Intel plans spinoff of FPGA unit
networkworld.com
networkworld.com
It also dampens their "US developed" technology pitch (which they had been pushing pretty hard vs other solutions that were fabbed at TSMC) I wonder if they will also give up on their third party use of their fabs.
Intel was the first "real" job I ever had and it was during Andy Grove's tenure as CEO and his "paranoid" attitude was really pervasive. I had been working on graphics chip which were the next big thing until the chip recession and then they weren't. And it followed a long series of Intel sending out tendrils outside its core microcomputer (eventually x86/x51 only) winners only to snap them back at the first sign of challenges.
I wonder if we can get better open source tool support from the new entity. That would be a win.
It's really quite impressive how bad Intel's track record has been. Would you say it's a broad cultural problem within Intel? Bad incentives?
(I'b be interested if someone could come up with a list of all of Intel's failed ventures. StrongARM, Optane, Altera... Obviously some of these are par for the course for such a large company as Intel, but Intel still seems to stand out)
(I know one reason Apple is so secretive about its ventures is because it knows many of them won't pan out, and it knows that due to its size anything it does has huge repercussions, and wants to avoid that)
It's a very tight market out there. TSMC was able to make it with very tight partnerships with OEMs and I doubt Intel can pull off something like that.
They have had lots of things they stopped doing, the graphics chip I was involved in (82786) was one, the speech chips, the iAPX431, the modem chips, their digital watch was an early one.
I always saw it as an exceptionally low tolerance for risk. When I was their the 386 had just fabbed and I was one of the people watching when the very first chip was powered up and it actually worked (which amazed pretty much everyone). Sun Microsystems was a little startup in Mountain View using the 68000 and getting some traction I suggested to Intel that if they used the graphic chip I was working on and the 386 they could build an Intel branded "Workstation" and both replace the VAXStations they were using for CAD work internally and sell them as part of Intel's initiative to become a "systems company." I got a response to my memo[1] from Dave House who was later CEO but at the time just the head of Microprocessor Operation (MIPO) Marketing explaining to this lowly NCG (new college grad) that Workstations were not much of a market and Intel was only interested in pursuing the "big" opportunities. Six months later I joined Sun (just after they IPO'd sadly) and didn't look back. :-)
[1] Yes memo, as in typed up and printed out and put in a manilla envelope where on the last line of the front had it coming from Dave and being addressed to me. And yes, it started as a memo I had printed out and given to my manager because while Intel had "heard" of electronic mail they preferred the "durability" of actual mail and forcing people to use actual mail cut down on "frivolous" communications.
Big laugh at the claim of intel having a exceptionally low tolerance to risk when they're the only semi company left still running their own fabs, decades after AMD sold off their own for being too risky of a venture to keep running by themselves. And every fab company out there will vouch for how risky that line of business is. Even more so to pour billions into a business like this while knowing Samsung/TSMC might beat you.
More correct would be that Intel has a low tolerance to risk for things that are way outside of their core competencies which have traditionally been designing X86 CPUs and fabbing them, and that's mostly it.
I commend them for not giving up on advancing their fab nodes despite the massive issues and setbacks(they could have spun it off like AMD did), and for building the ARC GPUs and keep improving them despite low sales.
He said this some time in the 80's to early 90's as a defence for their expensive commitment to still be running their own fabs when most big semi companies like Motorola were selling off their fabs and going fabless because it was eating their resources and it was impossible to compete with the likes of TSMC on node shrinks and value.
So even without Intel screwing with them in anti-competitive ways, they could not have kept their fabs competitive much longer. The writing was already on the wall way back then but Jerry Sanders was just being stubborn.
So I don't think it's as direct a link as you seem to claim. Maybe it was the right decision even if AMD had the money to keep it going. But it will likely forever be one of those questions that has no solid answer either way.
IMHO Ruiz cut back on engineering and increased marketing. Then purchased ATI which I though was an odd or even dumb move (it was not). But that expensive purchase along with the failure of bulldozer almost killed the company. That's why they spun off Global Foundries - to stay alive.
Yes and even within that they are risk averse, no need to look beyond 10nm and their reluctance to EUV.
So they are really risk averse and they just want their success to continue infinitely. In fact, contrary to Pat G claiming that Nvidia is lucky; it's really Intel which was lucky.
