The way that Jensen Huang runs Nvidia: 40 direct reports, no 1:1s
twitter.com
twitter.com
Bad CEOs will micromanage or only look mile-high.
Good CEOs will do deep dives, and use that to inform strategy, policy, and organizational design.
I can't say the same about my own a decade ago. it was due to the employer I was with at the time. they valued "attendance" and "billable hours" way more than quality or even meeting deadlines, let alone talking tech execution with a product manager
Everyone lies to leaders: https://www.rubick.com/everyone-lies-to-leaders/
They were gone 18 months later and only after the golden parachute had landed it turns out in addition to a multi million dollar salary they were padding the expense / travel account / banging their secretary.
It shows just how deeply imposters have infiltrated the industry. Somehow it has become conventional wisdom that people who are non-technical, or below average engineers, should manage the time of, and mentor engineers often much smarter, and more competent than them. That doesn't clear the bar of common sense.
There is value in having a hierarchy, some people are more competent, and better at planning, and time management, and interfacing with other people. Those people should be in charge of how other people spend their time. The confluence of those skills is rare. There are exponentially fewer people qualified to fill those roles than there are qualified to be engineers.
You shouldn't have managers unless you can find good ones. It is far better to spread management thin than to hire bad managers. You would never try to bring the quality of work, planning, and time management closer to that of a below average engineer, but that's exactly what happens when you bring in incompetent management.
It looks like Nvidia is doing it right. 40 is a high branching factor, and he can engage meaningfully with every one of them about their work.
It’s not true that NVIDIA doesn’t value professional development. It is true that Jensen is a special leader. He founded NVIDIA 30 years ago. No one leading a big tech company has the experience Jensen has. The whole company benefits from his vision. He is not always right, but he does learn fast and he works harder than anyone I know.
The moment you have significant turnover of these leaders (any overall as a result of flagging growth) you are in reorgs territory
Cycle: - Great Leader arises from a group of aligned leaders, more equal than his peers - Competing non aligned leaders are purged, creating strong execution momentum - The fellow leaders start hiring managers to take care of the day to day - Eventually leaders churn and managers take their place because they know the organization and finding aligned leaders is close to impossible. - Rinse and repeat - Eventually the Great leader (if he does not expire / retire) is surrounded by people who know how to run the organization within the box, but have never been allowed to think out of the box. All their success was in fact knowing how to maximize the inner workings of the box and not rocking the boat. - Leader leaves, game of thrones ensues.
This is in full display in most political parties. Take Germany or the UK: the qualifications to become leader are entirely different from the skillset to run the country (Hence the endless supply of prime ministers in the Uk who seem increasingly mediocre or downright incompetent). They absolutely know how to play the game to take the throne (chiefly loyalty to the cause) - but once on the throne, having to lead, requires entirely different skills they don’t possess.
The only interesting part is Nvidia getting their bets right repeatedly & the why behind that is likely interesting. Is lack of 1:1s the secret recipe…doubtful
I joined 17 years ago to work for "a startup within the company". We were ramping up our group fast to design a product for an new market. 4 months later, Jensen sees that product at CES. Ours would be only be ready 2 years after that. The next Monday, he joined our weekly meeting and announced that our group was cancelled immediately and we'd be spread to existing BUs the next day. His words at that meeting: "Every morning when I wake up, I look in the mirror and check if the core assumptions on which I based my decisions are still valid. If not, I make the change immediately."
So refreshing when you've come from a company where everyone knew a project was doomed, but nobody had the guts to pull the plug for months.
Same thing happened a year later: we had a chip that was ready to tape out. That project was canceled on January 10th, 2007, because something happened the day before.
With the iPhone, an SOC took center stage and the modem became secondary.
After the iPhone announcement, everybody realized that this model was a thing of the past. iPhone like phones were the future.
The companion chip market died that day, and the whole industry switched to developing only SOC models.
Is there some place I could read more about this shift?
