> Instead of making ASICS or FPGAs, Intel focused in recent years on making its CPU processors ever more powerful
If only, intel has been abusing their market position and pushing out "upgrades" that barely have a performance improvement over the previous generation.
AI is not going to eat intel's lunch, all those computers still require cpus. AMD on the other hand may eat intel's lunch by releasing powerful multicore processors for half the price all because they don't waste space on the die for things like integrated graphics.
>If only, intel has been abusing their market position and pushing out "upgrades" that barely have a performance improvement over the previous generation.
And why is that because of "abusing their market position", as opposed to plainly and clearly being more difficult to get faster processors at 14 nm and lower resolutions (and with the low power requirements of today)?
Besides, the trend the article points to is a BS fad as I see it (and I've seen 5-6 of those play out in the last 30 years). ASICS and FPGAs wont even come close to bringing in as much cash as general purpose CPUs do for Intel.
The real holdback for FPGA's and indeed any development of software is and how easy is it to translate requirments into reality. That is an area that general CPU's hold over FPGA's, but that will only narrow as it has over time. Just some distance to go. But then Intel recently brought a FPGA company and with that, they can cover any niche area's today as well as any expansion tomorrow in that area.
As for AI, that needs to run upon hardware and whilst Intel is known for hardware they equaly do software and like most large companies. Often easier to let the little chaps do all the work, pick the best before it becomes big and buy them. Which can be cheaper than starting from scratch yourself and reduce the financial risks, whilst still getting in on the returns using smaller companies as research for a much more diverse skills and talents along with increased financial budget combined than any large company could dream of expending. So in some area's it is easier to have in house skills that can keep track of what everybody else is doing in that area and how it is progressing so that you are ready to swoop in and borg that lego block for your collection and reduce your financial risk greatly. More so in area's that are still very much maturing.
>And why is that because of "abusing their market position", as opposed to plainly and clearly being more difficult to get faster processors at 14 nm and lower resolutions (and with the low power requirements of today)?
Okay, let's say that Intel hasn't been under investing in improving the performance of their cpus. They still have 63% margins across their entire product line, they still don't have 8 core cpu options below $1000. They still don't have consumer cpu options that ditch the integrated graphics.
All of which is only possible because they have had a monopoly.
>All of which is only possible because they have had a monopoly.
You got it backwards: them having a monopoly is only possible because it hasn't been possible for anyone else to produce more efficient desktop/laptop CPUs. On the mobile, where there are competent competitors, it's a different story.
It's not as if if they sacrificed their margins they'd magically be able to produce faster processors. Just cheaper ones.
And why would they have "consumer cpu options that ditch the integrated graphics"? What that would accomplish, when there's a ton of cpu being wasted otherwise?
>> You got it backwards: them having a monopoly is only possible because it hasn't been possible for anyone else to produce more efficient desktop/laptop CPUs.
Intel has patents over the x86 instruction set.
>> It's not as if if they sacrificed their margins they'd magically be able to produce faster processors. Just cheaper ones.
They have produced slower processor on purpose. Instead of soldering the chip to the heat spreader, they started using a very inefficient thermal paste. Replacing it leads to 20K better temps. It hurts both the speed and the expected life time of the processor. But what's wrong if the customer needs a new CPU in 3 years ...
>> And why would they have "consumer cpu options that ditch the integrated graphics"? What that would accomplish, when there's a ton of cpu being wasted otherwise?
The graphics units occupy more than 50% of the die on recent Intel mainstream CPUs. Instead of GPU they could put 8 cores in chips like 7700K and save money.
Yes, one can legally build and sell an 8086 clone. But there are tons of patented extensions in newer CPUs, e.g. http://www.patents.com/us-7499962.html .
They put 'crappy' GPU's on die which they continued to upgrade even when the CPU side stagnated. So, while the last upgrade may have been a process issue, the last 4 CPU refreshes have all been classic monopolist behavior.
PS: GPU's might seem like something you want on the CPU, but bandwidth is a killer worse they add heat which directly lowers CPU performance.
Frankly the updates to the GPUs have been very minor, over the past few generations. And for some reason, they don't put anything close to Iris or Iris Pro on the desktop CPUs.
