None of the other items Intel has let go of are an existential must have. GPUs are a must have in any computer big or small. Raw number crunching is core to the very profitable HPC market. Players like Nvidia have fully custom interlinks that threaten to dry up even the CPU sales Intel still gets in this market. GPU is a focus on mobile where expectations have been much raised.
This group is supposedly 6 years old. But reciprocally, the other players have been pouring money in over decades. GPUs are fantastically complicated systems. Getting started here is enormously challenging, with vast demands. Just shipping is a huge accomplishment. Time to grow into it & adjust is necessary. It's such a huge challenge, and I really hope AXG is given the time, resources, iterations, & access to fabs it'll take to get up to speed.
Intel should double down on its core business which is selling x86 chips for Windows PCs. If it loses to AMD+TSMC here, there's nowhere left to hide. IFS is very long term and a new business opportunity in a market where Intel is not the leader. x86 for Windows is their bread and butter. They need to throw everything else out and downsize to just excelling at what they've always done. My guess is AXG is a separate division than the division that makes the graphics for their core microprocessors.
> Intel should double down on its core business which is selling x86 chips for Windows PCs. If it loses to AMD+TSMC here, there's nowhere left to hide. IFS is very long term and a new business opportunity in a market where Intel is not the leader.
The very second someone else but Apple brings a competitive ARM desktop CPU to the market, it's game over for Intel. x86_64 literally cannot compete with modern ARM designs simply because of how much utter garbage from about thirty years worth of history it has accumulated and absolutely needs to support in the future because even the boot process still requires all that crap, whereas ARM was never shy about cutting out stuff and breaking backwards compatibility to stay performant.
The only luck that Intel has at the moment is that Samsung and Qualcomm are dumpster fires - Samsung has enough problems getting a phone to run at a halfway decent performance with their Exynos line and Qualcomm managed to completely botch their exclusive deal with Microsoft [1] (hardly surprising to anyone who has ever had the misfortune to have to work with their crap). A small startup that is not bound by ages of legacy and corporate red tape should be able to complete such a project - Annapurna Labs have proven it's possible to break into the server ARM CPU market well enough to get acquired by Amazon.
Apple will always be niche because they are too expensive.
ARM based PC client processors accounted for about 9% of the total market.
But I don't expect that number will grow a lot higher due to cost of Macs.
Also, you can't bring out a desktop PC based on Linux because that a very slowly growing market segment and MacOS is proprietary. Don't think Apple will start another PowerComputing type scenario where they open their platform to others.
It's not a matter of performance for this reason (price).
So, on Windows, assuming Macs don't take away any more market share from Windows,
they need to beat AMD+TSMC.
> So, on Windows, assuming Macs don't take away any more market share from Windows, they need to beat AMD+TSMC.
No. All it needs is
- Microsoft and Qualcomm breaking their unholy and IMHO questionably legal alliance
- an ARM CPU vendor willing to do the same as Apple did and add support for accelerating translation of x86 code (IIRC, memory access models/barriers are done differently between x86 and ARM, and Apple simply extended their cores to be able to use the same memory access/barrier model as x86 on translated-x86 threads)
- an ARM CPU vendor willing to implement basic functionality like PCIe actually according to spec - even the Raspberry Pi which is the closest you can get to a mass market general-purpose ARM computer has that broken [1]
- someone (tm) willing to define a common standard of bootup sequence/standard feature set. Might be possible that UEFI fills the role; the current ARM bootloaders are a hot mess compared to the old and tried BIOS/boot sector x86 approach, and most (!) ARM CPUs/BSPs aren't exactly built with "the hardware attached to the chips may change at will" in mind.
Rosetta isn't patented to my knowledge, absolutely nothing is stopping Microsoft from doing the same as part of Windows.
If you translate, you do face performance issues.
Apple or some other vendor cannot make an ARM chip so fast at a competitive cost
that beats an x86 chip in emulation mode.
The underlying acceleration techniques for both ARM and x86 are the same.
> Apple or some other vendor cannot make an ARM chip so fast at a competitive cost that beats an x86 chip in emulation mode.
This reminds me of Iron Man 1... "Tony Stark was able to build this in a cave! With a box of scraps! - Well, I'm sorry. I'm not Tony Stark."
Apple has managed to pull it off so well that the M1 blasted an i9 to pieces [1]. The M1 is just so damn well more performant than an Intel i9 that the 20% performance loss compared to native code didn't matter.
And for 99.9999% of users the M1 performance benchmarks vs. i9 _don't matter one bit._
The use case for the vast majority of laptops include I/O- and memory-bound applications. Very few CPU-bound applications are run on consumer laptops, or even corporate laptops, for the most part. CPU-bound applications should be getting run on ARM or GPU clusters in the cloud.
The use case for an M1 in laptops is the power benchmarks vs. an i9.
> The use case for the vast majority of laptops include I/O- and memory-bound applications.
Where the M1 just blows anything desktop-Intel out of the water, partially because they integrate a lot of stuff directly on the SoC, partially because they place stuff like RAM or persistent storage extremely close to the SoC whereas on desktop-Intel RAM, storage and peripheral controllers are all dedicated chips.
The downside is obviously that you can't get more than 16GB RAM with an M1 and 24GB RAM with the new M2's and you cannot upgrade either memory at all without a high-risk soldering job [1]... but given that Apple has the persistent storage so closely attached to the SoC to swap around, it doesn't matter all that much.
No.
The architecture is the biggest factor. Damn, even Jim Keller talks about how most programs use a very small subset of instructions.
It isn't like RISC makes miracles, but sure helps them when your power budget is small.
The "x86 translation support" bits are part of ARM64 ISA, they just weren't standardised in time - apple effectively implemented a WIP version of it AFAIK, though main changes are in interface for OS.
The fact you can run old software on any PC still is a pretty tremendous feature. Apple kicked of 32 bit apps from Mac and iPhones, its user accept that, business PC users might not. I have a number of IOS games I enjoyed that are no longer usable (yeah I could have stopped upgrading my i device, but the security problems make that not a viable option)
Conventional wisdom in the mid 90s was powerpc (RISC) would eventually be better the x86, but it never happened. They worked around the issues. And Microsoft eventually made an OS that wasn't a crash fest (I'm looking at you windows ME)
Also no one can afford to be on TSMCs best node when Apple buys all the production. Apple's been exclusive on the best node for at least a couple years now. Even AMD isn't using TSMCs best node yet.
> Apple kicked of 32 bit apps from Mac and iPhones, its user accept that, business PC users might not.
Yeah, but that was (at least on the Mac) not a technical requirement, they just didn't want to carry around the kernel-side support any more. IIRC it didn't take long until WINE/Crossover figured out a workaround to run old 32-bit Windows apps on modern Macs.
> Also no one can afford to be on TSMCs best node when Apple buys all the production. Apple's been exclusive on the best node for at least a couple years now. Even AMD isn't using TSMCs best node yet.
Samsung has their own competitive fab process, but they still have yield issues [1]. It's not like TSMC has a monopoly by default.
There is just one serious problem with the x86 architecture and that is the difficulty of decoding instructions which could be anywhere between 1 and 14 bytes. It's not hard to design a quick instruction decoder but hard to design a quick instruction decoder that is power efficient.