Thanks in advance!
And then the person who pulls the plug is blamed for it somehow.
Nvidia never went into cryptocurrency. Cryptocurrencies went into Nvidia. Nvidia actively tried to sell GPUs to gamers instead of crypto people because crypto people will flood the used GPU market when the bubble inevitably pops every single time.
Nvidia isn't abandoning the consumer GPU market. They're still the leader by far. The consumer GPU market has actually been declining for nearly 20 years now.[0] They should not invest more into a declining market.
AI is the correct market to put their eggs in.
[0]https://cdn.mos.cms.futurecdn.net/osFsEsoUthvCrzKiF5aoLn-120...
They are charging non-competitive prices for their consumer GPU market. This is opening the door to more competition.
The desktop discrete video market is not the same thing as the consumer GPU market. You are ignoring a huge sector of integrated GPUs and mobile GPUs.
They crippled normal GPUs and sold them to miners. They did this because once the crypto bubble pops, the used GPU market won't be flooded.
>They are charging non-competitive prices for their consumer GPU market. This is opening the door to more competition.
Ok, so what? All companies do this when they're the leader.
>The desktop discrete video market is not the same thing as the consumer GPU market. You are ignoring a huge sector of integrated GPUs and mobile GPUs.
Nvidia does focus on laptop GPUs. It's much bigger than desktop discrete GPUs. Nvidia doesn't have a way to get into phone GPUs because they're tightly controlled by Apple, Qualcomm, and Mediatek. That's why they tried to buy ARM.
Their leadership is going to be transitory since they are not being competitive. Intel would never had entered the discrete GPU market if nVidia keep prices low. Lack of mobile GPUs and integrated GPUs is a red flag that they're charging too much money. Discrete laptop GPUs is a niche of the market. It is not what most laptops use.
1. Low end GPUs have extremely small profit margins. It's smart business decision to not lower prices in a declining discrete GPU market. Business 101.
2. Nvidia did not sell GPUs directly to miners. They had no control/no way of verifying what the GPUs will be used for. And who cares if they sold to miners? It's just business at the end of the day - free market.
>Their leadership is going to be transitory since they are not being competitive. Intel would never had entered the discrete GPU market if nVidia keep prices low.
Their leadership is extremely competitive.
Nvidia prices GPUs at the price the market is willing to pay. It has nothing to do with "greed" that you seem to be suggesting. Business 101.
Intel needs to enter the GPU market because AI runs on GPUs as well. It's both gaming and AI. Intel needs to be in the AI market or they're toast.
>Lack of mobile GPUs and integrated GPUs is a red flag that they're charging too much money. Discrete laptop GPUs is a niche of the market. It is not what most laptops use.
No clue what you're talking about. Cite the source for lack of mobile GPUs? You mean laptop GPUs right? Nvidia doesn't make SoCs for phones or laptops.
Discrete laptop GPUs outsell discrete desktop GPUs. While it's not what most laptops use, it's still a much bigger than desktop GPUs.
2) They did, and more importantly did not try to stop it from happening. They could have banned bulk sales or whatever. Through inaction, they created huge shortages.
At the end of the day, their advantage is really via a few things: software support and somewhat more advanced technology. Few people really cares about the actual hardware behind their gaming rigs. In places like game consoles, nVidia already lost a lot of ground because they weren't willing to cut costs.
The vast majority of laptops have integrated GPUs. I don't have exact numbers for laptops, but for PCs in general, it is integrated graphics: https://www.pcgamer.com/intel-is-already-matching-amd-for-ga...
is it that mirror from snow white or does he base his decisions on whether he was sunburnt or not the previous day?
I could see a flat, engineer-focused hierarchy enabling that direction.
Look at Facebook before and after the ad money dried up. [1] then [2]
It's all fun and games until the investors aren't happy.
Nvidia is enjoying a strong cash influx of demand due in no small part to a certain blockchain boom. Now they are capitalizing on the ML surge (as they should!).