>> commend them for not giving up on advancing their fab nodes despite the massive issues and setbacks(they could have spun it off like AMD did), and for building the ARC GPUs
even if the turn around time in semi market is longer, the way they faltered on their Arc timelines and no encouraging signs of their new fab nodes is not commendable!
Sure, Intel fucked up along the way, and it's easy to point fingers and laugh from the comfort of the armchair at Intel tripping over massively complex engineering challenges, but if designing and building successful GPUs and sub-7nm fabs was easy, everyone would be doing it.
Further, neither do the facts nor my comment in any way implies building GPU's is easy. OTHO, when a company like Intel with their massive competitive advantage (when their fabs were the market leaders) thus leading to financial muscle fritter it away in what is supposedly their strength (fabs), don't you think that they in fact deserve to be laughed at?
Is it that unfortunate? Because if you knew how difficult semi fabrication or GPU design is, you probably wouldn't say these failures are such big deals worth pointing out as negatives, but part and parcel of the process of solving such complex problems only 3 players worldwide have mastered.
Do you see Nvidia trying to fab chips? Do you see TSMC trying to design GPUs? Imagine Intel is doing both at the same time. Of course they will fail at points along the way, the important thing is they stay in the game to provide us with competition.
Wasn't there an article on HN 2 days ago about Intel pushing 2 new technologies simultaneously in their next node? (ribbon gate and backside power?)
When I read "ribbon gate" just now, I assumed you were referring to some scandal.
Let's hope that's not prescient :)
Actually from what I heard, their failures with their Intel 7 was partly due to trying too many unproven technologies - i.e. they pulled a Zumwalt. With their current Intel 4 & 3 I heard they were more conservative.
Everything about bleeding edge semi fabbing is unproven technology.
What was more shocking is how long them seemed to be in denial before putting contingency plans in place.
As a former Intel-er myself (during Craig Barrett, Paul Otellini times) I would disagree. Intel took bold and foolish bets bying DiaLogic, selling Xscale, betting on WiMAX mobile technology while fighting LTE backed by all telcos. I would characterize Intel's risk management as "passive-aggressive".
I have wondered if leadership didn't understand the Gartner Hype Cycle, because a lot of projects seem to get killed after the investments but before significant success was even possible (the Trough of Disillusionment).
During my tenure I witnessed several failed initiatives. Itanium, Wimax, Digital Home, x86 phone (android), LTE modem. Imagine the billions wasted.
* They got their LTE modem into the iPhone 11 but just couldn't compete. * ARC GPUs would have been competitve if they had be launched a year or two sooner and its stagnancy is widley due to lack of funding since the market was dead during launch and we were in a massive market pullback where they were ultra hesitatnt to spend ANY money. (massive layoffs too) * Optane was a great product and had a good roadmap, but they were burning cash waiting for adoption and their customers had decoupled their memory and storage from the physical layer to mitigate limitations of NAND ssd so much that they didn't see the price as worth it, despite the obvious latency and endurance benefits that it had over even SLC NAND drives.
Sometimes they just miss the market for otherwise good products.
it wasn't "just couldn't compete", those modems were horrible. the speed was lower than the Qualcomm that managed to be shipped in some of the iphones (from 8-11), and the connections were extremely flaky. I had a 9 with an intel modem, and it was abysmal compared to the previous iPhones I had owned.
that apple went back to Qualcomm, bought the intel division, and still is shipping Qualcomm modems says a lot -- it's still just not there.
And this was even before being slower even with strong signals.
So slower and noticably worse range. Modems are hard and take a long time to have small things ironed out here and there. Small things add up.
I have a hard time believing that Intel will succeed at being a foundry this time around given how many failed prior attempts they have made. The foundry model is incompatible with the high margin duopoly CPU vendor model. What will prevent Intel from scheduling their own high margin products in fabs over foundry work from other companies? They will always have a business incentive to deprioritize their foundry customers when push comes to shove.
Reading some of the comments made by TSMC are enlightening. One of the reasons TSMC provided for their ability to succeed when Intel was flailing with their 10nm+++++ process is that TSMC runs experiments 24/7. This allows TSMC to run more experiments per day/week/month than Intel. I would really like to know when Intel starts running 3 shifts of process development engineers at a single fab to iterate faster, as that would be a true sign that meaningful change has occurred inside the beast.
I don’t care how well Intel executes and how well your launch market functions, if you’re selling a product that caters to business travelers and users, you had better actually be willing to sell the thing, across a large swath of the country, or it will fail regardless of how good it is.
