It is so weird that Intel and AMD blew their headstart so badly. I see so many players who come up with ARM implementations that come close or even exceed the performance / energy envelope of Intel/Amd.
This one counts as exceeding, at least on some measures. The chip has 1TB/s of memory bandwidth from the 48 cores to its on-package 32GB of memory. Looks like the smallest thing you can buy, though, is a 2U server with 8 sockets, 2 of them populated, for $40k.
Fujitsu has been making TOP500 supercomputers for years, maybe decades.
They're not a small player. They used to make UltraSPARCs, but since Sun / Oracle is a toxic wasteland now, its natural for them to move to ARM instead.
I'm not in that business/role, so I don't know how they compare to e.g. Dell or HP; especially regarding support. But I've got one of their older 1U rackmounts in my basement and it's a pretty nice machine.
I use at home since many years just Fujitsu Notebooks and also have an W550 workstation. Also for my mom, I bought for her many years ago a Fujitsu Notebook.
They come without any bloatware and just run and run forever. The design is nice, I like hiw they have mostly a red line over the device. But they are not cheap to buy.
I bought lately an U939X notebook and he is really light, but still everithing (touch and pen) works also with Linux.
https://m.youtube.com/watch?v=hS2uogVMK8U
They have tried. Here in Japan they have traditionally had an unassailable position as a domestic champion. The only serious competition was from other domestic companies such as NEC, Hitachi.
This model has been very successful for them domestically, but it also makes them uniquely unsuited to competing in global markets where they have no brand to speak of.
Another issue is cost vs. quality. Japanese manufacturers including Fujitsu despise low quality and assume their customers are always willing to pay for the best. Don't break things, even if at the expense of moving slow. This assumption generally holds domestically, but not in global markets.
It's not easy to take available data to confirm or reject such a hypothesis Mr. Fnoord.
I would guess this particular iceberg is at least 90% above the surface. It appears to be bog-standard industrial policy which in large part does not necessitate secret stuff. I make this conjecture partly by reference to other industries in Japan, whose comparable government support could certainly not be explained in that way.
I dont follow the PC market as closely now, but they used to compete in PC market. And they were All Made in Japan.
The problem with PC market is how they trained consumers into buying Spec. At first people would compare clock speed, and Memory capacity, later they learned CPU Generation was also important and Memory Speed, and then dozens of other things. And basically becomes a commodity market.
Nearly ALL PC manufacture moved their assembly to China. Fujitsu remained MIJ for as long as I could remember ( And may be that is still the case ).
There is also a culture problem. Which I think matter a lot more as they are factored into their decision making.
Wow. People may dislike IBM for all sort of reason ( Mostly Software ). But the IBM mainframes technology should certainly be something to be marvelled at.
If Fujitsu really did out engineered IBM counterparts, that would be an astonishing achievement.
"toxic wasteland" isn't the reason they gave for moving to Arm, though I don't have a reference to hand. I'm not sure about not missing a beat; as a component of Fukagu, this represents years of serious engineering. Sure, there's no great problem running typical HPC software on aarch64 or ppc64le per se, given suitable BLAS etc.
Fujitsu's previous "K" supercomputer was #1 on TOP500 and GREEN500 when it came out. Heck, it was #1 on Green500 for many, many years after its release. K held #1 on HPCG from 2011 through 2018.
Fujitsu just has good tech. Its kind of mysterious since its aimed at the Japanese market instead of the US market, so we don't hear much about it. But its clearly a world-class processor company.
My impression is that the Fujitsu K-machine && NEC SX-ACE equipment is designed, built, and sold almost exclusively for the Japanese market; courtesy of massive subsidies from the Japanese gov’t for the purpose of maintaining that expertise in-country.
K was basically a one-off, but you could buy systems with similar technology, though they weren't common outside Japan. A64FX machines are appearing in the West now. There was a recent substantial award for one to a UK academic/Met Office consortium, in which there's some interest.
Yeah, by K-machine I mean any of the PrimeHPC-FX class systems using SPARC64-V derivatives; I'm sure they sold a few outside Japan, but I don't think that not having the wider global market would've deterred Fujitsu from making them.