Where will that management style be in a few years? Let's stay tuned.
[1] https://www.fastcompany.com/3003771/how-mark-zuckerberg-and-... [2] https://theconversation.com/why-metas-embrace-of-a-flat-mana...
This is a pretty recent phenomenon, isn't it? For most of it's existence net margin was 0-15%, and only blew up to 25-35% with the push into data center around ~2016.
1) GPU intensive gaming
2) Crypto
3) AI
Regardless of margin % that's a stupidly good read on the future hotness. Luck maybe...but 3x luck idk
For sure. Was just pointing out that "money printing", which I could classify as disproportionately profitable, is more recent. As a stockholder I'm perfectly happy with their increase in CAGR over the last few years. If they made this many good bets with the income they had to invest in R&D, I'm hopeful that they can scale with larger budgets.
1:1s are problematic because they discourage/displace lateral communication between team members and encourage siloing. Everyone has to guess what 'the boss' said to the others, and to figure out which piece of the puzzle they've been entrusted with - classic oppressive divde and conquer strategy. Of course for people who may not be comfortable in social settings, 1:1s may be the only thing they can grapple with properly.
This isn't an "American" thing, that's a very inaccurate generalisation. Maybe more correct to say that 1:1s are useful in cultures where trust needs to be built before saying difficult things. Sure, people from some cultures can just bluntly say difficult things. That's fine. Many people cannot.
Both of which abound, in my experience. If your "boss" is paying you as an "employee", that's still a competitive situation between two businesses: the business of you selling your labour, and your boss, a.k.a. your only customer, who may try to use information warfare to get the best deal.
Basically, I try to engineer 1:1s in my favor if I can - usually only upside opportunity in my case. If I can't, I would probably avoid them.
In my experience, group discussions favor a very small subset of people, and it has less to do with comfort in social settings and a lot more with domineering personalities.
Most of us are "trained" not to "bother" our managers so allocating some dedicated time is valuable in my experience.
I'm honestly baffled at the pushback against 1:1s I'm reading. Assuming you have a relatively normal number of direct reports for software engineering (5-10) that adds up to not a lot of hours.
At companies without regular 1:1s, I often had no idea if my manager thought I was doing a good job or not. You can say that this is bad management or bad communication, and you're right but management and communication tend to be rather poor in our industry! 1:1s are more often than not a part of the solution.
I think if that kind of transparency at the team was put in place and people got used to it, I think you're right, it kind of aligns the whole team on everyone's goals, and everyone knows where everybody is and who's where in their career with regards to promotion and what not. And who the manager think should be given what opportunity and responsibility next, etc.
I had regular 1:1s when I worked for Huawei.
Then those longform discussions are summarised and points raised higher up the chain as needed.
At literally every company I've worked in Sweden the managers have sat down with everybody directly under them for a 1:1 twice a year.
Where you work are salary, promotion and general career development discussions done in open meetings with the whole team?
No status reports, instead he "stochastically samples the system"
- Doesn't use status updates because he believes they are too refined by the
time they get to him. They are not ground truth anymore.
- Instead, anyone in the company can email him their "top five things" with
whatever is top of mind, and he will read it
- Estimates he reads 100 of these everyone morning
I'm a software guy who for 1.5 years worked for a hardware company with a relatively strict upwards-only information reporting structure. It wasn't even for secrecy, it's just that the people running it didn't understand the concept of collaboration at all. They were a mix of physicists and economists. Really smart but also, as it turned out, very 'square' people. It was a relatively small company, around 100 people.
It felt quite soul-crushing after a busy week to have to spend mental efforts writing up a report to someone 1-3 levels up... especially if you had no idea if it was ever read. And then do this every single week. If colleagues could read and learn from this effort it would make a lot more sense.
Never again. Went back to the relatively speaking much saner software world after that experience.
There are ways to test that.
You could just not submit the report, and see if anybody complains.
