HNHacker News
TopNewBestAskShowJobs

brennanpeterson

252 karma · joined November 15, 2017

submissionscomments
brennanpeterson··on AMD Zen 5 CPUs Reportedly Made on TSMC's 3nm Process, Mass Production in Q3
It is a roughly 20% improvement to speed and power per area. And about 20% density.

So you get more, faster cores/units, in the same footprint.

sram.is the exception as it scales poorly, so don't expect much. If anything, this is the largest and most important hardware issue facing all the chip gains, since so much of performance is memory dependent.

brennanpeterson··on Intel, Samsung, and TSMC Demo 3D-Stacked Transistors
It is?

A factory makes transistors ,and if you increase a 'node', you make twice as much. If you do an amazing job, you might reduce cost 10%.

So by far the best way to maximize value in semiconductors is to enable shrink.

But you also just don't hear it in the popular or even engineering press. Most manufacturers and designers look at a PPAC curve (power, performance, area, cost) and find optimal design points.

As for spreading it out: the unit of production isn't a wafer, it is a lithographic field, which is roughly 25*35mm. You cant practically 'speead out' much more (ok, you sort of can with field stitching, but that is really expensive).

brennanpeterson··on ASML names Christophe Fouquet as new CEO
The stage, the mirror/aberration control, and metrology (overlay/,reg and height/focus) are massive projects

A fair case can be made that the twinscan concept is the key.

brennanpeterson··on ASML names Christophe Fouquet as new CEO
Having met the guy...he is pretty close to knowing all the bits at a pretty deep level.

That is part of the heart of asml, people are expected to know things at a very deep level.

brennanpeterson··on Engineering Statistics Handbook (2012)
It is old school in the sense that it is quite pragmatic. It is aimed at Mainstays of engineering like 'did I improve performance in this line' or 'ate these systems giving the same output's or even 'how do.i design an experiment in a really complex system'. And they generally work well.

Just do work using appropriate tools. There are reasons for all sorts of tools, use the right ones.

Reproducibility crisis is mostly about marginal results, publication bias, and humans.behavior. Bayesians can (and do!) The same things. It isn't the tool, it is the people.

brennanpeterson··on Embracing Chaos: The Imperfect Art of Semiconductor Manufacturing, Lithography
For a very low density experiment and. A few other cases, it is an option. Or for mask writing.

It is very informative to consider the data rate. To approximate (and perhaps do bad phone math) lithography tool has ~1bit per 10nm^2, doing 1 field ever 20s. Or 600mm^2/20s. 30mm^2/s, or about 3E12 bits/s. That's amazing.

brennanpeterson··on Sam Zeloof and Jim Keller start a new semiconductor fab
You have Lam, Applied, and KLA in SV which are 3/5 of the largest equipment suppliers. Along with that the entire supply chain and support. You also have most of your likely customers nearby.

So it remains expensive and still a great place to start.

brennanpeterson··on Ask HN: Engineers Running a Business?
Start by not calling it the boring stuff.

Atonse's comment is excellent. You can get excellent CFO and government finance help via contractors.

But in the end you will want to at the least understand it: what marketing, fundraising, sales, and financial commitments you make in your company are your responsibility.

brennanpeterson··on My Favorite Math Problem
I am curious. Are all boards that have equal white and black tiles counts solvable?
brennanpeterson··on The machine that saved Moore's Law
As a rule of thumb, about 10 unit cells (5nm). There are some exceptions, but this is a useful number.

That means that practically, the smallest conceivable metal pitch is 10nm, the smallest gate is 20.

From 5nm, it implies a factor of 10 is left, or about 3-4 nodes.i expect lower scaled nodes at larger delays, but another 15-20 years.

brennanpeterson··on What’s harder to find than microchips? The equipment that makes them
Masks are not whole wafer , except in some packaging cases. The mask illuminates a 25*35 area, for both 300mm and 200mm.

Masks are not a significant expense for most processes. Nor is the bare wafer. It is all the process steps that drive the cost.

300mm qafer

brennanpeterson··on Next-gen chips will be powered from below
I doubt this is true. Mike Mayberry (Intel) softly announced buried power rail last year, as a 'within 5 years' tech. That suggests it has been prototyped, since from final tests to HVM is at least 2 years.

Maybe not great.

Within 10 means an idea with simulation. Within 5 means prototyped. Interesting means never. Or that, at least, is my decoder ring.

brennanpeterson··on Next-gen chips will be powered from below
This is true but also wrong. The transistors themselves have a specific operating voltage.
brennanpeterson··on Richard Feynman’s Integral Trick (2018)
Mostly agreed, but I did like the last sentence: sometimes it is easier to solve the general problem than the specific one.

That is a useful trick to keep in mind all the time. The point of course, is not to learn to solve integrals, but to learn a transferrable piece off mental jiujitsu.

brennanpeterson··on Smart METL airless metal bike tires use a “shape memory alloy” for deformation
Woven metal tires are old. And nitinol is also old (it was not developed for the rover, a wiki check would show as much).

But the combination is new. I imagine the conversation went 'hey I am working on chainmail tires' 'areny those heavy' 'well. We need them light, strong and elastic, and you only get two...' ' hol up, there is an option, you heard of nitinol?'

And thus, cool history.

brennanpeterson··on I self-learned to code at 30
I have done a little bit of coding: macros, small programs, tool controllers. What kills me is that I can do 90% of what I need using other tools (excel, JMP, imageJ, matlab) and coding mostly makes things take longer. Most of the effort is plumbing: where are my files, what is the structure of the output, where is the file: the boilerplate stuff.