I've been waiting for a few generations now to get an Intel CPU with decent iGPU but it's really been stagnant. I gave up the wait and pre-ordered a Ryzen last week.
I do still think it makes sense for Intel to ship CPUs with those integrated CPUs, although maybe it wouldn't be a bad idea to offer a chip with more cores and no GPU for the enthusiasts.
> I do still think it makes sense for Intel to ship CPUs with those integrated CPUs
Why? iGPU's increase production costs AND reduce both CPU and GPU performance. If Intel wants to produce a separate GPU then more power to them, but that's not what they are doing they are a monopoly extending their dominance to another market at the cost of consumers aka us.
Remember, resistance increases with heat, so even it there where zero manufacturing or performance costs it still increases power consumption.
On motherboard integrated graphics is still a thing. They can still use main memory ect, but now the CPU is smaller increasing yield and lower temperature.
ASIC/FPGA or even GPU's aren't the future of neural computing. The future is analog(orders of magnitude perf/watt and perf/$). That require older fabs, optimized for analog, and having good embedded flash, which TSMC has and Intel mostly hasn't got.
And that same future applies not only for neural computing, but for a field called approximate-computing, i.e. computing where results aren't accurate.Some/many signal and image processing work well with that. I've also seen some research about doing scientific computing on approximate hardware and correcting errors.
What analog function exactly are we talking about here? I'm genuinely curious. There's certainly a fading art for lots of analog computing techniques, like taking the arctan or log of an input signal. But how is that useful for neural computing?
I worked for a while as the software/FPGA guy on a high-power RF amplifier for MRI machines. I worked with some very talented RF engineers. There were obviously some parts of the chain that were analog (like the preamps and high power amp itself). But lots of it was being replaced with digital (DSP techniques) in the next generation and the RF engineers loved it. Filters no longer needed a 15 minute tuning session by a technician with a network analyzer. A Hilbert transform is far easier and better in digital than analog. Design changes could be experimented with in a day instead of PCB spins that took weeks.
But analog computers do not permit error correction! Crosstalk of 10% does not affect digital data busses, but ruins analog. ECC memory is trivial with digital but not possible even in theory with analog. This (and programmability) is why we moved from super fast analog to slow digital half a century ago.
Given all these constraints I am actually quite excited to hear more of the dissenting view. Can you describe some of the research you mentioned in the last sentence? I have not heard of it and if true it would be at least interesting engineering, even if right now I doubt it would actually work.
A significant portion of any real-world AI system is dedicated to extracting relevant features from raw sensor information. Doing that digitally has proven to have a very high cost rated in both watts and operating temperature.
Most machine learning algorithms are by design resilient to random noise. They can even be learned to be resilient to systematic noise (due to say.. variable hardware performance in mass production).
In those cases; a low power analog device with lower reliability guarantees is 'ok'.
I suspect the "decision engines" (the computation enforcing complex logic based on sensor readings) will probably continue to be on CPU's for some time.
Any feature extractor worth using will be robust against some amount of Gaussian noise. The neural nets will probably never even see the bad readings.
It would surprise me if even unfiltered analog noise in the sensors significantly changed recall. I suppose I don't know off-hand know the magnitude of noise in typical analog systems.
I'm sure AC will be a major revolution in computing in general, though ANN-based systems may come there earlier.
Also, there's a huge application domain in image preprocessing: the pixel value is analog and noisy anyway, so building an analog primary visual cortex with a bank of common filters (line segments, stable features etc.) makes a lot of sense.
ADD: I'm also thinking about accelerating RANSAC-type algorithms, where you perform lots of preliminary evaluations to filter out the outliers. I bet one can cheaply implement an "100-argmins of 100 x 100 fixed10 distance matrix" computation in a handful of cycles with analog circuits. And high-dimensional RANSAC-type algorithms are what needed for robust and precise 3D visual navigation.
> AI is not going to eat intel's lunch, all those computers still require cpus.
But they might not need that powerful CPU anymore, if computational heavy task happens elsewhere anyway. So as time goes, CPU will become less critical, meaning less money for Intel
Let's see. As these early ML-heavy prototypes will start becoming industrial and safety-critical products, the emphasis will be on explainability and graceful degradation. This inevitably leads to a hybrid system with poorly vectorizable parts (expert systems, ad-hoc solutions), which will have to be tightly integrated. Also, full of various auditing and monitoring systems, which are "oligo-threaded" by nature.