... Oh yeah, and there are the junkware features like SGX and TSX and also the long legacy of pernicious segmentation that means Intel is always playing with a hand tied behind its back, for instance, the new laptop chips that should support AVX512 but don't because they just had to add additional low performance cores.
I may be showing ignorance of some important use case here, but I'd put AVX512 in the "junkware feature" category. It can run AI/HPC code faster than the main processor, but still so much slower than a GPU as to be pointless; when it does run, it lowers the frequency and grinds the rest of the processor to a halt until it eventually overheats anyway. Maybe, maybe it has a place in their big desktop cores, but why would anyone want it in a laptop? You can't even use it in a modern ultrabook-style casing without a thermal shutdown and probably first-degree burns, without even mentioning the battery drain.
If you have all the cores running hard the power consumption goes up a lot but if it is just one core it won't go up too much. If worse come to worse you can throttle the clock.
In general it's a big problem with SIMD instructions that they aren't compatible across generations of microprocessors. It's not a problem for a company like Facebook that buys 20,000 of the same server but normal firms avoid using SIMD entirely or they use SIMD that is many years out of date. You see strange things like Safari not supporting WebP images on a 2013 Mac while Firefox supports them just fine because Apple wants to use SIMD acceleration and they'd be happier if you replaced you 2013 Mac with a new one.
I worked on a semantic search engine that used an autoencoder neural network that was made just before GPU neural networks hit it big and we wrote the core of our implementation in assembly language using one particular version of AVX. We had to do all the derivatives by hand and code them up in assembly language.
By the time the product shipped we bought new servers that supported a new version of AVX that might have run twice as fast but we had no intention of rewriting that code and testing it.
Most organizations don't want to go through the hassle of keeping up with the latest SIMD flavor of the month so a lot of performance is just left on the table. Intel is happy because their marketing materials can tell you how awesome the processor is but people in real life don't experience that performance.
Slowdown/overheating are not inherit to AVX512, they are specific to older implementations. And no the GPU is not a replacement, the latency difference between GPU and AVX512 is off the charts.
Okay, there's a use case I didn't know about. Do you know any SIMD projects that require the lower latency?
I'm glad to hear some newer AVX512 implementations are better. I haven't used an Intel one for years, but I have used the current-gen Ryzen version. Unfortunately it's still suffering the same problems-last time I stress-tested it my CPU hit 100C and some absurd power number before I was too scared to continue. Granted, that was with Prime95, which I believe is close to peak utilization.
No one in their right mind will touch anything like Loongson or whatever else the Chinese government puts out. Maybe the Chinese military, but that's it.
>None of the other items Intel has let go of are an existential must have.
Except mobile ARM chips and mobile LTE modems, both of which Intel sold off, and those are some of the most desirable things to make right now. Just ask Quallcomm.
Yep. Competent mobile ARM and LTE modem cores would be very nice to have in their tech stack right about now. I think a credible (if not fully competent), GPU stack is pretty essential for strategic relevance going forward. This seems like something they have to make work.
> Since Q1’21 when Intel started reporting on its dGPU group, known as AXG or accelerated graphics, the company has lost a staggering $2.1 billion and has very little to show for it.
You know, except for a competitive desktop GPU. I'm actually impressed that it didn't take much longer and much more money to catch up with AMD and NVIDIA, given that those two were in business when 3dfx was still around.
I don't know why your painting Intel as some new entry GPU beginner here, but FYI, Intel had been making and selling integrated GPUs for over 20 years now, to the point they pretty much dominate the GPU market share in sheer numbers (most desktops and laptops, even older Macs, sold in the last 10-15 years are most likely gonna have an Intel chip with integrated graphics, regardless if it's used or not in favor of a discrete GPU).
Sure, their iGPU offerings were never competitive for gaming or complex tasks, but given it came for "free" with your CPU, it was good enough for most businesses and consumers and was a also a major boon during the GPU shortage where gamers who build systems with AMD chips were left unable to use their PC while those who went Intel could at least use their PC for some productivity and entertainment until they could buy a dGPU.
So it's not like they had to start absolutely from scratch here. In fact, Intel's latest integrated GPU architecture, Xe, was so good, it was beating the integrated Vega graphics AMD was shipping till the 6xxx series with RDNA2 in 2022, while also killing the demand for Nvidia's low end dGPUs for desktops (GT 1050) and mobile (MX350). Xe was also the first GPU on the market with AV1 decode support.
So given this, Intel is definitely not a failure in the GPU space, they're definitely doing some things right but they just can't box in the ring with "Ali" yet. Anyone thinking they can leapfrog Radeon and Nvidia at their first attempt would be foolish. Intel should take on the losses on the GPU division for a few more years and push through.
I'd have to agree with GP here. It wasn't at all obvious for almost anyone that Intel can get so quickly to the level of performance Arc offers. After decades of bottom of the barrel iGPUs in a span of 2-3 years they went to decent iGPUs, and then decent dGPUs (according to those who touched them). That's quite an achievement in a market that all but solidified around the 1.5 incumbents.
> they went from decades of bottom of the barrel iGPUs
Hah, I remember the time I spent a week dodging the software renderer to figure out why my shader wasn't working on an intel iGPU -- turns out that someone at intel decided to implement "float" with an 8 bit float. Not 8 byte, 8 bit.
"Sure, we support floating point, what's the problem?"
> during the GPU shortage where gamers who build systems with AMD chips were left unable to use their PC
This comment baffles me, both AMD and Intel have CPUs with onboard graphics and those without. You even noted the integrated graphics a sentence later.
If anything, this is more evidence that AMD is following the Intel playbook by having that integrated CPU/ GPU architecture plan.
>This comment baffles me, both AMD and Intel have CPUs with onboard graphics and those without. You even noted the integrated graphics a sentence later
Why does it baffle you? AMD has only been selling desktop chips with integrated GPUs only for a few years now (they called them APUs), and their APUs were not that stellar at either the GPU or CPU part due to compromises on both parts.
Most of the successful Ryzen chips AMD was selling for the desktop were exclusively without integrated GPUs, to save die space and cost, which hurt PC builders during the GPU scalpocalipse, while on the other hand, Intel's almost entire CPU product range for desktops had integrated GPUs for over 10 years now, enabling PC builders to at least use their PCs until a dGPU could be available.
Sure, Intel sold some CPUs without iGPUs but those were very few SKUs in comparison. Similarly, but in reverse, AMD also sold some Ryzen CPUs with iGPUs(APUs), but those were very few SKUs as their CPUs were weaker than the non-iGPU SKUs, and their outdated Vega iGPUs were pretty weak even compared to Intel's Xe.
So that's the major difference between Intel and AMD that was a game changer for many: Intel shipped most of its chips with iGPUs for over a decade while AMD did not, meaning you always needed to buy a dGPU, and if you couldn't, like in the past ~2 years, well ... good luck, your new tower PC is now an expensive door stop.
AMD's Zen 2 Apus were not made (widely) avaialble at retail, and Zen+ APUs only went to quad core. If you wanted 6 cores and an integrated GPU, you had to buy Intel or from Aliexpress or wait until August 2021 for Zen3 APUs to launch. The two 4650G chips I got in late 2020 work great, but having to call my card issuer to get them to turn off all fraud production for an hour so I can get an order in wasn't ideal.