Or prepare the report, but intentionally corrupt the file before submitting it, and see if anybody complains.
Or prepare the report, and in it ask if anyone is actually reading it, and ask that they contact you if they are.
I admit I never had the guts to submit a slide which asked if anyone was reading it. It did cross my mind.
(It was a Microsoft-by-default-company, as many small hardware/research-oriented companies often are, I guess.
There was also that one week epic fight needed to get approval to run MacOS/Linux in the new "software department".)
Facts. I worked in a place where one my primary work outputs was a weekly high-stakes status document, and boy this document was refined, and agonized over, and wordsmithed, and revised, and re-revised, and softened, and distilled, and lenghtened, and shortened, until it was precisely the exact narrative that the VP and SVP needed to hear. By the time it went through the hands of dozens of engineering managers and was submitted, it was far, far divorced from reality. Then, rinse and repeat next week. I don't know if/how leadership had a clue what was really happening.
Reminds me of a friend who didn't wear socks because Einstein didn't wear socks. The implication being "he doesn't wear socks and he's smart, so if I don't wear socks I'll be closer to being as smart as Einstein". Ok, we were like 8 years old, but still.
there are very very few industry-defining CEOs who are at the helm of the companies they founded 30 years ago and leading an industry that they essentially designed and built themselves, riding a money-train on an entire class of product and a product-use case they defined from the ground up. the fact that huang has done this multiple times means he's worth listening to. he's managed to run a (now) trillion-dollar corporation like an agile startup and maintained an extreme level of technical excellence and long-term realized customer value (every CUDA testimonial ever, etc), and that's cause he can run his OODA loop faster than anyone else and actually get good products out the end.
no, you don't have to wear the turtleneck or leather jacket and yell at underlings. yes, absolutely you should think about evidence-driven leadership, being agile and willing to pivot and throw away your product and move to something else if it's not going to work, eschewing formalized "busywork" 1:1s in preference for having the CEO actually "stochastically sampling" the IC work output, and making sure the voices from the boots on the ground are being heard.
there is almost no tech company for which those are not great principles that you should probably strive to incorporate. and "wow we have nothing in common with that" probably speaks more about your corporate culture than you think.
> I'm not saying these are good or bad. But I note that every time a company becomes "hot", people start mindlessly copying its quirks,
Also, nowhere did I say I work at a tech company, I do not.
I get a lot of cold emails from my website, asking me things. I ask those corresponding with me to ask the questions on my BBS instead, so when I take the time to answer, the answer benefits more than one person, and there is increased ROI on the time and energy spent constructing a thoughtful reply.
I honestly believe there was an entire generation of young managers who thought that it was OK to wear turtlenecks every day and be an abrasive asshole.
Things I liked about his stuff are the idea that every management position is a growth position. Some are, and in some you're just ensuring that superstars can play the right way. And you can land where you want on that.
This is also true of successful people and their secrets to success. With survival bias, it's hard to differentiate noise from cause
Yes, there may not be company-wide top-down 5-year plans or planning cycles. Don't mistake this to be your usual breed of agile. What folks work on are ultimately much longer-term visions. I think everyone sort of has/owns their own ideas and plans for how an area, product/service, project, whatever, should grow into, and everyone is asked to make an effort in communicating those.
Internally, it's one giant networked negotiation machine.
The company adapts well to changing circumstances and market needs, but NVIDIA is strategic and thoughtful - once it goes after something, it rarely backs off. And usually, NVIDIA is going after things others are actively avoiding because the problems in those spaces are too difficult. The consequence of such problem sets are visions that 10-20 year undertakings, if not more, that everyone chips away at bit by bit until one fine day things start taking solid shape.
This is also made possible because everyone up the management chain, from ICs till up to JHH are technically adept. This is why sending out "top-5 things" works; not because it's written for everyone, but because everyone can read it. These emails aren't trimmed down. They are sent regularly. They can act as status updates but that's not their intention; they are literally what the title says "top five things" (on a team's agenda right now).