Given the strong assumptions it was making about compile-time instruction scheduling, I'd have expected them to de-risk the project first. I.e., get a suitable compiler working, and get some benchmark results on a simulator.
"Anyway this chip architect guy is standing up in front of this group promising the moon and stars. And I finally put my hand up and said I just could not see how you're proposing to get to those kind of performance levels. And he said well we've got a simulation, and I thought Ah, ok. That shut me up for a little bit, but then something occurred to me and I interrupted him again. I said, wait I am sorry to derail this meeting. But how would you use a simulator if you don't have a compiler? He said, well that's true we don't have a compiler yet, so I hand assembled my simulations. I asked "How did you do thousands of line of code that way?" He said “No, I did 30 lines of code”. Flabbergasted, I said, "You're predicting the entire future of this architecture on 30 lines of hand generated code?" [chuckle], I said it just like that, I did not mean to be insulting but I was just thunderstruck. Andy Grove piped up and said "we are not here right now to reconsider the future of this effort, so let’s move on". I said "Okay, it's your money, if that's what you want." Suddenly this came up again later in another guise but again Andy shut me off, he said "we're not here to discuss it". Gordon Moore is sitting next to me and hasn’t said a word, he looks to all intents and purposes like he's asleep. He's got his eyes closed most of the time, you think okay, the guy's tired, he's old. But no, 20 minutes into this, he suddenly opens his eyes and he points to me and he asks, "did you get ever get an answer to your question?" and I said, "actually no, none that I can understand". Gordon looked around and says, "how are we planning to move ahead with this, if the answers don't make sense?" and this time Andy Grove said to him "We’re not here to discuss that, Gordon"."
Tragically, I wouldn't bet on it. In fact, I'd be shocked if that were to happen.
The whole thing definitely seems symptomatic of Intel's extreme caution. I wonder why they didn't apply the same caution when buying Barefoot. Intel obviously didn't have a great idea on how to leverage them.
This is all in stark contrast to AMD buying Pensando, a company that my SO works for. It makes sense for AMD to acquire Pensando to expand their data center offerings and compete directly with Intel. I think AMD made a smart buy.
> It still seems like a ridiculously foolish idea, considering how much money Intel spent acquiring Barefoot
Acquiring companies in order to kill them is a horrible business practice but not unusual. The thing that makes no sense to me is why they kept it going for three years before killing it. If they bought it in order to kill it, they should have done that before spending another hundred million on it.
At this point, if it's not about x86 cpus, listening to Intel's roadmaps seems foolish.
Internal politics?
Presumably the internal team(s) with overlap were against the acquisition.
But it was likely easier to see if the integration failed on its own before spending the political capital to kill it.
Folks forget executives at large companies usually optimize for "my career" over "the company."
So insane. It's always better in invent the next big thing that changes the market than a competitor, even if it kills another product of your own.
What do you mean by this? When working with a fab you get to know their design rules and that's about it.
The point being, there's no world where a fab would worry about sharing their design rules.
> Engineers needed manufacturing process information to verify that designs were free from systematic yield limiters, but foundries were reluctant to distribute process information that could reveal best practices and trade secrets. Encryption enabled the protection of selected design rules, process data and other information while still giving customers the details they needed.
> In this discussion, you must understand that the required protection is much more than just an obfuscation of rule definitions so they are no longer human readable. The process also includes the encryption of related files and control over all the outputs from the tool environment that provide information about any checks.
https://www.techdesignforums.com/practice/technique/protecti...
The design rules themselves could yield info to competitors about what processes are being used in the upcoming gen.
Unfortunately, this seems unlikely. The FPGA vendor tools are a complete shit show and open source tools would be greatly welcomed - they're doing really well in the Lattice/ICE40 space but those are small FPGAs. But Altera and Xilinx don't seem at all inclined to encourage the development of open source alternatives.
When it is almost impossible to develop for, you only get big contracts because nobody else has the resources to design products for it. On the other hand, with good and free design tools the toy projects done by hobbyists and schools can serve as the catalyst for using it in medium-scale projects.
There are plenty of applications imaginable for something like Intel's SmartNIC platform - but you're not going to see any of them unless tinkerers can get their hands on them.
Xilinx, for example, used to be very friendly about making really cheap or even free FPGAs available to academic users. This is because those users lead either directly to design wins (some academic designs go big, NSF style) or they become corporate users who are familiar with Xilinx gear.