> So it's not like they had to start absolutely from scratch here
No, and actually in some respects that's not a good thing either. Their existing iGPU driver was designed with the assumption of GPU and CPU memory being pretty much fungible, and with the CPU being pretty "close" in terms of latency, just across the ringbus. It wasn't even PCIe attached, like how AMD does it, it was directly on the ringbus like another core.
Now you need to take that legacy codebase and refactor it to have a conception of where data lives and how computation needs to proceed in order to feed the GPU with the minimum number of trips across the bus. Is that easier than writing a clean driver from scratch, and pulling in specific bits that you need? ....
One of their recent bugs in raytracing was literally due to one line of code in an allocator that was missing a flag to allocate the space in GPU memory instead of CPU, a one-line change produced a 100x speedup in the raytracing performance.
It is most likely much easier to do what AMD did and go from discrete to integrated than the other way around... again, they don't have a tightly-coupled design like Intel did, their iGPU is literally just a pcie client that happens to be on the same die.
(also, AMD paid the penalty like 15 years ago... there were terascale based APUs, and the GCN driver was developed as a dGPU/iGPU hybrid architecture from day 1, they never had to backport GCN itself from dGPU, that was all done in the terascale days.)
I have not seen any desktop GPU that is competitive with Nvidia or AMD yet. What surprises me is how naive Intel seems to be. It is extremely hard to enter the GPU market right now, so it should expect a lot of years of losses. Hopefully they will continue and fix their drivers and hardware problems so we can have more competition. For Alchemist, if they do not price it extremely aggressively ( like 200$ for a A770 ) they will not get much marketshare, I am afraid.
I have to believe that Intel's GPU division will be fine just because I refuse to believe that any executive over at Intel is short sighted enough to believe that Intel was going to leap frog AMD and Nvidia or even make a profit within the next 3-5 years. It took AMD years to get it right with Ryzen and AMD still hasn't had a "Ryzen moment" with their GPU. The fact that Intel's GPU can even compete with AMD and Nvidia is a feat of engineering magic. Intel should just take the L and make it a loss leader and grind it out for the long haul.
I would say RDNA/RDNA2 has delivered Ryzen-level improvements. It didn't result in Ryzen-level fanfare because Nvidia improves their GPUs every product cycle instead of sitting still like Intel was doing with processors. I don't think Intel has the leadership or, after their Boeingization (getting rid of senior engineers) the engineering ability to compete with Nvidia.
AMD has beat NVIDIA in terms of performance and/or perf/watt more than NVIDIA has beaten AMD. Just because NVIDIA has larger market share does not mean they were/are winning. If you need to go back and look at history, look at history. The majority of the time AMD beat the snot out of NVIDIA. The lead only stopped because of a certain prior CEO doing some really dumb stuff. The GPUs you see today (6900xt/6950xt) are the first of a generation of recovery. They are the 'Ryzen 1xxx' equivalents.
Note I'm not an AMD fanboy, but I've 3D in my lifetime was popularized by 3DFX, so I remember the entire saga.
Before delivering good products that customers buy, they need decent products that customers take seriously. Current ARC graphics cards are, reportedly, not even taken seriously by their manufacturers.
I'm not sure how you can give up the GPU market when Intel aspires to be a performance leader and must have some plan to get back into mobile. They have a GPU with bad drivers right now, but instead of axing the division, it would seem more appropriate to get some folks who know software. The duopoly of NVIDIA and AMD are a lot more vulnerable than other players in other markets Intel could expand into.
Intel somehow has the laptop makers absolutely flogging Intel CPUs, to the extent of actually concealing AMD builds from the online picker unless you specifically identify them by model number.
OEMs/ODMs/System Integrators get awards for shipping intel products, but instead of pure cash like in 2000 its hardware and rebates. Until very recently intel was giving out SSDs for free, probably switched to something else since they dont make those anymore.
Short term is may seem like a smart move to cut AXG but long term if it’s a must have.
All three AMD, nvidia and Apple are unifying general purpose compute with graphics. That is the direction the world is going. Unless intel had a new arch trick up their sleeves they will be soon made redundant
since everyone is on the hybrid computing bandwagon, might as well bring the axg on core and make a xeon with axg on the same package. that's what apple is doing, might as well give that a try.
Yes, the insight being that “graphics” is actually mostly just highly parallel compute. So if you don’t have a graphics solution then you’re unlikely to have a competitive compute solution either. It’s not quite there yet, but I think it will be soon.
And actually not just graphics but the world is moving towards these highly unified solutions in general, the hardware world is going "full stack" and companies need to be able to offer top-to-bottom solutions that solve the whole problem without relying on someone else's silicon.
Same reason NVIDIA wanted to buy ARM. Intel and NVIDIA are in trouble unless they can make that leap.
Long-term it's the same reason AMD bought ATI too, that vision just took a long time to come to fruition. Remember, they "acquired" their way to success as well, RTG wasn't something that AMD indigeneously developed themselves either... just like AMD bought Xilinx and Intel bought Altera.
I really don't understand how Intels drivers for Alchemist are so flawed, they have integrated GPUs that run esports titles at 100+fps.
It seemed to me scaling the existing architecture to have more processing and bandwidth should have yielded a very competitive GPU while reusing their existing talent pool.
Instead we end up with a very buggy, very specialized (needs resizable bar?, DX12 works well but not other runtimes?) that aren't required today for Intel graphics.
One pattern that has emerged is many of the performance problems come from VRAM no longer being carved out of main system RAM. The drivers now have to be careful about which pool of memory they allocate from, and for CPU access to VRAM without ReBAR the access patterns matter a lot more than they used to.
DX12 works well cause it was designed to make driver development easier. Last gen APIs required specific optimizations in drivers for every single game in order for it to run well. And as far as I heard the problem isn't in drivers, problem is in hardware scheduler that have some kind of limitations that become more apparent with scale. It might be impossible to fix this without retape.
Hope it does not end up like Larrabee - really don't get why Intel is competing in this market segment - Integrated graphics yes but surely there is bigger and more important fish to fry (Amd Zen) elsewhere ???.
They don't want to depend on third party companies to complete their stack offering on the datacenter. It's the same way as why Nvidia wanted ARM, but on the other direction (CPU/GPU)
>Started in 2016 the dGPU group snatched showboater Raja Koduri away from AMD
We are in 2022 I still dont understand why Raja is popular. One of the reason why I have been extremely sceptical of Intel's GPU since the very beginning. To the point I got a lot of bashing on Anandtech and HN.
And I have been mentioning drivers as the major concern since 2016. Citing PowerVR Kyro on Desktop as an example. Even pointing that out on Twitter. With one of the Intel Engineers on the GPU team replied something as "GPU Drivers is a solved problem".
I do love to be wrong. But it is increasingly looking like another item to be added to my book of prediction that came true.
I'm with you here. At AMD, Koduri released GPU after GPU that was slow, terribly hot and inefficient, plagued by driver problems, and overpriced. He left on terrible terms for a number of reasons, but one of them was his complaint that Radeon was underfunded and deprioritized, so he couldn't compete with Nvidia. Here we are, six years of practically unlimited funding and support from a company with a greater R&D budget than AMD's entire 2016 gross...