I know some of Fujitsu's A64 modules were making it into Cray systems as well; isn't that what U. Bristol has?
I'd forgotten the name PrimeHPC-FX. You would think they'd be worth looking at if you knew about K. I don't remember exactly what Bristol have got now. (Its not so easy to get detailed info on UK Tier 2 systems, even if you're supporting one.) I know it's using Infiniband, not Tofu as they're mostly interested in single-node performance for some reason, but I don't know what other than the chips come from Fujitsu. It may be right that they're not doing much with systems. There does seem to be more interest this time round, at least.
Fugaku, for which A64FX was designed, has no DDR memory, just HBM, but yes because of bandwidth. Its Tofu fabric has decent bandwidth but particularly low latency.
As I see it, Intel and AMD see the super computer market as "niche" and while it gives good publicity, there are perhaps 1 or 2 such machines sold each year. As a result they spend their time and "innovation beans" on things that make desktop and laptop computers more desirable. IBM's Power architecture is another computer ecosystem that is really innovative and cool but you hardly hear about it. Different target markets.
The second part of the issue is "head start." For literally decades the "cost" of a computer architecture was perhaps 3% chips and 97% software support for those chips. The "Windows" lock (by having the #1 OS run only on x86 architecture machines) meant that Intel really didn't have to do anything for years but shrink their transistor size and they could "win."
Linux took away the Windows "moat", networking took away the "enterprise server" moat, and gcc took away the "compiler/tools" moat. Between the three of them it gave everyone a chance to play if they wanted to stake out a niche.
ARM took "low power" and captured phones for which battery power is key. By licensing the architecture widely and inexpensively they enabled parallel investment that well exceeded anything a single manufacturer could apply to a particular architecture. By some estimates you have over $50 billion dollars / year being spent on R&D that accelerates ARM compared to Intel's $15 billion R&D budget for 2020? Not so much "blowing their head start" as is it getting mowed down by not being able to invest as much as the other guys.
It used to be a big deal when intel machines started to show up and then rule the supercomputer benchmarks. It used to be all specialized chips like IBM Power. It was when Intel was ascendant. Now ARM is doing the same thing.
Arm is attacking Intel at the top and bottom and middle leaving less and less space for them to operate.
Intel lost embedded to arm first. Then the lost mobile. Then they lost the Apple laptops and later this year the Apple desktops. They are starting to lose the super computers. The nail in the coffin will come with the cloud servers if those vendors could get something close to Apple’s custom Arm cpu performance.
The only market left unchallenged for Intel/AMD is the consumer dektop / gaming. If Samsung / Qualcomm/ Mediatek make a new motherboard design and introduce a new ARM cpu with double the single core performance (M1 showed that this is not science fiction), then the x86 folks are in deep trouble.
Cloud isn’t really challenged yet and is the most lucrative market for Intel, just minor inroads from fairly slow arm chips. The problem is single core Arm performance outside of Apple chips isn’t there. I wonder how long that will take to change?
Honestly it would be brilliant for Apple to start offering Mx cpus in the cloud at affordable rates to the various cloud providers.
I think Intel/AMD's cloud presence is definitely getting challenged by ARM and it's momentum is just increasing. A lot of cloud workloads are not CPU bound. More often they're IO bound.
Amazon's Graviton2 and Ampere's Altra are tackling that IO bound space. They have dozens of single thread cores, lots of memory channels, and lots of PCI Express lanes (64 and 128 respectively IIRC). They're not slouches in CPU power but their IO is really impressive.
Having high power and high bandwidth chips at a fraction of the power of Xeons is a win for cloud providers. They can either get higher density in the same power/cooling envelope or reduce their power/cooling requirements for the same capacity. At cloud provider scale power is a major cost driver.
HyperScalers are estimated to be 50% of Server and Cloud Business.
HyperScalers are expanding at a rate faster than other market.
HyperScaler expanding trend are not projected to be slowing down anytime soon.
AWS intends to have all of their own workload and SaaS product running on Graviton / ARM. ( While still providing x86 services to those who needs it )
Google and Microsoft are already gearing up their own ARM offering. Partly confirmed by Marvell's exit of ARM Server.