But the software was crap, and they would have gotten more wins if the software wasn’t so awful.
"EDA software is crap" is so close to an axiom around here that I think Vivado doesn't get the credit it deserves. The synthesis flow has been rapidly modernizing and supports a chunk of VHDL-2019. The bundled simulator still lags on support and features, for reasons that are pretty easy to understand.
That’s too bad. I once got a small pile (8?) of max-spec Virtex-5 chips, for free, FedExed from Xilinx. Xilinx apparently valued them so little that they didn’t even bother putting the full address on the envelope. Tracking it down was fun.
The only remotely supportable way to synthesize images for them was to run a primitive CentOS 5 container on a big server — the Windows version of the tools were still 32-bit, and 4G of address space was too little for the synthesis workflow. Even with the magic 64-bit RHEL/CentOS-only build, it would take 45 minutes or so for the tools to notice any errors during synthesis.
Fun times.
Thanks for the stock tip. The size of a chip can change, and will. Stupidity really tends not to. Great dev tools are incredibly important, and the people using them actually can influence decisions on which parts to use.
Recent hindsight with TSMC makes it look "easy", but it's incredibly uncertain simply due to the complexity and number of pieces that have to align.
https://www.eejournal.com/article/intel-plans-to-spin-off-fp...
I remember finding it rather funny at the time that Altera would go "full circle" so quickly.
The amount of market I saw dissolve for Altera in a puff was incredible. To this day one of the most shortsighted corp decisions I have ever witnessed.
Intel is my daily driver for gaming, due to it's incredible (but sadly cancelled) Extreme Compute Element form factor with its incredible customization.
What they're struggling with is adapting to a changing world and to find new business centers. But this is a hardware/software world built on Intel, even if it changes daily.
https://www.pcmag.com/picks/the-best-cpus-for-gaming
For server marketshare, intel vastly dominates (80%?). For gaming marketshare, Intel still has 60-70% the marketshare.
Even if it's a mixed bag, if given an Intel (or AMD) processor, there's no way I would throw it in the trash and if a current-gen chip, you can find things that either are better than the other with.
Both are top-tier accomplishments. AMD has a signifigant fan-following though, which tends to distort the story a little.
> For server marketshare, intel vastly dominates (80%?). For gaming marketshare, Intel still has 60-70% the marketshare.
Marketshare is one thing, but saying they are the best but the marketshare is 60-70% is a totally different metric.
Its saying oh Toyota has the best cars because they ship the most... but they cant be as good as a Lambo / Ferrari because they dont ship the most.
In terms of server cpus, if you are looking at raw performance per socket, intel is no where near amd -- https://www.phoronix.com/review/intel-xeon-platinum-8592/10 . This is based on Emerald Rapids / Bergamo.
In terms of gaming cpus, they definetly are neck and neck with performance.
But yes you are right, Intels chips arent something you would throw away in the trash. They are good chips. They need to stick around. Its just that you cant say marketshare makes them the best.
That’s exactly where the Unix processors were the year before Intel put them out of business.
(Except the old Unix machines also won by large margins on multithreading and Intel isn’t winning there anymore).
No. They're definitely behind AMD. AMD has significantly better power efficiency[1], better performance in gaming for most of the titles[2], and vastly more cores and better multithreaded performance (with Threadrippers[3]).
[1]: https://gamersnexus.net/megacharts/cpu-power
[2]: https://gamersnexus.net/cpus/intels-300w-core-i9-14900k-cpu-...
[3]: https://gamersnexus.net/cpus/best-cpus-2023-intel-vs-amd-gam...
Intel bet on a fabrication process that did not succeed. I will bet in the next 2 years Intel will be on par with AMD again.
Why ?
For example, the chiplet approach that AMD embraced allowed AMD to put together far more powerful CPUs, i.e. with more cores and at a competitive price, than Intel could.