That's how failing upwards works in this racket. As long as you have a great looking resume at a few giants in the industry with some fancy titles to boot, you're set for life regardless of how incompetent you are.
I was surprised he got promotion and even more resources under Pat.
I dont think Intel should axe their GPU, but considering Pat Gelsinger has zero BS tolerance, I would not be suspired if someone replace Raja if he doesn't deliver within the next 12 months or so.
To be fair, in the consumer space, he seemed to be holding them back. In the data centre acceleration space, they're still essentially using a descendant of the Vega uArch.
Honestly, I'm in the same boat as you, until I realized I don't know anything he's done before he became head of Radeon Group. According to his Wikipedia entry:
> He became the director of advanced technology development at ATI Technologies in 2001.[3] Following Advanced Micro Devices's 2006 acquisition of ATI, he served as chief technology officer for graphics at AMD until 2009. At S3 and ATI he made key contributions to several generations of GPU architectures that evolved from DirectX Ver 3 till Ver 11.[4] He then went to Apple Inc., where he worked with graphics hardware, which allowed Apple to transition to high-resolution Retina displays for its Mac computers.[5]
So he has a history of launching some really good products but I'd say it's been at least a decade since he's been involved in anything industry leading.
He has experience, I dont deny that. But nothing about S3, ATI or AMD in his era was industry leading or innovative. And considering his fall out with Lisa Su, I am firmly on Lisa Su 's side.
To be successful as a new entrant against powerful incumbent players you need to have some sort of competitive angle / advantage. Intel used to have a fabrication lead but chose not to make dGPUs whilst they had that and allowed Nvidia and AMD to own the market.
Really struggling to see this leading to a successful outcome - even if they produce a reasonable product it’s likely to be third placed which is not a comfortable place to be.
I'm still convinced by Intel's and Pat's XPU strategy. Having a well rounded portfolio of CPUs, GPUs, SoCs, and FPGAs that all work flawlessly, alongside a comprehensive software environment to make cross-platform development easier (their OneAPI stack) is a dream for any OEM/ODM; provided it works.
Their issue at the moment isn't strategy. It's execution. Axing their GPU division would only hurt their current plan, and do nothing to fix the the systematic problem that they're missing deadline's and shipping incomplete products. From the outside looking in, it seems like there's some fat that needs trimming and people aren't pulling their weight. If they can scale back to efficient team and org sizes, cut the side projects, and focus on excellent software and hardware validation, I can see them pulling this off and Pat being lauded as a hero.
People I know who develop high performance systems are so indifferent to OpenCL it's hard for me to picture what could make people develop for OpenCL.
There is such a long term history of failure here that somebody has to make a very strong case that the next time is going to be different and I've never seen anyone at Intel try that or even recognize that history of failure.
This depends on the degree to which the underlying hardware remains compatible across generations. nVidia can sell monster GPU systems today to run 5-year-old CUDA code faster without any recompilation or other mandatory engineering work. I don’t know whether this is possible with direct hardware access.
Yeah it is in practice not really true. Maybe after hacking the code you can watch it compile for _days_ and then perform probably worse than a GPU without extreme tuning, which removes the point anyway.
Depends... sometimes its built specifically to kernels.
One common instance is, blender's Cycles renderer. Every time there is a new NVIDIA GPU. It needs to be recompiled to support it. Sometimes that also requires a new CUDA version to be able to do it. CUDA versions over the years have deprecated different operations and what not.
That's a "did you just tell me to go fuck myself?" solution. It didn't work for AMD and it won't work for anyone else. Developing a CUDA-to-GPU compiler and runtime stack is an immense amount of work and customers or "the community" can't afford it.
One of the big attractions of CUDA is not just the core (programming, API, compiler, development tool) but also the libraries (cuDNN, FFT, cuSignal, … It’s a very long list). These libraries are usually not open source, and can thus not be recompiled even if you have core CUDA support.
Intel doesn't implement CUDA, but they're working on a CUDA to DPC++ (their parallel programming language) migration tool to make the jump less painful. I've tested some of the NVIDIA samples with it and it seems to do the vast majority of the grunt work, but there are still some sections where refactoring is required. You can find it in the OneAPI base toolkit.
No way 30-50% is enough. You need 2x at least, if not more. There is a deep rooted ecosystem of CUDA apps. You would need to have at least 2-3 generations of 2x lead to get people to switch.
With all accelerators, GPUs, FPGAs etc., the solution to me seems pretty simple. Give it away. Like literally give them away like usb sticks in a parking lot. All these companies and startups work on raising massive rounds and focus on enterprise sales and deals with national labs etc., while the average joe has nothing to tinker with. Everybody who mattered had a cheap GPU, and a cheap x86 system before that. That's what drives adoption.
Back in the early 2000s you just had to put your email into a developers' sign-up list and tiny startups nobody had ever heard of like NVIDIA would ship you a GPU nearly a full year before it was available to mainstream customers. Fun times...
There are FPGA dev kits that are fairly affordable. It's probably still cheaper to get a breadboard and a bag of 54xx chips if you want to learn digital logic but I'd say the problem for hobby FPGAs is that the developer tools are poor and the FPGA vendors keep the bitstream format secret so there is no competition to produce better developer's tools.
I was also going to mention some dev boards NVIDIA had come out in the last few years that were affordable (I think) but they are all out of stock now and the ones you can get are $2000+ now.
Intel tried that in its own way. Galileo/Edison Quark based products, lasted 2 years and was pure shit. They entered Arduino hobby embedded world with microcontroller like x86 without dedicated fast IO (it could do something like 1KHz max) and no support/documentation. Intel version of giving it away was pricing slightly below? competition with similar spec (400MHz) while targeting 8bit 16MHz market :-)
Nowadays only Chinese seem to be able to give it away to grow market share. $1 Espressif is a good example, Govin $7 GW1N-1 fpga devboards. Raspberry <$1 RP2040 is one western exception I could come up with.
For me it's like watching a slowly unfolding car crash. Opencl for all it's flaws is the only compute API that is open, yeah oke Vulkan compute exist as well. All of them using proprietary CUDA has already bitten them in the ass many times. Yet there is nothing beeing done to remove CUDA from the stack at all. Quite the opposite in fact.
Execution is about engineers.
Intel can't retain their engineers.
Without good engineers you can't compete.
So you downsize and try compete only on one thing -- better x86.
Pat Gs return as CEO does help with engineer retention though.
For the last 20 years there have always been stories in the press about how Intel is laying off 10,000 engineers. It never struck me as a very psychologically safe place to work.
Meh, not really. Google has some of the best engineers in the world yet they fail at nearly everything that isn't related to search, android, ads and e-mail.
Even Jensen Huang said that it's more about vision and not all about execution. He said in the early days of Nvidia all of their competitors at the time had engineers good enough to execute, yet only they had the winning vision that enabled them to make and sell exactly what the market wanted, nothing more, nothing less.
Yeah they do, Intel has the FPGA division they bought from Altera. That's a huge business on its own.
And my point still stands: having great engineers is not enough for great execution. You need great leadership with a vision a-la Steve Jobs or Jensen Huang.
Oh yeah I'm waiting for their 200G NICs while NVIDIA is ramping up connectx 7 with 400G on PCIe5...
The low end AI chips are a mess. Myriad-X can be used only through openvino, and they've been closing details about the internals... Keembay is... Years late?