A key relevant item that's left out in the tweet but JHH mentions in the referenced interview, is "strategy". Leadership, from the bottom-up ensures everyone's going in the right direction for a body of work in a given area. All of which funnel up to a general strategy that NVIDIA executes.
Wouldn’t in large meetings mostly the loudest and most political be heard? How do you hear everyone in a large meeting? Who takes credit, assumes responsibility, be more or less paid, etc?
1:1 are important, at least to connect to people.
Nvidia bet on a good fast growing market early on (AI). Maybe that explains their success. Also, take these stories with a grain of salt, til the sources are clear.
1. CUDA. They noticed that people were using graphics shaders to do compute (GPGPU). This was extremely (really extremely) painful to program in. They created CUDA which is only somewhat painful to program in. Now, at this point, it wasn't that clear that GPU compute was a great market, but it opened up a bunch of opportunities.
2. After CUDA came out, Alex Krizhevsky came around and used it to implement a neural network. Now, this wasn't the first time someone had made a neural network with a GPU, but CUDA made it a lot easier to make a sophisticated one.
3. NVIDIA noticed that people were using CUDA to make neural networks, and created CuDNN and started sending out free GPUs to all of the academic researchers they could find, further cementing their lead in this area over AMD.
In big corporations it tends to be hard for some seemingly insignificant thing like people using your product for something it wasn't intended for to bubble up to top management where resources can be allocated to capitalize on it. It's especially hard to do this quickly. I wouldn't be surprised if this habit of Jensen's was responsible for NVIDIA being able to take advantage of these opportunities.
Problems with meetings: people don’t express themselves because they might offend others or create enemies, people follow each other, no defined rules on how the meeting operates, problems with credit, diluted responsibilities, opportunities for shifting the work and credit, as someone said in comments formation of cliques competing, etc.
Both are probably needed. You need to mix.
Sampling group emails likely has a similar bias to sampling large meetings.
Emails being push rather than pull means you can't self-filter as easily - something more like a company forum may be better than email groups as it finer grained categorization than emails, while still being visible to anyone if there ends up being useful discussions between groups which wouldn't be the case if you had a thousand different email groups.
Nvidia is a /huge/ company doing a very large number of different things, there's probably not a lot that's actually relevant to everyone. Lots of people think what they say may be, often the "louder" people I mentioned, but it's probably not actually true.
Generally I find emails are a great place for 1-on-1 communication, OK for 2-3 people discussing things, but very quickly spiral into a mess if there's any discussion by a group - things get lost, threads get mixed up based on who happened to respond first, and taking things "off thread" exclude people based on arbitrary criteria of the responder anyway and are often invisible to anyone else.
Maybe OK for "declarations" that aren't actually for discussion, but that doesn't fit with the idea of a "flat hierarchy" engineer-driven workplace either.
> Emails being push rather than pull means you can't self-filter as easily - something more like a company forum may be better than email groups as it finer grained categorization than emails, while still being visible to anyone if there ends up being useful discussions between groups which wouldn't be the case if you had a thousand different email groups.
I mean its stated fairly clearly that CEO reads these emails or rather randomly picks some to read every day.
It seems you're arguing based on some hypothetical view of Nvidia divorced from what the linked text and other comments state Nvidia actually runs like. I don't think it's fruitful to have a discussion where you simply make up something in your head and then argue based on that.
It's going to happen.
At these "open" nVidia meetings, I'm sure cliques are texting on MSFT Teams or SMS.
I'm skeptical that 1:1 here means no 1:1 contact and rather no recurring meetings with reports.
The top-5 mailing lists seem like the way that is achieved. My personal experience is that as a line manager at a company I actually know a ton more about who is doing well across departments than the CEO probably does. My view is direct and unmoderated while the CEOs is not.
>1:1 are important, at least to connect to people.
Why is connecting to people important?