I usually know what I want and how, but it takes quite a bit to structure. Back in the day, labview+scripts/c were great. You could get to a point, and then extend as needed.

The other problem, of course, is finding meaningful problems to solve. Would I like to build a semiconductor process simator: yes, yes I would. Will I....not likely.

brennanpeterson··on Product Manager vs. Product Marketing Manager vs. Product Owner
Capital equipment is even more barren in information.

It might be better said that there are sales, development, and operational aspects to product management. You can split this into a product owner, sales interface, and engineering interface. Or not.

Even b2b is very different depending on what the market is. If you have 10 customers, the marketing path is different than if you have 1000, or 100,000. The roles shift dramatically.

brennanpeterson··on Intel outsources Core i3 to TSMC's 5nm process
Tsmcs r&d spend has been equivalent to Intel's for the last decade. TSMC did a great job, most particularly on EUV introduction.

I don't think wall street cares about outsourcing, except as a potential savings. The hollowing out is real, but it is fixable.

It will take a really different approach, though.

brennanpeterson··on Intel is dying and we can’t save it
Everything in M1 and everything on Intel 14nm. Go read wikichip. It is there. IEDM just happened, tons of detail there.

3nm, sure, good luck.

brennanpeterson··on Intel is dying and we can’t save it
Intel 14nm is 42/70/52 (fin/cgp/mmp) with a 7.5 track height. HD stamBitcell is 0.05um2 Very restrictive design rules

Tsmc 5nm is 27/54/36 with a 6T and 7.T option. More open design rules. HD SRAM cell is about 0.025um2.

This is almost exactly what you expect for 1 node of litho scaling.

The more difficult part is that the designs and related process optimizations for intel are different.

brennanpeterson··on What do RISC and CISC mean in 2020?
You are right, I misread 'high bandwidth' as HBM. I was surprised at the cost as well.
brennanpeterson··on What do RISC and CISC mean in 2020?
A good design + a 1 node (15% transistor level benefit) advantage + embedded(ish) HBM. You kind of expect 15-30% benefit in the same design. Whether CPU or bandwidth bound. Some latency bound benchmarks will be at par.

It is a perfect hit on all cylinders: a good design, a node advantage, and better memory.

This has happened before. We have been at a recent lull in performance, but annual performance increases used to be about 30%.

brennanpeterson··on Artist creates nanosculptures smaller than a human hair (2014)
Some yes. Often residual carbon will deposit on the surface and change it (depositing a bit of a surface layer. If you do ion beam imaging, you do etch the surface away with each image.
brennanpeterson··on Civil Liberties in Times of Crisis [pdf]
This has been said for the history of the ACLU. And there are difficult balances there: do you defend the Nazi March? What if that March includes armed marchers aiming guns at others? Is aiming a gun speech? Or speech the ACLU has interest in defending? Does it maximize the freedom for marchers, counter marchers, or citizens?

The ACLU had to come to terms with consequences, and harder answers. After defending and enabling a march where people died, the ACLU had to answer whether it was responsible...and it was.

So the organization is a bit different, but I am not so sure that the ACLU could have continued. It had avoided really hard problems, and then had to come to terms with it.

I prefer it this way: it is messy and tricky and also much less likely to be a disaster.

brennanpeterson··on Transistor Options Beyond 3nm (2018)
There are various areas. 3DNAND is a fully 3D design, and scaled reasonably past the limits of 2D. Crosspoint memory is a simplified 3D structure.

In logic, people (Unity, for instance) talk about fully 3D, with stacks of transistors layered. This turns out to be capability and thermally limited.

The most interesting variant today are the N over P nanosheets. Instead of having N and P transistors next to each other, putting them on top saves area. Contact and interconnect are still limiting, but this gives a node shrink without lateral shrink.

It isn't completely obvious to me that this gives enough design flexibility to be worth it, but maybe?

So anyway, there are good attempts, and the stacked NPN is the most likely.

brennanpeterson··on Transistor Options Beyond 3nm (2018)
What is the magic technique?

Selective growth and patterning already have precision in the sub angstrom range, over mms of distance. This is what 'TMU' is, when you read an ASML or Lam paper on the topic. Current target is sub nm, for total variation cross die per layer.

I strongly recommend a read of the Tennant's law series onnlithoguru, or Tim Brunner's paper on why optical lithography wins forever.in short, optical is better (including EUV).

There are whole conferences dedicated to atomically precise manufactue, and hundreds of not thousands of people working in the field.

brennanpeterson··on ‘Better Yield on 5nm Than 7nm’: TSMC Update on Defect Rates for N5
Sure! It is a huge mix. ASICs don't, mobile cannot, and you can't really just ignore.

But there are also other tricks, like redundant vias, or SRAM spares, which do make 'yield' less listed by defects than it might first appear.

brennanpeterson··on ‘Better Yield on 5nm Than 7nm’: TSMC Update on Defect Rates for N5
I would add lithoguru.vom and the related courses (Chris Mack) as well as semiengineering.com
brennanpeterson··on ‘Better Yield on 5nm Than 7nm’: TSMC Update on Defect Rates for N5
Agreed on binning, but all DRAM and all NAND are fault tolerant. Between all memory, GPU and GPU-like, and CPU....I have all but mobile, or about 70% of the market.
brennanpeterson··on ‘Better Yield on 5nm Than 7nm’: TSMC Update on Defect Rates for N5
Yes, and I believe TSMC does. Because the dose scales as resolution to the 4th power (or so), it is better to double pattern than to try to single pattern at massive dose. There are also cases where double patterning can reduce defects in self aligned schemes.
← PreviousPage 2 of 4Next →