Is there anything coming out of Intel these days?
And what happened to nervana systems, became plaidml then disappeared after bought by Intel? Maxas was really great and now, crickets.
Even profiling tools, which they used to be on top of, don't seem to work well on Linux w/ the TigerLake gpu. It's so painful to debug and program, they might as well have not put it in...
Google didn't have Android because they had a good engineering team who created it though. They have Android because they had executives with vision who bought Android Inc., and because they had the subsequent humility and vision to override what the engineers had initially developed resteered the project in the wake of the iPhone.
I've noticed it's very hard for a lot of technical people to give credit to executives, marketing, finance, or anyone else when it comes to the success of a technical product, but very easy for them to assign blame to those groups. Whereas those latter groups often seem to be the first to credit engineers with successful products, and willing to share blame for failures.
Maybe it's because I'm in the technical group so I hear insider gossip and rantings, and the other groups project an overly generous facade, but unfortunately sometimes it doesn't feel like it. Ask a sample of nerds about failed projects near to their hearts. DEC, Sun, Boeing, whatever Google chat app they liked, the failed company or project they worked on, etc. 9 times out of 10 you'll hear rantings about greedy management, incompetent finance, the idiots in marketing, etc. who wantonly desecrated the Mona Lisas and Sistine Chapels of Engineering.
Many engineers are very good at taking a problem given to them by their company or one they've dreamed up and optimizing the hell out of it. Very few are capable of identifying problems that will be valued by others, and coming up with approaches to solving them in ways that are cost competitive and marketable. Very few people at all can do that well because it requires a measure of creativity and multi disciplinary vision, but at least executives are supposed to be thinking about these things so they have a chance.
If you look at the details, I would lay the blame on Ghemawat and Jeff Dean and their insistence on using C++ for everything. Nobody could argue against them, even VPs.
When YouTube joined Google, their engineers were shocked to find that Google engineers were writing 1000s of lines of server side C++ code, that could be accomplished in 100s of lines of python code. Which explained why YouTube was able to move a lot faster and pump out features, while Google video was struggling to keep up.
It was not the leetcode engineers, but Dean and Ghemawat insisting on C++ only on the server. Its still true today. Go might be making inroads now.
My impression is that most of Google’s failures are strategic. Stadia, for example, has been executed very well in the technical sense. It just doesn’t make a lot of sense strategically. I feel every failure I can think of fits this mold of great technology solving the wrong problem or held back from solving the right problem.
Google’s pricing model didn’t even make sense. Pay for a subscription and then buy the games that will forever be locked on the Stadia service? Why would anyone want to do that. Microsoft was just rolling out Gamepass, why would you not do that in the cloud?
Google is very engineering led. From my experience working with software developers, and as a software developer myself, they're really bad at the business aspect of things. They're too idealistic and not practical.
That and their culture of "move fast, fail fast" means they pull the rug just as fast as they lay it out.
> Meh, not really. Google has some of the best engineers in the world yet they fail at nearly everything that isn't related to search, android, ads and e-mail.
Seems to me that Google didn't fail to execute in other things they tried. The products worked fine, but they didn't have a market.
That's not failure of execution, it's failure of strategy.
The sweetspot secret sauce is for leadership to superposition the talent's output in the product. I don't know if ever Microsoft's ecosystem did this. While watching Google's acquired Deepmind defeat China's Go-champ I was playing with everything top of the line from Microsoft's ecosystem and the whole was less than the individual parts in quality, design, etc, put together. Apple has at times achieved this while under Steve Jobs. Steve Jobs's vision was decades and decades ahead of the curve if you go through his interviews. You get this sixth sense is in Elon Musk.
It's a bit of everything but yeah, what I've seen is basically that Intel has a combination of absolutely insane talent working on some really hard problems and a lot of mediocre talent that they churn through.
And the reality is that yeah, with their comp for engineers so abysmally bad why would anyone really go there? Especially when it's not like they're getting great WLB.
Part of strategy is knowing what your organization is capable of, and not burning time and money trying to do something you can't execute effectively.
I mean, "make the fastest CPUs and GPUs and charge a premium" would be a great strategy if they could deliver. But in a world where they're struggling to keep up on the CPU side, pouring resources into GPUs could lose them both markets.
I can't remember the last time I sincerely used an integrated GPU, Intel or AMD, since they're too weak for my purposes. I only ever use them as a Lowest Common Denominator failsafe when troubleshooting.
And for use cases (not necessarily mine) where low-end GPUs make sense, it's not like Nvidia and AMD don't sell bottom of the barrel GPUs that you (or OEMs) can buy for a Franklin each or something.
There is nothing stopping them from using discrete GPUs on the lower end. Don't tell me they don't have the physical space for another chip on the mobo either, because they do.
Integrated GPUs with shader support (Intel HD graphics, not GMA) are fine for web browsing and programming work, which is what I spend most of my time on. In laptops with switchable graphics, sticking with integrated graphics can save power consumption. And additionally they're not anywhere near as scalped/overpriced as discrete gaming GPUs, or as horrible value as low-end GPUs (I suspect some of the price comes from the cost of a display adapter and media decoder/encoder). Though integrated graphics are subpar for gaming or compute for sure.
"Of the groups Gelsinger got rid of was Optane (started in 2017, never made a profit), sold McAfee (bought in 2010, never made a profit), and shut down the drone group (started in 2015, never made a profit). Last year, Intel sold off its NAND business to Hynix, giving up its only Chinese fab, and sold off its money-losing sports group; and this year, the company shut down its Russian operations. Since Gelsinger’s return, Intel has dumped six businesses, saving $1.5 billion in operating costs and loses."
None of these groups are remotely as close to Intel's core (pun intended) business, as this one. It is certainly true that this division needs to produce better results, but that's true of the company as well. Chopping this one, unless there is some plan for a replacement strategy in the GPU space, would be a bad sign, as it would suggest a company circling the drain/milking the existing winners, unable to make any new winners.
The drones were great, and expensive. I don't get how they couldn't make a profit. They only had one model too, it took DJI another 4 years before it had an acceptable rival in the industrial class. It was a simple drone too, it couldn't have been that heavy of a burden on their financials, why not just keep it around for a couple years longer, the drone industry is clearly still maturing.
What the heck about GPU development costs $2 billion? That seems obscenely expensive. Anyone who has detailed knowledge of the industry care to weigh in on what they might actually spending that money on?
Intel needs to buy talent, make shit loads of mistakes, try as hard as they can to squeeze as much driver optimization as they can in reduce the driver gap as much as possible.
Intel quit making ARM cpus right before smartphones took off. They sold off their LTE modem division essentially giving Qualcomm a monopoly. Now they might give up on GPUs.
Intel is way too quick to give up. Imagine if they had continued to make ARM CPUs. Apple may have never pursued making their own chips. Intel CPU's + modems could be competing with Qualcomm for every mobile device sold today.
I think it was more hubris that caused them to stop making ARM chips as opposed to "giving up". They were still in the "x86 everywhere" mindset back then. Remember when they were trying to get x86 into phones & tablets?
Intel missed a huge opportunity by killing Larrabee and the Xeon Phi. They aren't going to be able to beat CUDA with the "oneAPI" software layer they are trying to offer: an outgrowth of OpenCL is not going to be popular with people who care about the performance of their devices. The only programming API that Intel has and CUDA can't beat is x86 assembly.
