1) Arc GPUs. Looks like they will be seriously disruptive.
2) Hedging their process node bets by using TSMC too. The only magic in the wins AMD and Apple have had the last few years have been in them being effectively a manufacturing process node ahead of Intel.
3) Intel will likely be first to market with getting a large set of on-package memory that serves as a bridge between CPU cache and DRAM in terms of latency. Think 8GB+
I am no Intel partisan. My core interest is actually more in seeing performant, open (that is, blob-free [1][2]) hardware. RISC-V and Power10 [3][4] are what I am looking at in that regard.
I expect the reports of Intel's impending descent to be largely exaggerated. Still, it is good drama to fuel a hearty compute war. That is to the benefit of all, so have at it.
1: https://raptorcs.com/content/base/faq.html
2: https://www.osnews.com/story/133093/review-blackbird-secure-desktop-a-fully-open-source-modern-power9-workstation-without-any-proprietary-code/
3: https://en.wikipedia.org/wiki/Power10
4: https://www.hpcwire.com/2021/09/08/ibm-introduces-power10-based-server-the-power-e1080-targets-hybrid-cloud/The basic problem is that Intel hasn't lost ground in mobile, it's not even on the playing field, a huge strategic blunder. That leaves it defending its dominant positions on server and desktop / laptop. At the same time billions of $ pour into TSMC from mobile and those same facilities are now being used to make desktop and server CPUs that compete with Intel.
Its execution has been poor too but this massive strategic issue is a bigger problem.
If servers can run on cheap ARM hardware that uses less electricity and runs far cooler without much of a performance difference, that's a massive improvement. Electricity savings would more than justify a switch. Apple has shown that it's possible at the desktop/laptop level. Now they or someone else can prove it works at the server level as well.
For consumers, compatibility is a big issue and hard to switch unless a myriad of things work properly.
For servers, if just a single commonly used system (e.g. mariadb or wordpress) works and you can do it 10% cheaper, companies can replace large quantities of hardware to a completely different architecture.
While I mostly agree with that, the fact that Intel are also moving to a chiplet-style architecture that AMD adopted beforehand means I think there is at least _some_ other bits AMD was ahead of Intel on aside from purely the process node.
I'm nowhere near as knowledgeable about this stuff as others, but from my understanding from Dr Ian Cutress' articles and YouTube channel, Intel is somewhat following AMD in this area. I could be misunderstanding, of course.
Things are hesitantly shifting towards a more AMD-oriented lineup now. The enormous kickbacks and discounts don't (always) weigh up to the increasing performance differences.
Except for all the features on the chips, differing design philosophies, and peripheral interconnects, sure the only difference is manufacturing node. But what have the Romans ever done for us?
I've been hearing the same story for 10+ yrs on how Intel is going to disrupt the GPU market any day now. For real.
Remember Larrabee?
It's not going to happen. Intel is not good at disruption. They're good at several things. But not disruption.
The server marketplace is shifting, slowly, but it is happening.
I'm curious to read more about this. Everything I've seen about Intel's new discrete GPUs makes it sound like it'll be more of the same when compared to AMD and nvidia, and potentially not even as powerful.
They'd been successful for decades, were raking in cash, and remembered the last time they tried something new AKA Itanium AKA the itanic AKA the most mocked architecture of all time. It's a failing, but I can totally understand how they got there.
Everyone seems to be excited about the newer AMD chips which are x86.
ARM is nice for low power, low performance workloads like embedded, phones, and laptops.
I have a lot of criticisms of Intel, but held back from reimagining isn't quite a great one.
Intel knew where their money makers were, x86 servers. Their attention and investment showed a lack of foresight, combined with their famous bloated engineering teams, is a story as old as time itself.
Timing is hard.
i don't recall hearing about arm in other places really... maybe the occasional home router or cable set top box...
Same as when Microsoft had IE6, or the French phone market company was a one provider deal, or when Blackberry was eating the smartphone market.
They stop to innovate. They milk the cow.
It works for some times, then a new guy arrive in town, Firefox for IE6, Free for French Telecom, iPhone for Blackberry...
