2,002 karma · joined March 9, 2013
The article does get it right that you need to reduce the parasitic elements to the chip, but you have to consider like, everything - not just the wires to the package, but also the lead frame, bond wires, the wiring in the chip. Usually the chip designers modeled that all out and they put decoupling capacitors on chip and did PCB model simulations that includes some specific assumption about the impedance curves of the off-chip caps, and they probably used 0.1uF in their simulations.
If anything, you need closely placed decaps to prop up higher frequencies, not lower frequencies. Remember if you have a SPI bus clocking at 25MHz, 25MHz is just the fundamental - you have the whole fourier series going up to 100's of MHz on the edge.
The answer I had always seen looking at the chip models is that there is an off-chip capacitance value below which it does not make sense to use because the bond wires effectively isolate the chip above certain frequencies (i.e., while smaller value capacitors have higher SRF it doesn't matter because the chip can't "see" the capacitor due to the bond wires screening it out).
If you knew where that roll-off was, and you knew the curve of your capacitors + board parasitics, you'd place a cap as close as you can to the chip in the frequency band right below the roll off of the bond wires to prop up that zone. Then, you'd place larger caps farther away because, as the article notes, the inductance goes up but also you're just looking to prop up the higher impedance-at-lower-frequencies curve of the tiny cap that's close to the chip.
So a lot of it depends on the exact chip you're working with and how well designed it is. A classic chip design team would have an expert who did all the parasitic modeling of the package, board, and then they'd do a noise analysis on the chip and recommend a minimum on-chip decap so the board designers don't have to worry too much, they can get away with "almost anything" in the 0.1uF range. Unfortunately chip design teams are getting leaner and leaner these days and I don't see the same level of care being put into chips. I think we more or less get away with it because there is so much margin in the chip timing; also modern chips are "mostly" (>50%) fill cells -- e.g. decoupling capacitors -- that are placed right up against the logic gates so you can get away with bloody murder on the package and power distribution networks (background: modern chips are wiring-limited, not transistor-limited, but for process stability reasons you still need to make transistors everywhere at a uniform density, so they instantiate dummy transistors that are wired as capacitors between power and ground).
Where it really starts to matter is if you had e.g. a PLL and you're trying to reduce noise that the loop filter can't get rid of and in those cases often times you need much smaller capacitors because they have much better performance at higher frequencies. Yes, they suck at low frequencies - but your noise problem isn't in the 1MHz band anyways; the loop filter can track that out. It's going to be in the 100MHz+ range.
And as someone noted elsewhere, in-rush current is a real problem, and too much capacitance can cause a problem for regulator stability; especially the extremely high performance ceramics. And, if you're doing an extremely power efficient design you may need to consider factors like leakage and losses due to CV-energy cycling if you shut down significant portions of the design when not in use.
(edits for clarity)
The amount of leakage is on the high side, so it does point to some device that is wired to the standby power that is probably 'powering' the 3.3v line via some ESD network between the power rail and pads. this is usually indicative of a part that was not chosen correctly for the job, or the correct part was wired to the wrong power bus.
If I were the designer I'd put the resistor in, power down the system, get a cup of coffee, come back and image the system with a thermal imager and the problem chip will be the 'hot one' (slightly warm, if you're unlucky its hot enough to just feel with your finger).
2 watts static current is nothing for a beefy PC motherboard but it will need a special heat sunk resistor which is annoying. The bigger problem is whatever is supplying the leakage is possibly an i/o pad and eventually that thing will fry if you keep pulling power from it (but from the description this feels possibly like a buck regulator that is 'off' and just leaking current, in which case it's fine to draw that amount of power).
It's one of those annoying things where it's hard to say who did what wrong. On the one hand, you could say the motherboard is out of spec by not bringing 3.3v to 0; but in reality 'we all know' that at power off things always hang out around 0.7v (a diode drop above ground) for a very long time without active countermeasures. on the other hand you can say the drive is out of spec for not going into reset when 3.3v goes out of regulation. I haven't looked at the atx spec specifically so maybe this edge case is specified, but when you are doing a fully integrated design this is just one of the things you learn to check for in integration testing.