Therefore you'll need a capable several-core CPU anyway, and unless the memory is shared (which means you have a SoC, where there's no clear leader AFAIU), moving data in and out becomes a major bottleneck.
I think for most systems where explainability and audit trail are important it will be easier (I mean cheaper) to build on a powerful CPU from the start, and when the system is successful (i.e. approved by real customers), then the highly-parallel parts may be factored out to special hardware. Using them from the start is very risky.
These thoughts are from experience building an all-weather computer vision system for traffic control applications (99.5+% vehicle classification, with errors requiring human intervention and explanation), and being rather familiar with a company which makes specialized ID-recognizing mobile apps.
>AMD on the other hand may eat intel's lunch by releasing powerful multicore processors for half the price all because they don't waste space on the die for things like integrated graphics.
If that's all it is isn't that going to be pretty much a simple knob for Intel to turn to respond? They already have Xeon which doesn't use the space on graphics.
The main reason for pushing iGPU on desktop so hard was so the same chips could be pretty much used on mobile I guess?
Intel would still loose even without amd. Yesterday people that were ordering 100 Intel chips will be ordering maybe just 10% of these as the major processing s has been passed on to a different processor. They could even go for cheaper intel CPUs as the major processing power requirement would be met by gpus.
AI/ML/etc. may be part of it. But the other factor is that you can't just wait 18 months any longer for a new generation of x86 to be a lot [EDIT] faster. That was the big problem with specialized architectures historically. The volume architecture would catch up soon enough without you having to rewrite software to optimize for some different design of processor.
That's no longer the case so specialized designs for the compute-hungry workload du jour (which happens to be AI at the moment) are starting to look a lot more attractive.
It is more a case most cpu intensive demands being addressed by dedicated chips, as we always have had. Many area's of information technology move from general cpu's towards dedicated silicon. Even CPU's adapt and add instructions and with that small area's of silicon space for some dedicated demands (think MMX, AVX, AES,...).
This is no change at all in what we already have. It is when we finally dedicate all tasks down to dedicated silicon that the glue of a CPU processing wise will diminish.
But then CPU's of today are constantly adapting and I'd say the C in CPU is better defined as Centralised rather than Central.
For me, I'm looking forward to a AI grammar and contextual spelling checker that will make all grama nazi's obsolete.
So my perspective upon this for intel is that I foresee no trouble for Intel, who already adapt to change and are not to be dismissed any time soon, just yet.
>For me, I'm looking forward to a AI grammar and contextual spelling checker that will make all grama nazi's obsolete.
I'm not. I feel like human language is meant to be fluid and to evolve as we do. This would potentially lead to more of a global monoculture than we are already starting to get.
That is a very fair observation and one in which I had not considered. I was somewhat biased in thinking about the aspect that it would make universal translation closer to becoming a reality. So it is somewhat a chaos aspect in languages I suppose that does allow evolution of the language.
With that the whole aspect could voice translation add's another aspect to this and perhaps tackling keyboard inputs upon a keyboard layout that is designed to be the worst possible ever layout is an area due to die off sooner than we think. Though the prospect of forcing generations to endure, even if they do not know about typewriters does bemuse me.
In the Classical era, grammar study was focused on a few idealized languages (Latin, Greek, Sanskrit); when people in the Early Modern decided to apply these rules to vernaculars, they ran into the problem that many modern languages don't follow such clean rules. The most common response was to try to insist that things that didn't look Latin shouldn't be considered "grammatical." Only quite recently did people begin working out how to describe the grammar of languages like English.
One consequence is that almost everything you're taught about English in school is completely and totally wrong. There's the completely bogus prescriptivisms that have no grounding (e.g., thou shall not end a sentence with a preposition). But even basic things like "what are the parts of speech" are pretty much wrong, being derived primarily from "this was what this Latin scholar said 2 millennia ago about a language that has a distant linguistic relationship to English."
It's also worth pointing out that the trend to conserve spelling and the written form probably obscures the underlying grammar as it changes. The French clitic pronouns seem rather more like recently-introduced inflections to the verbs rather than clitics, and the 's of English acts rather more like a particle than a genitive case marker.