In my opinion, when they wanted to get back into graphics, they should have brought back Xeon Phi, maybe doubled the vector width (and added some special units), and hired some engineers from HFT firms to figure out how to make it pretend to be a fast GPU.
170 comments
[ 3.2 ms ] story [ 245 ms ] threadThis group is supposedly 6 years old. But reciprocally, the other players have been pouring money in over decades. GPUs are fantastically complicated systems. Getting started here is enormously challenging, with vast demands. Just shipping is a huge accomplishment. Time to grow into it & adjust is necessary. It's such a huge challenge, and I really hope AXG is given the time, resources, iterations, & access to fabs it'll take to get up to speed.
The very second someone else but Apple brings a competitive ARM desktop CPU to the market, it's game over for Intel. x86_64 literally cannot compete with modern ARM designs simply because of how much utter garbage from about thirty years worth of history it has accumulated and absolutely needs to support in the future because even the boot process still requires all that crap, whereas ARM was never shy about cutting out stuff and breaking backwards compatibility to stay performant.
The only luck that Intel has at the moment is that Samsung and Qualcomm are dumpster fires - Samsung has enough problems getting a phone to run at a halfway decent performance with their Exynos line and Qualcomm managed to completely botch their exclusive deal with Microsoft [1] (hardly surprising to anyone who has ever had the misfortune to have to work with their crap). A small startup that is not bound by ages of legacy and corporate red tape should be able to complete such a project - Annapurna Labs have proven it's possible to break into the server ARM CPU market well enough to get acquired by Amazon.
[1] https://www.xda-developers.com/qualcomm-exclusivity-deal-mic...
No. All it needs is
- Microsoft and Qualcomm breaking their unholy and IMHO questionably legal alliance
- an ARM CPU vendor willing to do the same as Apple did and add support for accelerating translation of x86 code (IIRC, memory access models/barriers are done differently between x86 and ARM, and Apple simply extended their cores to be able to use the same memory access/barrier model as x86 on translated-x86 threads)
- an ARM CPU vendor willing to implement basic functionality like PCIe actually according to spec - even the Raspberry Pi which is the closest you can get to a mass market general-purpose ARM computer has that broken [1]
- someone (tm) willing to define a common standard of bootup sequence/standard feature set. Might be possible that UEFI fills the role; the current ARM bootloaders are a hot mess compared to the old and tried BIOS/boot sector x86 approach, and most (!) ARM CPUs/BSPs aren't exactly built with "the hardware attached to the chips may change at will" in mind.
Rosetta isn't patented to my knowledge, absolutely nothing is stopping Microsoft from doing the same as part of Windows.
[1] https://www.hackster.io/news/jeff-geerling-shows-off-an-amd-...
This reminds me of Iron Man 1... "Tony Stark was able to build this in a cave! With a box of scraps! - Well, I'm sorry. I'm not Tony Stark."
Apple has managed to pull it off so well that the M1 blasted an i9 to pieces [1]. The M1 is just so damn well more performant than an Intel i9 that the 20% performance loss compared to native code didn't matter.
[1] https://www.macrumors.com/2020/11/15/m1-chip-emulating-x86-b...
The use case for the vast majority of laptops include I/O- and memory-bound applications. Very few CPU-bound applications are run on consumer laptops, or even corporate laptops, for the most part. CPU-bound applications should be getting run on ARM or GPU clusters in the cloud.
The use case for an M1 in laptops is the power benchmarks vs. an i9.
Where the M1 just blows anything desktop-Intel out of the water, partially because they integrate a lot of stuff directly on the SoC, partially because they place stuff like RAM or persistent storage extremely close to the SoC whereas on desktop-Intel RAM, storage and peripheral controllers are all dedicated chips.
The downside is obviously that you can't get more than 16GB RAM with an M1 and 24GB RAM with the new M2's and you cannot upgrade either memory at all without a high-risk soldering job [1]... but given that Apple has the persistent storage so closely attached to the SoC to swap around, it doesn't matter all that much.
[1] https://www.macrumors.com/2021/04/06/m1-mac-ram-and-ssd-upgr...
Wouldn't bet on that one
Conventional wisdom in the mid 90s was powerpc (RISC) would eventually be better the x86, but it never happened. They worked around the issues. And Microsoft eventually made an OS that wasn't a crash fest (I'm looking at you windows ME)
Also no one can afford to be on TSMCs best node when Apple buys all the production. Apple's been exclusive on the best node for at least a couple years now. Even AMD isn't using TSMCs best node yet.
Yeah, but that was (at least on the Mac) not a technical requirement, they just didn't want to carry around the kernel-side support any more. IIRC it didn't take long until WINE/Crossover figured out a workaround to run old 32-bit Windows apps on modern Macs.
> Also no one can afford to be on TSMCs best node when Apple buys all the production. Apple's been exclusive on the best node for at least a couple years now. Even AMD isn't using TSMCs best node yet.
Samsung has their own competitive fab process, but they still have yield issues [1]. It's not like TSMC has a monopoly by default.
[1] https://www.gsmarena.com/samsung_claims_that_yields_from_its...
... Oh yeah, and there are the junkware features like SGX and TSX and also the long legacy of pernicious segmentation that means Intel is always playing with a hand tied behind its back, for instance, the new laptop chips that should support AVX512 but don't because they just had to add additional low performance cores.
If you have all the cores running hard the power consumption goes up a lot but if it is just one core it won't go up too much. If worse come to worse you can throttle the clock.
In general it's a big problem with SIMD instructions that they aren't compatible across generations of microprocessors. It's not a problem for a company like Facebook that buys 20,000 of the same server but normal firms avoid using SIMD entirely or they use SIMD that is many years out of date. You see strange things like Safari not supporting WebP images on a 2013 Mac while Firefox supports them just fine because Apple wants to use SIMD acceleration and they'd be happier if you replaced you 2013 Mac with a new one.
I worked on a semantic search engine that used an autoencoder neural network that was made just before GPU neural networks hit it big and we wrote the core of our implementation in assembly language using one particular version of AVX. We had to do all the derivatives by hand and code them up in assembly language.
By the time the product shipped we bought new servers that supported a new version of AVX that might have run twice as fast but we had no intention of rewriting that code and testing it.
Most organizations don't want to go through the hassle of keeping up with the latest SIMD flavor of the month so a lot of performance is just left on the table. Intel is happy because their marketing materials can tell you how awesome the processor is but people in real life don't experience that performance.
I'm glad to hear some newer AVX512 implementations are better. I haven't used an Intel one for years, but I have used the current-gen Ryzen version. Unfortunately it's still suffering the same problems-last time I stress-tested it my CPU hit 100C and some absurd power number before I was too scared to continue. Granted, that was with Prime95, which I believe is close to peak utilization.
Larrabe was shown at GDCE 2009 as if going to render the competition useless, then faded away.
Their GPU debugger marketing sessions used to be focused on how to optimise games for integrated GPUs.
I really don't get how they keep missing the mark versus AMD and NVidia in both GPU design and developer tooling.
Except mobile ARM chips and mobile LTE modems, both of which Intel sold off, and those are some of the most desirable things to make right now. Just ask Quallcomm.