Compare to toms' hardware 2023 listing earlier in the year.
https://www.tomshardware.com/reviews/best-cpus,3986.html
Category | Winner | Alternate
Overall Best CPU for Gaming: Intel Core i5-13400 (Buy) [More] AMD Ryzen 5 7600 (Buy) | Ryzen 5 5600X3D
High Performance Value Best CPU for Gaming: AMD Ryzen 7 7800X3D (Buy) [More] Intel Core i7-14700K (Buy) | Ryzen 7 5800X3D (Buy)
Highest Performance Best CPU for Gaming: AMD Ryzen 9 7950X3D (Buy) [More] Intel Core i9-13900K (Buy)
Mid-Range Best CPU for Gaming: Intel Core i5-13600K (Buy) [More] AMD Ryzen 5 7600X (Buy)
Budget Best CPU for Gaming: Intel Core i3-12100F (Buy) [More] AMD Ryzen 5 5600 (Buy)
Entry-Level Best CPU for Gaming: AMD Ryzen 5 5600G (Buy)
This is a neck-and-neck race, with AMD a fan favorite, and year over year changes in leadership. No way is Intel out of this as a viable competitor of equal-ish standing. I don't mean to denegrate AMD in my original post, but I do mean to say that Intel is still producing top-tier results.
https://www.tomshardware.com/news/amd-and-intel-cpu-market-s... In particular, Intel still has a 5x lead on AMD in server market, which, of course, is a pretty big market.
> (summarizing: Intel best CPU, Best mid and budget entries, so winner in 3/5 categories. And what does it mean to give AMD "Best high performance" and "Highest performance?").
Intel as much more brand recognition than AMD. So calling AMD fan favorite is kinda strange to me. There are only two x86-64 chip designer right now, and from a pure market strategy it doesn't make sense for AMD to offer much more value to customer than what they can get from the only competition in town. If you want to really understand how intel is strugling you have to did into the execution speed of both companies, the profit margin on each SKU etc... etc...
Intel is only neck and neck with AMD if you don't look at things like power efficiency and how much money they are actually making per chip.
> In particular, Intel still has a 5x lead on AMD in server market, which, of course, is a pretty big market.
This a function of market inertia more than anything. From a technical perspective AMD Zen 4 workstation and server offering seem to be much better these days.
Gamersnexus is great! But they and the other YouTube benchmarkers don’t tell the whole story. when you are playing games that are bottlenecked by single thread performance single core performance and therefore memory speed becomes important. This is very difficult to benchmark as most of those games where it matters (CoD warzone) don’t have an easy way to benchmark. When I was testing my OC I literally would change the ram frequency in the bios, and drop into a pubg match, and screen record my fps and manually add it to a spreadsheet.
So take those benchmarks and then add some performance gains from memory overclocking that I (and others) been able to get meaningful gains in fps beyond just overclocking the cpu. 5-10% more fps is not amazing, but it’s definitely worth it to me for games like warzone or valorant.
AMD’s support for high end ram has always lagged. Happy to use amd for non-gaming, but if you’re trying to build “the best gaming rig” I’m not sure AMD can take that crown. And until people are willing to benchmark single threaded bottleneck games like warzone with different ram speeds, I doubt this argument will be settled.
The gap between those non-3D cache chips and the newer 3D-cache ones is generally very big for games because they touch huge amounts of data every frame and thrash memory.
[0] https://www.techspot.com/article/2579-cod-warzone-2-benchmar...
It's just spinoff and their people are at executive levels
Eh, people have been saying this for over a decade, and I think that opportunity has passed.
FPGAs are not going to beat GPUs anytime soon due to the software ecosystem (among other things), and ultimately they are not going to outrun ASICs that are now economically viable (especially in the embedded space).
Completely true as long as the chip companies stonewall open-source toolchains.
If Xilinx or Altera published the bitstream format for their highest-volume high-end chip (i.e. one or two generations back from the bleeding edge) you'd see the effect on the entire AI space within twelve months. I'm not kidding.
The interest in this kind of access from developers is enormous, and the problems in this space are extremely regular. There are massive opportunities for FPGAs to avoid spilling to DRAM or even to SRAM -- you have an ocean of tiny register-speed memories in LUTRAM mode.
But it will never happen. And so FPGAs will continue to be trinkets for weapons manufacturers and not much else.
Instead, FPGAs are sold to big entities like governments and big corps who can throw bodies at the crappy toolchains until they produce what they want.
I'm not an expert... But in my naiive opinion it seems entirely reasonable to expect an acquisition to start paying off within ten years?
Would love to hear counterexamples
It would be a lot more viable to use a chiplet approach: bake the FPGAs on a slightly-cheaper node, and just glue them right next to the CPU itself. Unfortunately Intel is still miles behind AMD when it comes to chiplets, and their tech is only just now barely starting to ship.