>The problem is single core Arm performance outside of Apple chips isn’t there.
Cloud computing charges per vCPU. On all current x86 instances, that is one hyper-thread. On AWS Graviton, vCPU = Actual CPU Core. There are plenty of workloads, and large customers like Twitter and Pinterest has tested and shown AWS Graviton 2 vCPU perform better than x86. All while being 30% cheaper. At the end of the day, it is workload / dollars that matters on Cloud computing. And right now in lots of applications Graviton 2 are winning, and in some cases by large margin.
If AWS sell 50% of their services with ARM in 5 years time, that is 25% of Cloud Business Alone. Since it offer a huge competitive advantage Google and Microsoft has no other choice but to join the race. And then there will be enough of a market force for Qualcomm, or may be Marvell to Fab a commodity ARM Server part for the rest of the market.
Which is why I was extremely worried about Intel. The lucrative Server market is basically gone. ( And I haven't factored in AMD yet ) 5 years in Tech hardware is basically 1-2 cycles. And there is nothing on Intel's roadmap that shown they have the chance to compete apart from marketing and sales tactics. Which still goes a long way if I have to be honest, but not sustainable in long term. It is more of a delaying tactics. Along with a CEO that despite trying very hard, had no experience in market and product business. Luckily that is about to change.
Evaluating ARM switch takes time, Software preparation takes time, and more importantly, getting wafer from TSMC takes time as demand from all market are exceeding expectations. But all of them are already in motion, and if these are the kind of response you get from Graviton 2, imagine Graviton 3.
People I work with test my software on MS Azure cloud VM with some recent AMD Epyc CPU.
> for Apple to start offering Mx cpus in the cloud at affordable rates
Large part how these Apple's ARM chips are so fast is integration. Just removing DDR4 wires improves memory latency by multiple CPU cycles due to speed of signal in copper. For servers you probably wouldn't want integration, you'd want DIMMs to replace just the failed parts, with ECC.
Depends on the server. Some server software can be easily scaled horizontally. For them what you proposed indeed can be a good deal.
Other things, like RDBMS servers, can’t do that or at least not so easy. For them you actually want loads or RAM in the servers.
I don’t think it’s practical to manufacture SOCs with hundreds of GB of DDR4 in the package. The package gets too large, also economy of scale is not that great compared to phones or laptops.
Absolutely. That ARM has been gnawing on the dominance of the x86 architecture (and by proxy Intel) first became obvious to me when Intel tried desperately to make an x86 chip that was competitive in the Android ODM space. I still have a Lenovo Android tablet with an Intel processor in it[1]. Intel's business model of being 30% of the BOM cost for the base board just doesn't fly in the phone/tablet world. If I were still at NetApp I would be all over a filer design based on 64 bit multi-core ARM architecture with a bespoke I/O architecture to support 1000+ drives per filer.
The arrival of Intel (and AMD with the Opteron) on the TOP-500 supercomputer list demonstrated the importance of channel bandwidth (memory and I/O).
[1] It sucks BTW, Android gave up trying to fix all of the byte ordering problems and so many of the apps fail to execute correctly.
> The arrival of Intel (and AMD with the Opteron) on the TOP-500 supercomputer list demonstrated the importance of channel bandwidth (memory and I/O).
Top500 is based on the HPL benchmark. Its problems are specifically not being sensitive to memory and networking (specifically network latency). That's why half(?) of it is things like "cloud vendor", where you're not going to run typical HPC codes past three nodes.
There may be only a couple of CORAL-size systems (Summit et al) a year but there are many >~$1M HPC systems sold. Intel have so far failed to deliver for CORAL.
That's true, but microarchitecture docs and optimization manuals aren't like regular documentation. Apart from the intellectual dick-waving, it's just plain interesting.
Apple hasn't even documented their matrix extension, apparently thinking that people shouldn't compile against them directly but instead use their accelerate framework.
Further, they use their private llvm branch and haven't pushed the pipeline description tables for AFAIK any of their arm chips to upstream llvm.
So while I agree it would be interesting, I wouldn't hold my breath.