You know, except for a competitive desktop GPU. I'm actually impressed that it didn't take much longer and much more money to catch up with AMD and NVIDIA, given that those two were in business when 3dfx was still around.
Sure, their iGPU offerings were never competitive for gaming or complex tasks, but given it came for "free" with your CPU, it was good enough for most businesses and consumers and was a also a major boon during the GPU shortage where gamers who build systems with AMD chips were left unable to use their PC while those who went Intel could at least use their PC for some productivity and entertainment until they could buy a dGPU.
So it's not like they had to start absolutely from scratch here. In fact, Intel's latest integrated GPU architecture, Xe, was so good, it was beating the integrated Vega graphics AMD was shipping till the 6xxx series with RDNA2 in 2022, while also killing the demand for Nvidia's low end dGPUs for desktops (GT 1050) and mobile (MX350). Xe was also the first GPU on the market with AV1 decode support.
So given this, Intel is definitely not a failure in the GPU space, they're definitely doing some things right but they just can't box in the ring with "Ali" yet. Anyone thinking they can leapfrog Radeon and Nvidia at their first attempt would be foolish. Intel should take on the losses on the GPU division for a few more years and push through.
Hah, I remember the time I spent a week dodging the software renderer to figure out why my shader wasn't working on an intel iGPU -- turns out that someone at intel decided to implement "float" with an 8 bit float. Not 8 byte, 8 bit.
"Sure, we support floating point, what's the problem?"
This comment baffles me, both AMD and Intel have CPUs with onboard graphics and those without. You even noted the integrated graphics a sentence later.
If anything, this is more evidence that AMD is following the Intel playbook by having that integrated CPU/ GPU architecture plan.
Why does it baffle you? AMD has only been selling desktop chips with integrated GPUs only for a few years now (they called them APUs), and their APUs were not that stellar at either the GPU or CPU part due to compromises on both parts.
Most of the successful Ryzen chips AMD was selling for the desktop were exclusively without integrated GPUs, to save die space and cost, which hurt PC builders during the GPU scalpocalipse, while on the other hand, Intel's almost entire CPU product range for desktops had integrated GPUs for over 10 years now, enabling PC builders to at least use their PCs until a dGPU could be available.
Sure, Intel sold some CPUs without iGPUs but those were very few SKUs in comparison. Similarly, but in reverse, AMD also sold some Ryzen CPUs with iGPUs(APUs), but those were very few SKUs as their CPUs were weaker than the non-iGPU SKUs, and their outdated Vega iGPUs were pretty weak even compared to Intel's Xe.
So that's the major difference between Intel and AMD that was a game changer for many: Intel shipped most of its chips with iGPUs for over a decade while AMD did not, meaning you always needed to buy a dGPU, and if you couldn't, like in the past ~2 years, well ... good luck, your new tower PC is now an expensive door stop.
Still baffled?
https://en.m.wikipedia.org/wiki/Intel_Graphics_Technology (introduced 2010)
No, and actually in some respects that's not a good thing either. Their existing iGPU driver was designed with the assumption of GPU and CPU memory being pretty much fungible, and with the CPU being pretty "close" in terms of latency, just across the ringbus. It wasn't even PCIe attached, like how AMD does it, it was directly on the ringbus like another core.
Now you need to take that legacy codebase and refactor it to have a conception of where data lives and how computation needs to proceed in order to feed the GPU with the minimum number of trips across the bus. Is that easier than writing a clean driver from scratch, and pulling in specific bits that you need? ....
One of their recent bugs in raytracing was literally due to one line of code in an allocator that was missing a flag to allocate the space in GPU memory instead of CPU, a one-line change produced a 100x speedup in the raytracing performance.
https://www.phoronix.com/news/Intel-Vulkan-RT-100x-Improve
It is most likely much easier to do what AMD did and go from discrete to integrated than the other way around... again, they don't have a tightly-coupled design like Intel did, their iGPU is literally just a pcie client that happens to be on the same die.
(also, AMD paid the penalty like 15 years ago... there were terascale based APUs, and the GCN driver was developed as a dGPU/iGPU hybrid architecture from day 1, they never had to backport GCN itself from dGPU, that was all done in the terascale days.)
Note I'm not an AMD fanboy, but I've 3D in my lifetime was popularized by 3DFX, so I remember the entire saga.
OEMs/ODMs/System Integrators get awards for shipping intel products, but instead of pure cash like in 2000 its hardware and rebates. Until very recently intel was giving out SSDs for free, probably switched to something else since they dont make those anymore.
All three AMD, nvidia and Apple are unifying general purpose compute with graphics. That is the direction the world is going. Unless intel had a new arch trick up their sleeves they will be soon made redundant
Same reason NVIDIA wanted to buy ARM. Intel and NVIDIA are in trouble unless they can make that leap.
Long-term it's the same reason AMD bought ATI too, that vision just took a long time to come to fruition. Remember, they "acquired" their way to success as well, RTG wasn't something that AMD indigeneously developed themselves either... just like AMD bought Xilinx and Intel bought Altera.
It seemed to me scaling the existing architecture to have more processing and bandwidth should have yielded a very competitive GPU while reusing their existing talent pool.
Instead we end up with a very buggy, very specialized (needs resizable bar?, DX12 works well but not other runtimes?) that aren't required today for Intel graphics.
We are in 2022 I still dont understand why Raja is popular. One of the reason why I have been extremely sceptical of Intel's GPU since the very beginning. To the point I got a lot of bashing on Anandtech and HN.
And I have been mentioning drivers as the major concern since 2016. Citing PowerVR Kyro on Desktop as an example. Even pointing that out on Twitter. With one of the Intel Engineers on the GPU team replied something as "GPU Drivers is a solved problem".
I do love to be wrong. But it is increasingly looking like another item to be added to my book of prediction that came true.
and he's done the exact same thing.
That's how failing upwards works in this racket. As long as you have a great looking resume at a few giants in the industry with some fancy titles to boot, you're set for life regardless of how incompetent you are.
I dont think Intel should axe their GPU, but considering Pat Gelsinger has zero BS tolerance, I would not be suspired if someone replace Raja if he doesn't deliver within the next 12 months or so.
There needs to be accountability for these failures of execution.
> He became the director of advanced technology development at ATI Technologies in 2001.[3] Following Advanced Micro Devices's 2006 acquisition of ATI, he served as chief technology officer for graphics at AMD until 2009. At S3 and ATI he made key contributions to several generations of GPU architectures that evolved from DirectX Ver 3 till Ver 11.[4] He then went to Apple Inc., where he worked with graphics hardware, which allowed Apple to transition to high-resolution Retina displays for its Mac computers.[5]
So he has a history of launching some really good products but I'd say it's been at least a decade since he's been involved in anything industry leading.
Really struggling to see this leading to a successful outcome - even if they produce a reasonable product it’s likely to be third placed which is not a comfortable place to be.
Their issue at the moment isn't strategy. It's execution. Axing their GPU division would only hurt their current plan, and do nothing to fix the the systematic problem that they're missing deadline's and shipping incomplete products. From the outside looking in, it seems like there's some fat that needs trimming and people aren't pulling their weight. If they can scale back to efficient team and org sizes, cut the side projects, and focus on excellent software and hardware validation, I can see them pulling this off and Pat being lauded as a hero.