Till they fix their culture problem, they should be unloading these business units and profit by holding majority stakes in them.
Maybe we can look forward to all the ZIRP semiconductor consolidations to unwind.
Xilinx was fabless before the acquisition. I'm missing how that made less sense.
with all the zero interest rate money floating around
Hehe...as one of my finance-world pals said: "everyone's doing M&As like drunken sailors".
Intel marketing built the market of x86 server + tightly integrated FPGA. Large customers built products based on their promises. Then they didn't deliver.
Then AMD bought Xilinx and started shipping to the market.
To my mind the more interesting stuff are the PCIe FPGA boards like https://www.xilinx.com/products/boards-and-kits/alveo/u200.h...
One particularly interesting research project was using the FPGA fabric to remap addresses, allowing database tables to be "virtually" rearranged (eg. making a row-major data source into a column-major source for easier searching). https://disc.bu.edu/papers/edbt23-relational-memory https://arxiv.org/pdf/2109.14349.pdf
More optimistically something along the lines of hands-off offloading, where if the scheduler sees enough of the same type of calculation (e.g sparse matrix multiplication) it can reconfigure the fpga and offload.
For instance, SHA-3 is quite slow on x86 CPUs, which, unlike the recent ARM CPUs, do not have any hardware instructions to accelerate it.
If there had been an included FPGA, it would have been easy to implement a very fast SHA-3, as its hardware cost is very low.
FPGAs excel in parallelizing very simple operations. One example where FPGAs are good is where you might want to look for a specific string in a network packet. Because FPGAs are electronic circuits you can replicate the "match a byte" logic thousands of times and have those comparisons all run in parallel and the results combined with AND & OR gates into a yes/no decision. (I think the HFT crowd do this sort of thing to preclassify network packets before forwarding candidate packets up to software layers to do the full decision making).
I don't think that's a good characterization. FPGA are good at what they end up being programmed for. And in the final analysis, everything a chip does is broken down to simple operation.
The FPGA selling point has always been around perf/W and efficiency with regard to a "set of task". An ASIC will always be faster for a specific task, and CPU will always be faster on average on everything. However, when considering say "compression" or "check-sum" as general class of algorithm, and FPGA with a set of predefined configuration could be better and cheaper.
FPGAs being selected for performance per watt was only a fairly recent phenomenon, when they were deployed on semi large scale as password crackers/cryptocurrency miners.
Their real strength is ultimately real time processing (DSP or networking), with reconfigurability often quite valuable for networking applications. For DSP applications it's usually because a MOQ of custom silicon can't be justified.
The moment you are memory heavy the FPGA loses it's edge because CPUs and GPUs are designed around hiding memory latency really well and they overall have higher memory bandwidths.
Acceleration of new hashing and cryptographic algorithms, new compression algorithms, new codecs, and many other similar things, without adding special-purpose instructions for them.
Implementation of fast virtual peripherals. Accurate emulation of special-purpose hardware.
Implementation of acceleration modules for well-established software. Imagine if popular libraries or databases or other engines didn't just come with acceleration using SIMD and other instruction sets, they also came with modules loadable on an attached FPGA if you have one.
"Intel should have dominated this space but Xilinix etc got lucky"
That does seem to be how these kind of deals are usually structured. Spinco is 60% owned by the parent or wherever.
Intel needs the cash, so this is understandable.
I’m being pedantic, I know
Also some company seems to be significantly worst than other, MSFT/microsoft/Dell come to mind. My suspicion is those type of acquisition are mainly driven by C/executive level employee as way to hide the real struggle of the company.
Is there a report analyzing bit tech acquisition say for the last 30 years, and they economical impact ? That would be an interesting read.
Maybe it's time for a new form of regulation around acquisitions
Their competitors don't seem to do this sort of thing.
The rumour mill through the distributors suggests they will be available on standard though, as they're targeted at customer with lower cost centres.
To give credit to Pat Gelsinger, his stated goal is to shed non-essential units and refocus on the fundamentals. But I'm not sure how well that's going.
Intel now, without Andy Grove and the founders steering the culture[0] it seems there is no-one who is able to steer this ship past quarterly results anymore.
AMD had suffered decades of mishaps and near bankruptcy to finally find competent leadership in recent years which turned the company.