The "3,854 cores" versus "28 cores" is dubious as always. 3,854 I think counts just the individual fp32 ALUs; a true similar comparison would be number of warp schedulers or maximum number of warps resident at once, or even just SM count (which share a cache).
Apples to oranges (a super-hyperthreaded 1024/2048-bit wide vector machine with minimal cache to a minimally hyperthreaded 128/256-bit wide vector machine with lots of cache).
Apples to oranges? It's more like how a hundred thousand squirrels can't write a novel no matter how long they're given but one person can given a few months.
Not all compute devices are equivalent and "core" vs. "core" is a totally absurd comparison.
> Crazy idea: buy the rights to sell AMD's Vega GPUs fabbed out of Intel
1) I don't think Intel has a lot of spare fab capacity. Certainly not on the nodes AMD is looking to produce Vega on.
2) Intel only just announced a deal to start manufacturing ARM chips on their fabs. [0]
Honestly I can't believe it took Intel so long to wake up and realize that their x86 business is okay, but if they want to survive long term they have to accept that they need another business segment to bring in money after x86 stops being as relevant as it is today.
Just look at TSMC [1] if you want an example of why Intel is foolish to think they can keep being top dog with only x86. TSMC was nobody in the 90's, and now their market cap is within ~10% of Intel's [2] (TSMC @ 160B versus Intel @ 175B).
TSMC doesn't even design their own chips. I'm not saying building semiconductors is easy, or that TSMC has no R&D costs, but you're talking about a company which specializes only in manufacturing some of the most advanced chips on the planet, and doing it at volumes I doubt Intel can match. I predict unless Intel does something major in the near future (<24 months), TSMC will surpass Intel's market cap.
The former CEO of Intel Paul Otellini captured it best himself:
"It wasn't one of these things you can make up on volume. And in hindsight, the forecasted cost was wrong and the volume was 100x what anyone thought." [3]
Intel still thinks they can kill it by selling expensive CPUs. TSMC is proving that thinking is outdated. You don't have to have a 60%+ margin on your chips, you just have to make it up in volume.
Where do you think the next billion chips are going to be sold? It's not going to be $500 x86 CPUs. It's going to be <$5 ARM chips in embedded devices, and that's exactly the market segment TSMC is appealing to.
The strategy that nervana is taking is to reduce precision to 16 bit fixed point and then accumulate in 48 bits (which appears to be unnecessary and 24 bits should be sufficient).
For deep learning or chips for deep learning?
With an EE/Chem/Math background you should be set to go :)
For deep learning, I highly recommend the cs231n course materials (available for free online) and the Deep Learning Book by Goodfellow et al.
For chip design for deep learning specifically, it's a fairly new field, and a lot of commercial interests, hence why a lot of it isn't available in the form of an "Intro to Deep Learning Chip Design" course. But, the basic point is that deep learning can both train and perform inference in astonishingly low precision.
Other than that, stripping out the cache hierarchy, HBM memory, etc. are the obvious steps to take in targeting deep learning. Note that it's my opinion that deep learning chip startups (such as ours) need another "secret sauce" beyond simply lowering precision, we're not like Intel which can just say "okay, here's $100mil, make me an 8-bit GPU, Go."
I had the same thing happen, but it stayed open fine once I tried the link in incognito mode. Probably just a mis-coded "you've reached your article limit for the month" deal.
Do most of these application (machine learning, vision, etc) rely on OpenCL? It seems to me that GPUs are better suited to OpenCL than a regular CPU, but if that's what all the excitement is about I suggest reading up on some of the work on adding vector extensions to RISC-V and the corresponding flops/watt they're may achieve. They are basing some of the work on results from here: http://hwacha.org although they make it clear that hwacha will not be the standard vector instruction set.
> But the GPUs also have new destinations: notably data centres where artificial-intelligence (AI) programmes gobble up the vast quantities of computing power that they generate.
Should not it be "programs"? Or is "programmes" used in some dialect of English I am not aware of?
If only, intel has been abusing their market position and pushing out "upgrades" that barely have a performance improvement over the previous generation.
AI is not going to eat intel's lunch, all those computers still require cpus. AMD on the other hand may eat intel's lunch by releasing powerful multicore processors for half the price all because they don't waste space on the die for things like integrated graphics.