Imagine, if Intel would swallow their pride and start building ARM SOCs… they could show their engineering prowess, and use X86 synergies to their advantage. As a buyer, I’d buy ARM cpus from intel if they also provided me a slow path for legacy x86 workloads that I haven’t migrated yet. Also, ADM is just the ISA, so, once a customer locks in to intel’s ARM cpu ecosystem, I’m sure they’ll come up with lots of special additions that then the customer would be hesitant to walk away from.
Or, they could be this decade’s IBM and make x86 look like the next generation’s mainframes - niche, even great in some dimensions but out of mainstream use.
I would wager mobile semiconductor revenue worlwide is bigger than server semiconductor revenue worldwide.
But I agree with you that Intel is still dominant in the server world.
https://semianalysis.com/the-semiconductor-heist-of-the-cent...
(discussed recently: https://news.ycombinator.com/item?id=28329731)
IIRC, the problems with ARM China started long before the launch of ARMv9-A. Alibaba has certainly received technical information some time before the public announcement, but it is unlikely that any IP would have been transferred to ARM China before that.
They meant in house ARMv9 design as in Amazon in house Graviton 2. In reality it is an ARM N2 design. Although likely clocked slightly lower due to TDP usage.
I dont think the ARM China issues has been solved ( And possibly never will be ). So the ARMv9 design is likely coming from ARM UK. Alibaba also operate in SEA and expanding outside of China.
The linked video (https://occ-oss-prod.oss-cn-hangzhou.aliyuncs.com/share/risc...) is impressive! Open source Android running on an Open Source CPU.
https://en.wikipedia.org/wiki/Chinese_mythology
https://en.wikipedia.org/wiki/Wuxia
Xuantie probably came from Xuxia, and is widely used in pop culture (games especially) for metals that have special mystical properties.
Picking 玄铁, which also comes from Jin Yong's novel, seems rather natural in that context.
Jin Yong's novels are very popular in Chinese speaking countries and regions. Most readers won't associate them with China. I don't think it has anything to do with national pride. Culture pride? I don't see much either.
Since all these high core count, cheap ARM processors are only available via cloud providers I might end up using spot instances to spin up very large build servers but it feels overkill for what are currently just personal projects (NixOS stuff).
Does that mean that we are going to see soon a data center class RISC-V chip with hundreds of RISC-V cores? That would really be something.
When there are more eyeballs looking at something it can become better.
https://www.anandtech.com/show/15991/hot-chips-2020-live-blo...
I should have one of the eval boards in mid November and I'll be able to do real-word tests then.
The C906 is a little bit faster than a Raspberry Pi Zero, except it has a fairly useful vector unit which can double or triple the speed of many things.
I published memcpy and strcpy benchmarks on Nezha six months ago:
https://hoult.org/d1_memcpy.txt https://hoult.org/d1_strcpy.txt
I also have results for it in my primes benchmark. It beats out a U54 at the same clock speed and is not far off the higher clocked A53 in a Pi 3.
The current "Nezha" board at $99 is obviously expensive compared to a Pi Zero. SiPeed are promising a board with with same D1 SoC with 512 MB RAM for under $20 within the next month.
https://twitter.com/SipeedIO/status/1443486484112183298
The C906 is very comparable to SiFive U54 (as in the HiFive Unleashed, and the Microsemi "Icicle" FPGA) except it has a vector unit and a much better DRAM interface than the FU540 had. But the D1 is only single core.
The C910 is comparable to the SiFive U84, which has not yet been seen in public in actual silicon.
There are getting to be a good number of SoCs reaching actual silicon in small batches, and available on relatively expensive boards ($100 to $665). They work. You can buy them.
Hopefully some of them go into mass production, which should drop the production cost for the actual chips to $5 or so and boards to something like Pi prices (there are dozens of companies that can make boards once chips are available).
Having the core and chip designed and progressed to working test silicon is by far the hardest part already done.
Price then comes down to production volume.