[0]: After all, one core virtue that Andy Grove had was "only the paranoid survive", which in context, was all about never getting too comfortable with your market position. Granted, they also engaged in many illegal practices as essentially a monopoly in the 90s. I guess they lost his second virtue of Competitive Mindset which as Grove saw it viewed competition as the key driver of innovation and progress. This was out the window by then, though I don't know if he was effectively leading the company by 96. He stepped down in 98, but he had health problems before that - eventually diagnosed with prostate cancer. This is all to say that Intel's culture is a mixed bag of unhealthy paranoia and being used to being #1 in their categories, which is all incoming executives and major shareholders see when they think Intel. There's no-one to faithfully steer the ship back to clear waters it seems
It might be the case that Andy Grove "only the paranoid survive" was really a factor. Or simply the fact that under Andy, the tech sector was really different : X86-64 market was the dominant and growing platform of compute, they pretty much had a strong set of patent around X86-64. The initial military and gov contract give them enough run away money to build a moat around having a fab.
Also let's not forget despite all those advantage, intel was also fine 3 or time, for more than a billion each time for very anti competitive practice. It's unclear to me that intel was never the great company they think they are.
Fast forward to now, X86-64 is no longer growing as fast, having a fab is no longer a moat with TSCM around (might even be seen as a liability), the WinTel monopoly doesn't have strength he had...etc... etc... The game is just much harder now.
For nvidia, they had a triple boom market with first PC gaming, then crypto,the now AI... Hard to distinguish CEO performance from market performance.
Now the market moving to mobile ARM chipsets definitely damaged Intel, though Intel had years to come up with a viable alternative and ways to re-capitalize its fab infrastructure etc. They let their own moat run dry. Theories abound as to why, and since I'm not CEO nor do I have access to Intels internal records, I can only speculate, however from available information I've been able to read, it seems that they became more risk adverse as time has gone on, due in part, because there is immense pressure for any new thing to quickly produce margins similar to that of their main CPU business. Which of course, is a conundrum, as barring any major breakthrough that puts them well and above any alternative, these sorts of things to diversify into strong businesses often take alot of time.
Compare that to Nvidia: the market dynamics is what is allowing them to capitalize on 2 booms that ultimately leveraged GPUs for compute over traditional processors, but the foundation for all this was laid in ~2007 (CUDA first came out then) and Nvidia had to both improve CUDA over time and invested in research around parallelized computing. If were not for their long horizon investment in this, they would not have been primed to take advantage of both crypto and AI.
I'd argue that gaming while lucrative no doubt, is comparatively modest to what they are raking in over crypto and AI, all the while still investing in CPU / GPU integrated chipsets[0] and CPUs tailored to massive data center requirements around AI specifically[1], all of which they wouldn't be able to really take advantage of it not for their long term investment in things like CUDA and parallelized computing. I suspect quite strongly that without the founder at the helm, there would have been little internal cover to continue investing in these things at the rate Nvidia was prior to the boom times.
[0]: https://www.reuters.com/technology/nvidia-make-arm-based-pc-...
Without CUDA and Nvidia silicon, I'm not convinced the AI boom would have happened any time within this decade.
They seem to only have been useful for prototyping and military applications (low volume and infinite budget).
So sure, prototyping and military, but there are other uses as well. But none of them are super high-volume because once you’re selling millions of something you should be designing your own chips.
When complicated realtime signal processing is to be done, FPGAs shine - in particular if there exists no DSP that is competitive for the task.
Unfortunately their CAD tools suck, but that's beside the point.
Too expensive in volume only applies to Xilinx and Altera and with every node shrink, the amount of designs that fit on an FPGA grow while the non recurrent development costs grow for ASICs. Due to this, the maximum volume at which FPGAs remain cost competitive keeps growing with every generation.
Smart NICs make extensive use of reconfiguration because things such as protocols are not set in stone. They can change all the time. It is also possible to build designs that make extensive use of partial reconfiguration.
MPSoCs have been a thing for a long time. If you want to get into those Google Kria 260 and you will be pleasantly surprised.
When I take a look at Effinix FPGAs those are designed specifically for vision applications and massive amounts of I/O. A CPU would struggle with multiple camera streams and consume too much power.
Yeah ok but an 100k LUT FPGA Chip from Effinix costs what? 25€? You're going to need a really high volume to get an overall cost saving. 40000 FPGAs is only a million dollars. The masks of an 22nm ASIC cost 1.5 million dollars without the rest of the development costs.
And finally your last sentence contradicts the first. Next time stick with a story.