This is a lot like the manuals that ARM publishes for its IP. For example, for the A76 [1] they publish the Arm Cortex-A76 Software Optimization Guide. But I haven't seen this for any non-ARM ARMv8 implementation such as Apple, NVidia, ...
52 comments
[ 2.9 ms ] story [ 111 ms ] threadThey're not a small player. They used to make UltraSPARCs, but since Sun / Oracle is a toxic wasteland now, its natural for them to move to ARM instead.
The previous Fujitsu supercomputer, the K, was also #1 when it came out in 2011. (https://en.wikipedia.org/wiki/K_computer)
-----------
The main thing that been proven is that a company like Fujitsu can transition between ISAs: the SPARC ISA into the ARM ISA, and not miss a beat.
I'm not in that business/role, so I don't know how they compare to e.g. Dell or HP; especially regarding support. But I've got one of their older 1U rackmounts in my basement and it's a pretty nice machine.
And this? https://www.fujitsu.com/global/products/computing/pc/desktop...
it's been 15 years since I've seen a fujitsu desktop in EU. I mean I regularly see way more panasonic toughbooks around and it's actually rare.
This model has been very successful for them domestically, but it also makes them uniquely unsuited to competing in global markets where they have no brand to speak of.
Another issue is cost vs. quality. Japanese manufacturers including Fujitsu despise low quality and assume their customers are always willing to pay for the best. Don't break things, even if at the expense of moving slow. This assumption generally holds domestically, but not in global markets.
What I said above may be changing. But slowly.
The rumor was NSA still was going with DEC Alpha fabs long after DECs demise.
My government has to buy American/Chinese stuff AFAIK. Housing ASML doesn't give an equal advantage, right?
I would guess this particular iceberg is at least 90% above the surface. It appears to be bog-standard industrial policy which in large part does not necessitate secret stuff. I make this conjecture partly by reference to other industries in Japan, whose comparable government support could certainly not be explained in that way.
The problem with PC market is how they trained consumers into buying Spec. At first people would compare clock speed, and Memory capacity, later they learned CPU Generation was also important and Memory Speed, and then dozens of other things. And basically becomes a commodity market.
Nearly ALL PC manufacture moved their assembly to China. Fujitsu remained MIJ for as long as I could remember ( And may be that is still the case ).
There is also a culture problem. Which I think matter a lot more as they are factored into their decision making.
PCs and mobile devices have recently been moved into separate subsidiaries. Presumably ahead of a divestiture / major reorg.
If Fujitsu really did out engineered IBM counterparts, that would be an astonishing achievement.
Fujitsu just has good tech. Its kind of mysterious since its aimed at the Japanese market instead of the US market, so we don't hear much about it. But its clearly a world-class processor company.
I know some of Fujitsu's A64 modules were making it into Cray systems as well; isn't that what U. Bristol has?
The second part of the issue is "head start." For literally decades the "cost" of a computer architecture was perhaps 3% chips and 97% software support for those chips. The "Windows" lock (by having the #1 OS run only on x86 architecture machines) meant that Intel really didn't have to do anything for years but shrink their transistor size and they could "win."
Linux took away the Windows "moat", networking took away the "enterprise server" moat, and gcc took away the "compiler/tools" moat. Between the three of them it gave everyone a chance to play if they wanted to stake out a niche.
ARM took "low power" and captured phones for which battery power is key. By licensing the architecture widely and inexpensively they enabled parallel investment that well exceeded anything a single manufacturer could apply to a particular architecture. By some estimates you have over $50 billion dollars / year being spent on R&D that accelerates ARM compared to Intel's $15 billion R&D budget for 2020? Not so much "blowing their head start" as is it getting mowed down by not being able to invest as much as the other guys.
Arm is attacking Intel at the top and bottom and middle leaving less and less space for them to operate.
Intel lost embedded to arm first. Then the lost mobile. Then they lost the Apple laptops and later this year the Apple desktops. They are starting to lose the super computers. The nail in the coffin will come with the cloud servers if those vendors could get something close to Apple’s custom Arm cpu performance.
Honestly it would be brilliant for Apple to start offering Mx cpus in the cloud at affordable rates to the various cloud providers.