Tens of high quality OSS HDL implementations that target FPGAs and ASICs. Many tens of successful hardware implementations, already shipping multiple billions of cores.
The board Bruce mentions is an amazing value. A board in that form factor that you can bring up a graphical Linux and the distros are already targeting it.
You can get a 4-Stage 160Mhz RV32IMC with 400K of SRAM for $1 in the form of an ESP32-C3. Dev board for $15.
But yes, if you have sufficient funding and expertise you can take their CPU core -- approximately as good as the performance cores in the Samsung Galaxy S8 or the Raspberry Pi 4 -- and without permission or payment or even notification use them in your own chips.
They are some of the best existing RISC-V cores. One of them was the fastest existing RISC-V core when it was introduced.
You can use the Verilog code to either synthesize it for a FPGA and run it in a FPGA board with a large enough FPGA, or if you have access to an ASIC manufacturing process, you can synthesize it for that process and include the RISC-V cores together with whatever else is needed in a custom IC, without paying any royalties.
Using one of these cores for a FPGA board seems very attractive, because most other open-source cores that are available have a much lower performance.
They have also provided versions of the gcc compiler, of the glibc standard C library, of the boot loader, of the Android Bionic standard library and a few other software packages that are needed to run programs for Linux or Android on these RISC-V cores, which have many extensions over the base RISC-V specification.
Alibaba appears to have played for a few years with RISC-V, but even if they have succeeded to design the fastest such cores, eventually they have decided to use ARMv9-A for their real high-performance server CPUs.
There were only 10 publicly available from the first batch, out of I think a total of 80.
I managed to snag one. Hopefully I'll have it mid November.
It pretty clear to anyone that is not a fanboy that both (like is normal in these situations as they're CEOs) have no idea what they're saying.
Impressive. I did not expect riscv to progress so quickly. Intel is much weaker than we all thought after all. Hats off to these clever engineers.
AWS launched their own ARM based processor, Graviton a while ago.
Now they have begun nudging them to serverless fleet by making them available at a lower price point.
Not too different from their parent, Amazon, pushing their own branded products, cutting out the middleman.
Seriously. "More cache on die?" reminds me of the scene in "Smokin' Aces" ... "we did that. all of it."
Intel is ossified so totally that the top i dunno how many layers will have to be brutaly cleaved off it before the remaining pieces can hope to be productive again.
All they have to do is bring these to market without bugs and they'll sell like hotcakes.
Also, Intel's tiling tech is more versatile than chiplets. Though AMD is adopting TSMC's SoIC, which should be comparable. https://www.hardwaretimes.com/intel-believes-its-tiles-are-a...
Most modern "CISC" processors internally are composed of RISC-like elements with instruction decoders in front.
ARM is a specific almost-RISC instruction set architecture.
The original Xbox was codenamed Midway for obvious reasons. :^)
It took all x86 platform warts, and replaced it with even harder to get right "Poulsbo warts" like SFI, and etc.
But timing is hard here too, and lots of things that seem inevitable take many attempts and a long to happen. ARM server chips have near history marked by struggle, there have been several over the last decade. Remember eg AMD in 2012? https://arstechnica.com/information-technology/2012/10/amd-a... Or this ARM Ltd announcement: https://www.tomshardware.com/news/Server-CPU-Xeon-Opteron-AR...
There are still a load of jobs you wouldn't do that for, but the list is short and getting shorter.
Actually if you look at the current top 10 supercomputers, only 2 are Intel. The rest are custom ARM or RISC CPUs (in Japan and China), IBM Power9, or AMD EPYC.
That may be a piece of it, but I think there's a little more. AMD released the first generation of Epyc processors in June 2017. They didn't see much adoption for years.[1]
[1]: https://cdn.mos.cms.futurecdn.net/vono9miWH83ZVeqRfE9CCV-970...
Is this finally an accepted truth now? I have been banging on about it for years that AMD are not doing good enough on Server. They will still gain a little more market share with Zen 4, but now Intel is coming back. The perfect opportunity windows is closing down.