A dirty fix could be to just put a static load on the 3.3v line to ground. I'd start with a 1/4w resistor around 100 ohms and just stick it from 3.3v to ground to see if that does not soak up the stray current. if it works just leave it, it's about 0.1 watts of static power and no big deal for a non portable setup.
The larger picture is that the controller on the nvme might not hit its power on reset condition because it may be rated to run at 1.8v (just a guess), so 3.3v is not going low enough for the controller to perceive the system has been power cycled. Usually a supplemental power monitor is needed in those cases to ensure a reset is generated in case of leakage problems like this.
The intended answer was to simply write the equation for ohm's law: I=V/R, P=IV => 5 volts / 1 ohm = 5A, P = 5A*5V = 25 watts (or something like that). Several students wrote their response in this way: "To find how many watts of heat are generated given a voltage and a resistance, you need to use ohm's law which is I=V/R. ... (about a paragraph of hand written text) ... in summary, the heat generated by a voltage source of 5 volts across a resistor of 1 ohm would be 25 watts"
I would have thought that the students were sneaking their phones in and copying results straight from chatGPT except she said the devices were left at the door and it was a strict pencil and paper test. It's remarkable to me how much text they wrote to answer such a simple question, and how much their impression of a response to a question is modeled after what ChatGPT does...
We couldn't figure out how to monetize it, but that's alright with me; not all returns on investment are measurable in dollars.
I didn't always have a rep -- the first step is the hardest in any journey, and the criticism was just as strong (if not stronger) back then. But, I was also a hardware guy in a software world. I'm stubborn, idealistic, and a rule-breaker, so I didn't fit in to any traditional corporate roles; no boss could manage me. That made it easier to walk out the door.
The most bold critics were always VCs -- not ones that I met with because I needed investment or anything, just folks I would run into at conferences or do an odd job for as a mercenary. You'd wrap up the job, and they'd make some quip about how it's a shame I'm wasting my talent in a "lifestyle business" and how I should consider finding some honest work at one of their companies. It was like some sort of weird negging tactic. I eventually learned to shrug it off but really, thought I was the only one who heard that term applied so pejoratively.
Thankfully, I had a few lucky breaks, and I'm very happy about where I'm at. However, it helps that I found residence in a place with affordable public health care, a functional pension system, oodles of public housing, and low taxes. No way could I do what I'm doing and also be so chill about my future without a functioning social safety net. Unfortunately, "just move somewhere that suits your needs" is not scalable advice; I wish I had a more practical blueprint for others who want to do a lifestyle business, but I don't.
If they encourage a thousand lifestyle business bloom, one might turn into a boot-strapped "unicorn" that didn't take their money, so it's in their own interest to make lifestyle businesses seem like a terrible idea.
The more talented people are convinced the only way to start a business is with VC, the higher a value VCs can extract from talented people. If we actually had more socially acceptable options, more folks would have the leverage to walk away from the table and shove the VC's non-competes, non-disclosures, preferred shares and meddling board seats in their face.
I always assumed it was just me being overly sensitive to folks expressing concern over my life choices, urging me to snap out of my "mid-life crisis" and do something "useful" like raise money and scale up, or "apply myself" at famous institutions with "real impact".
Thus the GPU is storing a 110M model in gigabytes of external RAM, and paying the power penalty associated with the excess capacity, while the chosen model 110M fits neatly within the FPGA's on chip RAM, and the design can trim all that overhead accordingly.
A more fair comparison would either run a larger model that had both systems hitting external RAM, or they would compare power/performance against some sort of inference ASIC that had all the RAM on chip (maybe a cerebrus, but scaled according to the portion of the wafer actually used for the model).
That being said, it's neat that they open sourced their work, and it's worth looking down this path more.
Probably the next thing to do is to put a diffuser on the LEDs to improve the uniformity of lighting. I think some of the hot-spotting has to do with the radiation pattern of the LED itself, if you just look at it on a blank sheet you can see a bit of a halo on the pattern.
Thus to get a high resolution of individual slices, you have to do something like ptychograpy or CT scanning, where you move the light source around to get a better idea of what's doing the absorption. These types of scanners are substantially more expensive than a dental X-ray.