There is such a long term history of failure here that somebody has to make a very strong case that the next time is going to be different and I've never seen anyone at Intel try that or even recognize that history of failure.
It also claims to do this across radically different architectures like FPGA and all I can say to that is "I find that very hard to believe"
One common instance is, blender's Cycles renderer. Every time there is a new NVIDIA GPU. It needs to be recompiled to support it. Sometimes that also requires a new CUDA version to be able to do it. CUDA versions over the years have deprecated different operations and what not.
Although for completeness I'll note that Intel's GPU architecture is documented: https://01.org/linuxgraphics/documentation/hardware-specific...
https://gpuocelot.gatech.edu/
I could see Apple and AMD, working with TSMCs latest node, stepping up to the challenge.
The fact that a similar API can be used for training on servers, inference from laptops to phones, is an appealing proposition
Best part: most devices have decent vulkan drivers. Unlike openCL.
I was also going to mention some dev boards NVIDIA had come out in the last few years that were affordable (I think) but they are all out of stock now and the ones you can get are $2000+ now.
Nowadays only Chinese seem to be able to give it away to grow market share. $1 Espressif is a good example, Govin $7 GW1N-1 fpga devboards. Raspberry <$1 RP2040 is one western exception I could come up with.
But yes, that probably does not make it a psychologically safe place to work...
Meh, not really. Google has some of the best engineers in the world yet they fail at nearly everything that isn't related to search, android, ads and e-mail.
Even Jensen Huang said that it's more about vision and not all about execution. He said in the early days of Nvidia all of their competitors at the time had engineers good enough to execute, yet only they had the winning vision that enabled them to make and sell exactly what the market wanted, nothing more, nothing less.
Yeah they do, Intel has the FPGA division they bought from Altera. That's a huge business on its own.
And my point still stands: having great engineers is not enough for great execution. You need great leadership with a vision a-la Steve Jobs or Jensen Huang.
They have competitive NICs, although they don't seem to be maintaining the lead there they once had.
They bought competitive network switches. These have largely languished, in part because they sat on the IP and then targeted it at weird niches.
They bought Altera. I feel like it had lost some momentum vs. Xilinx, but with AMD acquiring the latter, it's probably going to end up a wash.
The AI chips are kinda too early to tell, but at least they're playing the game.
Overall, I think they have squandered the massive advantage they had in CPUs for the last 3 decades.
The low end AI chips are a mess. Myriad-X can be used only through openvino, and they've been closing details about the internals... Keembay is... Years late?
Is there anything coming out of Intel these days?
And what happened to nervana systems, became plaidml then disappeared after bought by Intel? Maxas was really great and now, crickets.
Even profiling tools, which they used to be on top of, don't seem to work well on Linux w/ the TigerLake gpu. It's so painful to debug and program, they might as well have not put it in...
That's a reflection of their company culture. They're a bit stiff.
I've noticed it's very hard for a lot of technical people to give credit to executives, marketing, finance, or anyone else when it comes to the success of a technical product, but very easy for them to assign blame to those groups. Whereas those latter groups often seem to be the first to credit engineers with successful products, and willing to share blame for failures.
Maybe it's because I'm in the technical group so I hear insider gossip and rantings, and the other groups project an overly generous facade, but unfortunately sometimes it doesn't feel like it. Ask a sample of nerds about failed projects near to their hearts. DEC, Sun, Boeing, whatever Google chat app they liked, the failed company or project they worked on, etc. 9 times out of 10 you'll hear rantings about greedy management, incompetent finance, the idiots in marketing, etc. who wantonly desecrated the Mona Lisas and Sistine Chapels of Engineering.
Many engineers are very good at taking a problem given to them by their company or one they've dreamed up and optimizing the hell out of it. Very few are capable of identifying problems that will be valued by others, and coming up with approaches to solving them in ways that are cost competitive and marketable. Very few people at all can do that well because it requires a measure of creativity and multi disciplinary vision, but at least executives are supposed to be thinking about these things so they have a chance.
And YouTube.
No wonder they are 10X as valuable as Intel
It was Google engineering that allowed YT to scale.
Google video engineering wasn't up to snuff and management decided to move fast and just acquire YouTube.
So hiring the best 1337-coders in the world didn't automatically give them the best product? Shocking.
When YouTube joined Google, their engineers were shocked to find that Google engineers were writing 1000s of lines of server side C++ code, that could be accomplished in 100s of lines of python code. Which explained why YouTube was able to move a lot faster and pump out features, while Google video was struggling to keep up.
It was not the leetcode engineers, but Dean and Ghemawat insisting on C++ only on the server. Its still true today. Go might be making inroads now.
That and their culture of "move fast, fail fast" means they pull the rug just as fast as they lay it out.
> Meh, not really. Google has some of the best engineers in the world yet they fail at nearly everything that isn't related to search, android, ads and e-mail.
Seems to me that Google didn't fail to execute in other things they tried. The products worked fine, but they didn't have a market.
That's not failure of execution, it's failure of strategy.
And the reality is that yeah, with their comp for engineers so abysmally bad why would anyone really go there? Especially when it's not like they're getting great WLB.
Ay, there's the rub!
I mean, "make the fastest CPUs and GPUs and charge a premium" would be a great strategy if they could deliver. But in a world where they're struggling to keep up on the CPU side, pouring resources into GPUs could lose them both markets.
I can't remember the last time I sincerely used an integrated GPU, Intel or AMD, since they're too weak for my purposes. I only ever use them as a Lowest Common Denominator failsafe when troubleshooting.
And for use cases (not necessarily mine) where low-end GPUs make sense, it's not like Nvidia and AMD don't sell bottom of the barrel GPUs that you (or OEMs) can buy for a Franklin each or something.
None of these groups are remotely as close to Intel's core (pun intended) business, as this one. It is certainly true that this division needs to produce better results, but that's true of the company as well. Chopping this one, unless there is some plan for a replacement strategy in the GPU space, would be a bad sign, as it would suggest a company circling the drain/milking the existing winners, unable to make any new winners.
Who are the six startups? It mentions four are in China, and two in the US.
https://www.mobygames.com/images/covers/l/51358-pod-windows-...
https://www.hardware.fr/articles/95-5/jeux-optimises.html
https://news.ycombinator.com/item?id=28237085 "I can tell you that Intel gave companies $1 million for "Optimized" games for marketing such."
Intel needs to buy talent, make shit loads of mistakes, try as hard as they can to squeeze as much driver optimization as they can in reduce the driver gap as much as possible.
Intel has to slog through this. Forget about CUDA support - Apple is going to kill that monopoly for them.
Intel is way too quick to give up. Imagine if they had continued to make ARM CPUs. Apple may have never pursued making their own chips. Intel CPU's + modems could be competing with Qualcomm for every mobile device sold today.
This phrase does it, this guy has no clue whatsoever. Baffling.
In my opinion, when they wanted to get back into graphics, they should have brought back Xeon Phi, maybe doubled the vector width (and added some special units), and hired some engineers from HFT firms to figure out how to make it pretend to be a fast GPU.
Err... what? Who's got the $billions to compete with NVIDIA and AMD?
If Intel couldn't do it with $3.5B invested and existing silicon design experience and a bunch of fabs... who can!?