But the open ISA levels the playing field and allows for upstart hardware designers to make compatible hardware that is royalty and/or blob free.
Following the source code for the C910 core upload 24 hours ago, Olof Kindgren is live-tweeting getting it going in an FPGA using the existing FuseSoC framework. He made significant progress in the first session before going to bed. It would not surprise me if it's working in the next 24 hours.
That's a big "when". The road from an ISA to a core is long, and unless the core is copylefted (this isn't), then you aren't going to get the source code for something someone else manufactured either.
(2) having researchers with anonymous retail samples verify through decapping and inspection. I believe der8auer in Germany does die shots -- it would not be a bad idea to kick start this kind of research for community assurance purposes.
It's difficult to prove there's no hidden logic, but it's also not trivial to hide complex logic needed to introduce covert undetectable vulnerabilities (probably around things the RNG source or crypto).
Also assuming you're a high value target, otherwise this is mostly going too far.
And there is certainly growing interest in blob-free computing [1], so some at least will exist to fill that demand. There is some hope for video with Linux landing blob-free hardware encode/decode very quickly the last couple of years [2].
1: https://www.crowdsupply.com/mnt/reform/updates/post-campaign-orders
2: https://www.youtube.com/watch?v=E9JLxjYlIWgArm servers have been sold for some years now. AWS, Oracle and others already made the switch. And the fastest super computer (Fugaku, Japan) uses ARM.
The fact that apple is making arm mainstream means if I switch to arm, so are a lot of others working to make the arm port of my stack a first class consideration.
Suddenly the gavatron cores aws has are becoming a whole lot more viable. I personally plan to move our api instances over to them once the erlang vm's arm port matures a bit.
Phones bit of a mixed bag but their computing gear has always been a bit old hat since the 1980s.
The M1 is the first time since PowerPC they took a risk. My guess is that it will turn out the same PowerPC. I really want a linux powered ARM workstation but Apple aren't going to make that.
PowerPC was developed by IBM. Apple got involved when Motorola could not deliver a faster 68k to beat Intel's 486 which was passing 50MHz. So they joined IBM along with Motorola and formed the AIM alliance (Apple, IBM, Motorola) to build better processors.
> The M1 is the first time since PowerPC they took a risk.
There's little to no risc (buh-dum-tish) in moving to Arm these days. It's a well supported and understood architecture and found everywhere.
> I really want a linux powered ARM workstation but Apple aren't going to make that.
Apple has the power to steer its own ecosystem. When AMD was working on its Arm A series server processors I kept thinking "This sounds like a backwards approach destined to fail. Why not start by making a performant Arm SoC with 2-4 cores and GPU with a TDP of 10-20W? Target it at consoles/tv/laptops/desktop computing and jump start the desktop Arm market which will naturally lead to demand for Arm servers." The Idea was an Arm SoC that could fill the gap between low power/performance Arm SoC's for mobile/embedded and the power hungry yet performant x86 chips. Basically an AMD version of the M1. That could have really changed things but the big issue AMD would face is where's the Arm Desktop software ecosystem? That's why Apple can take these "risks", they have full control over the whole stack.
The complexity of developing validation tests suites for PowerPC alone would have sucked up all the software resources inside of Motorola at the time, as all the old 68k OSes and support code, etc had to be rebuilt from scratch for PowerPC. Not at all a small undertaking.
The 6502 was introduced at Wescon in September 1975.
Apple I was out in April 1976. That's seven months.
The KIM1, a board made by MOS to demonstrate their 6502 chip, was also released in April 1976. Even they didn't beat Apple to it.
Commodore Pet was December 1977. Rockwell AIM65 was 1978. Acorn System 1 was March 1979. Atari 400 was November 1979
In short: I don't know what the heck you're talking about.
Similarly, the Lisa was a very early 68000 machine. The Amiga and Atari ST were years after the Lisa and Mac. Only very expensive workstations from HP, Apollo and Sun were before the Apple Lisa.
If the pattern holds the M1 is due to be a dead end.
Not always. Firewire, for example.