That being said, the back side power delivery stuff that is currently in the pipe for the sub-"2nm" nodes would block viewing the transistors.
In the end, the actual LEDs themselves are tiny; a PCB containing the LEDs + driver ICs is smaller, lighter, and brighter than any mirror/lens assembly I could easily source with the requisite coatings to operate efficiently in this frequency band.
For a laser source, a fiber optic source can make a lot of sense, but fiber optics optimized for this frequency of light turn out to be pretty expensive, and the termination of the fiber optic still needs a collimator lens. At the end of the day, the net weight and cost budget still works out in favor of mounting a laser diode directly on the microscope head, so long as the power requirement is under a watt. Beyond that, the weight of the heat sink starts to be a factor, and an off-board laser starts to make more sense.
This package type is fairly popular in mobile phones and many types of CPUs.
The technique does not see through regular black plastic IC packages.
(you should also check out Charles Papon's webinar tomorrow, also listed at the link above, if you're into RISC-V or open source hardware!)
Whereas in the 2018 crisis, junk debt was rolled into CDOs but there were no 'knobs' the bankers could turn to actually improve the fundamentals of the junk debt. By repackaging the debt they did not get any right to fix the bad actors that made it bad. Things were actually bad, and when it defaulted there was insufficient firewall to prevent the contagion from affecting the encapsulating derivative.
Of course it is up to the maintainer but by absorbing an abandoned project, they are actually reducing the amount of unmaintained code, versus simply hiding it.
Probably the more accurate analogy to 2018 would be if a private equity firm started a company you could pay to upgrade the rating of an unmaintained Rust package, but no actual change happened and all they did was swap labels on things until the problematic rating disappeared.
And your eyes are both covered so there isn't a good way to non-verbally communicate waning interest levels...I suppose a solution is to simply care less if I hurt my friend's feelings but I'd also like a way to spend time with friends without feeling trapped.
At least in an f2f or video call meeting that I'm bored of, I can zone out or look at my phone or tap on my laptop, or do anything but stare at the slides. With eye tracking, the headset knows (and presumably everyone else could know as well) when you're tuned out.
The eye tracking thing also kind of weirds me out from a privacy standpoint. It's already bad enough that web pages track how long you engage with different portions of content. Now they know what parts of images I stare at, and can algorithmically feed me content based solely on my gaze alone. Does that prospect not weird anyone else out? Or are most normal users like "plug me into AR TikTok, but with gaze mechanics now!"
The lack of X-prop in verilator is one of the reasons why a lot of OSS hardware designs are risky or unsuitable for ASIC integration... verilator is popular in OSS, and X-prop isn't a problem for FPGA because every bit in an FPGA is guaranteed to start at a known state.
So, Verilator assuming 0 by default instead of X means that a lot of OSS developers get away with bad coding habits that work perfectly in FPGAs, but can't reliably power up in an ASIC.
Most of the fancier schemes take advantage of the fact that traditional binary signalling has excess noise margin, ie, they were throwing away energy to start with, and they are encoding extra bits in the energy budget. But to maintain noise margins as you cram more bits in, you have to up the amount of energy per bit.
The other half is that the physical layer implementation that does the encoding and decoding consumes more energy because it's doing more computation. This also figures into the energy per bit metric if you're being honest about your comparisons (and because it is not always clear if this is included in a metric you find papers where people cherry pick numbers to make their case). This number can become quite big because the baseline of a binary tx/rx is so low compared to doing effectively a DAC/ADC and phase recovery system.
What you find is that QAM or more schemes are certainly possible, but they can consume more power than the CPU just to keep the link idle and trained. The real art is picking the implementation and developing new circuit tricks that we hadn't thought of before to wring a little more bandwidth without killing the power budget.
Which leads to the corollary that, over time, all good engineers are eliminated from an organization (or are re-homed to a new position where they aren't as effective, until they do become effective, in which case they are re-homed again).
An alternative resolution of the paradox is that if every engineer solved their problem such that there were less problems than when they started, there would eventually be no more need for engineers.
The PI is Rosalind A. Coleman. I did not thoroughly investigate the preferred pronouns of this individual but canonically Rosalind would be assumed a 'she' and the top Google images returned for the name present as female.