No. Early on they realized that supporting developers for GPGPU is important, and many industries have been well served by it, starting with computer graphics and gaming.
I don’t believe that. There’s really no confidential matters he discusses with team members on a 1-on-1 basis?
There is no thing as so confidential that only 2 people need to know it.
That seems ludicrous at any company I've worked for absent some serious HR-related issue.
It also fostered a lot of trust and cohesion. Everyone knew they were there because they deserved to be there, and so it made it very easy to make fast, correct decisions without a lot of bullshit. They were a great group and I felt lucky to have them on the journey.
Executives are expected to be mature enough to not need hand holding.
I think explicit PD makes sense at a big corporation or even a medium sized firm, but not in a startup environment where every second counts.
If you do not do Professional Development, and make this clear to your hires whom you expect to be pretty good at their job functions already, I do not see how it should be treated as cultivating yes-men. 'I don't do Professional Development' is also not the same as 'I can't take feedback or let someone call a spade a spade'.
No, he simply claims the team was already competent enough to execute at the level required and without egos weighting them down.
The culture we needed in that moment was one of high standards, of directness and honesty, of pressure to do the best work. My message fit the necessary culture, and the outcome was exactly the team I had hoped to shape. I still remain friends with these folks to this day.
We were hardly self-congratulatory; quite the opposite, we were completely without ego and laser-focused on doing right by the company and our teams, and I’m proud to this day of how we ran things.
When did that change for you?
What's in it for me?
What's on the table here?
This to me sounds like a culture of no support for technical staff, no listening, and helicopter management when things go wrong.
And no training budget.
> and no training budget
You’re just speculating. And this GP was referring to his executive reports, not the ICs.
You are talking in a past tense, and I can’t help but wonder about the reason for that.
What I see described here is the exact opposite of what I witnessed at "the greatest management training company in the world." What a quagmire.
They had the CEO from Cinnabon come in to give a talk and she laid down what sounded to me like a dress down of the management culture at GE....you could hear the whooshing jet noises as it went right over their heads.
I haven't watched https://www.youtube.com/watch?v=h5xY_kRKHxE but perhaps someone will be interested enough to.
The old-school SV management style of actually being a leader is lost these days.
You're wrong. Not a single piece of working software (literally zero) was shipped before a combination of Leetcode + agile + performance reviews + feature factory culture + microservices was invented in the last decade.
It's simple, you're either part of the group or you're not, there's no "I" in group. I don't see past great leaders (military and not only) doing 1:1s with the members of their groups.
The modern 'open space' office makes it impossible because it is effectively a panopticon where everyone is in full view and hearing distance of others all the time so effort has to be made to schedule 1:1s and that may seem forced and contrived.
When one had to go to their manager's private office to report of discuss something it was easier to exchange in confidence without any specific planning.
My personal experience is that 1:1s are very useful not just as an opportunity to raise and discuss concerns but to create rapport and talk shop openly and informally.
You have a relationship with your manager.
You have an optional relationship with a mentor.
There's no reason for both to be the same person. In all probability, your manager might not even have the skills to be a good mentor.
And they certainly don't have aligned incentives while also people-managing you.
I think no matter what structure you put in place as a manager you still have to get to know your reports well enough to know what you can and can't confidently ask them to do.
Not to mention, you also get to just talk nonsense and get to know each other as humans a bit. Some folks enjoy that, me included.
Even a chat exchange of ideas or mails qualify as a 1:1.
You don't think Titus Labienus discussed anything one on one with Julius Caesar? That just sounds like insanity.
I want 1:1s, it's the best way that I can make my boss work for me. That stuff gets me raises and promotions.
And back to your question, no, I don’t think that Zhukov or Napoleon had 1:1s with their second people in command, I don’t think they were laying down advancement opportunities nore were those people asked about how do they feel and what’s troubling them about the team (the army in this case).
Notably, when you have trust, your reports will tell you what’s actually wrong.