Besides that it's fun to debug with on older systems, or even new ones if they have it. If not it can be plugged in from anything between 30 to 50 universal credit units from 1 to 4 port Firewire800 as PCIe 1x to 4x. It's fun to have in a homelab. You can even tunnel IP over it.
All of these ARM server chips are going to be very low on the pecking order for TSMC fab time, so they won't have the edge that Apple had by buying up the first slot on the latest manufacturing node. Even being a process node ahead, M1 only just squeaks by with comparable single threaded performance to Zen 3 cores, so ARM still has some catching up to do in the performance realm.
I don't think this is true. It is not only power consumption but also power backup. If you can use smaller diesel engines and smaller battery packs this will lower the cost.
And why do you think datacenters have been raising temperatures? It saves a ton of money: https://www.datacenterknowledge.com/archives/2008/10/14/goog... HP estimated they saved 8 million. That is not a rounding error.
A server that uses less power will also generate less heat.
I have to admit it is years ago I worked for a company owning datacenters but at that time the highest costs were always: power and connectivity.
100,000 sq. ft. is roughly enough space for about 250,000 1U servers. Say conservatively the servers are in the ballpark of $2,000 each. That's $500 million just in server hardware costs. If the total electric cost is around 3*$8 = ~$24 million over their lifetime, then we're talking about <5% of just the server expenses. Never mind the facility and staff costs (and, as you've said, connectivity).
So maybe not a rounding error, but way down the priorities list.
There's a caveat to that in the planning phase of a data center: A lot of the site infrastructure costs are a function of the total power requirement. So if - when building out a data center - you can get more power efficiency, then that does translate to a significant cost savings.
Cooling tends to be somewhere between a 20 and 60% add on to the direct power consumption of a server.
1: https://www.missioncriticalmagazine.com/ext/resources/MC/Home/Files/PDFs/(TUI3011B)SimpleModelDetermingTrueTCO.pdfhttps://www.irishtimes.com/news/politics/data-centres-could-...
Norway seems poised to build out significantly. They're actively inviting datacenters [1]. They also seem to have lower electric costs for large businesses [2] vs Ireland [3]. FAANG are already populating the Nordics too [4].
1: https://www.datacenterdynamics.com/en/news/norway-wants-data-centers-to-locate-there-and-be-more-sustainable-too/
2: https://www.statista.com/statistics/595859/electricity-industry-price-norway/
3: https://www.statista.com/statistics/595806/electricity-industry-price-ireland/
4: https://www.zdnet.com/pictures/the-nordic-datacenter-boom/The same climate benefits apply to northern England and Scotland, but Great Britain's grid's capacity has around 10 times the capacity.
(Ireland's grid's peak demand was 6.8GW, Great Britain's 63GW.)
Although the same argument then applies for siting the datacentres in continental Europe, where the grid is around 667GW. Perhaps this is part of Facebook's reasoning for a second datacentre in Denmark.
(Most of Denmark, including Facebook's existing datacentre, is connected to the continental European grid. The eastern islands are part of the Scandinavian grid.)
[1] https://www.thelocal.dk/20211013/facebook-eyes-second-danish...
units '500W * 1 year * 0.20 €/kWh' '€'
€ 876
Not as large as some may expect once you factor in power delivery guarantee and redundancy. But basically yes. Every single HyperScaler has been pointing out electricity as one of their largest item on their TCO, 2nd only to hardware cost.
Problem is, that sounds high... the coefficient of performance for chillers is around 4 to 7 [1]. That puts ~15-30% of the total energy demand from chillers, though often datacenters do budget a factor up to about 60% of equipment power demand for cooling power demand. Sum energy related expenses for datacenters tends to be around 10-15% of the total costs [2]. So it would be odd to have such high cooling costs. A TCO over 4 years seems more in line with typical figures.
1: https://www.energy.gov.au/sites/default/files/hvac-factsheet-chiller-efficiency.pdf
2: https://www.missioncriticalmagazine.com/ext/resources/MC/Home/Files/PDFs/(TUI3011B)SimpleModelDetermingTrueTCO.pdf