hardware programming stack.

The 2024 Stack Overflow developer survey puts Matlab as the least admired programming language among 50 different ones.

Lower than Cobol, Visual Basic, Prolog and Fortran.

Read full post. 👇🏽
www.viksnewsletter.com/p/the-declin...