Amazon's Graviton2 and Ampere's Altra are tackling that IO bound space. They have dozens of single thread cores, lots of memory channels, and lots of PCI Express lanes (64 and 128 respectively IIRC). They're not slouches in CPU power but their IO is really impressive.
Having high power and high bandwidth chips at a fraction of the power of Xeons is a win for cloud providers. They can either get higher density in the same power/cooling envelope or reduce their power/cooling requirements for the same capacity. At cloud provider scale power is a major cost driver.
AWS are estimated to be ~50% of HyperScalers.
HyperScalers are estimated to be 50% of Server and Cloud Business.
HyperScalers are expanding at a rate faster than other market.
HyperScaler expanding trend are not projected to be slowing down anytime soon.
AWS intends to have all of their own workload and SaaS product running on Graviton / ARM. ( While still providing x86 services to those who needs it )
Google and Microsoft are already gearing up their own ARM offering. Partly confirmed by Marvell's exit of ARM Server.
>The problem is single core Arm performance outside of Apple chips isn’t there.
Cloud computing charges per vCPU. On all current x86 instances, that is one hyper-thread. On AWS Graviton, vCPU = Actual CPU Core. There are plenty of workloads, and large customers like Twitter and Pinterest has tested and shown AWS Graviton 2 vCPU perform better than x86. All while being 30% cheaper. At the end of the day, it is workload / dollars that matters on Cloud computing. And right now in lots of applications Graviton 2 are winning, and in some cases by large margin.
If AWS sell 50% of their services with ARM in 5 years time, that is 25% of Cloud Business Alone. Since it offer a huge competitive advantage Google and Microsoft has no other choice but to join the race. And then there will be enough of a market force for Qualcomm, or may be Marvell to Fab a commodity ARM Server part for the rest of the market.
Which is why I was extremely worried about Intel. The lucrative Server market is basically gone. ( And I haven't factored in AMD yet ) 5 years in Tech hardware is basically 1-2 cycles. And there is nothing on Intel's roadmap that shown they have the chance to compete apart from marketing and sales tactics. Which still goes a long way if I have to be honest, but not sustainable in long term. It is more of a delaying tactics. Along with a CEO that despite trying very hard, had no experience in market and product business. Luckily that is about to change.
Evaluating ARM switch takes time, Software preparation takes time, and more importantly, getting wafer from TSMC takes time as demand from all market are exceeding expectations. But all of them are already in motion, and if these are the kind of response you get from Graviton 2, imagine Graviton 3.
People I work with test my software on MS Azure cloud VM with some recent AMD Epyc CPU.
> for Apple to start offering Mx cpus in the cloud at affordable rates
Large part how these Apple's ARM chips are so fast is integration. Just removing DDR4 wires improves memory latency by multiple CPU cycles due to speed of signal in copper. For servers you probably wouldn't want integration, you'd want DIMMs to replace just the failed parts, with ECC.
Other things, like RDBMS servers, can’t do that or at least not so easy. For them you actually want loads or RAM in the servers.
I don’t think it’s practical to manufacture SOCs with hundreds of GB of DDR4 in the package. The package gets too large, also economy of scale is not that great compared to phones or laptops.
The arrival of Intel (and AMD with the Opteron) on the TOP-500 supercomputer list demonstrated the importance of channel bandwidth (memory and I/O).
[1] It sucks BTW, Android gave up trying to fix all of the byte ordering problems and so many of the apps fail to execute correctly.
What byte ordering problems could affect x86??
The only big endian platform Android was ported to was MIPS.
Maybe what you're seeing is apps with native components that are only compiled for armv6/7.
Top500 is based on the HPL benchmark. Its problems are specifically not being sensitive to memory and networking (specifically network latency). That's why half(?) of it is things like "cloud vendor", where you're not going to run typical HPC codes past three nodes.
Further, they use their private llvm branch and haven't pushed the pipeline description tables for AFAIK any of their arm chips to upstream llvm.
So while I agree it would be interesting, I wouldn't hold my breath.
[1] https://developer.arm.com/documentation/swog307215/
(search for TRM)
Xavier has 8 Carmel cores and Parker/X2 has 2 Denver2 cores and 4 Cortex-A57s. The SoC manual includes instruction timings for the NV cores.