My favorite part is they claim x86 has “low memory bandwidth” yet they have the exact same bandwidth as EPYC Rome. I will continue to be skeptical until they offer something concrete and testable.
AMD Ryzen series, does have quite some latency, due to their auto frequency settings with xmp timings and infinity fabric and core auto overboost; much more than Intel, which is why AMD absolutely NEEDS fast memory to perform. This latency can lead up to several seconds, before auto config has stabilized, and the system runs pretty optimal and lag-free. It doesn't affect compute by much, save for the beginning load times, but it does affect VMs, and adds to the latency of cloud interactions. I guess they're saying that this will not be the case with thunderx3.
Six months ago I lost my job and after that I was fortunate enough to stumble upon a great website which literally saved me• I started working for them online and in a short time after I've started averaging 15k a month••• iｃash68.ｃｏM
Basically why it hasn't taken off is nobody is developing on those systems. Any benefit that ARM provides is negated by the fact that you need to change everything that is already developed, so there is no cost benefit. People will pay more for an x86 box simply because it's what they developed their load on. He points to example of why RISC vendors died off.
Same on software side. There was no cross-development. It's too costly and relatively painful. And developers go to where the hardware and software already exit and is easily develop on.
Yup. People keep going on and on about how Arm hardware is so much cheaper, when hardware is only a portion - often small - of the TCO. x86 is entrenched and the inertia to overcome that entrenchment is massive. Hence why the only companies that are actually interested in Arm servers are the companies that don't have to pay that massive software debt - i.e. hosting providers like Amazon.
I think the idea was the platform provider would port the entire toolchain, back when AMD bought SuperMICRO. The platform providers need to port the entire platform to ARM. Like the OS, the database software, the entire software ecosystem so when folks are selling SAAS, it doesnt matter what the CPU type is, because the customer doesn't care.
Look at Android SW development, you dont know what the base CPU Type is. The linker deals with it.
Ultra dense with high IO is mainframe IMO. The mainframe space has blurred a lot over the last decade with many new companies entering the market with these ultra dense x86 servers that behave very similar to mainframes of the past. Maybe the IO isn't quite high enough to qualify as mainframe but they are close enough I wouldn't qualify it.
That’s fine. I don’t think ARM makers expect total domination upon release.
I think this is a 5-10 year shift.
You first need significantly better hardware to convince some software makers to switch. That’s how it gets started. You also have an 800 gorilla in Amazon pushing this.
This is a never will be shift. It's will have its targeted areas and no it will not replace the advancements of and the massive ones to come from AMD, never mind Intel.
If ARM and it's licensees would want to jumpstart development, make a thousand or more free workstation-type systems with your newest and hottest server chips available longer-term free of charge for the top developers in that space. Even the big boys will start playing with a free toy. It would mean even more for smaller, independent developers, and don't forget to give special consideration to those individuals involved in open-source software, starting with Linux. I know such a "giveaway/free loan" would cost tens of millions of dollars, but it'd give them buzz and likely help getting that software out there.
Unlikely. There aren't just a handful of software companies. Even if they did get working with the top 10 commercial applications they'd still be handicapped by the broad support x86 offers.
For Nvidia, ARM is rapidly becoming do or die. AMD CPUs are no longer a joke, and Intel is pushing its dGPU ambitions to survive as well. IBM is faltering without a good foundry, and they chose their old partner Samsung, who is still largely unproven at HPC equipment.
Nvidia is out in the cold, here, and are desperate enough to have ported their CUDA development stack to ARM and are now trying to roadmap the Jetson series. None of that matters, IMO, without a laptop somewhere, and I don't see Nvidia entering that market, ever. The other ARM laptop attempts by MSFT/Qualcomm has been largely misguided whiffs.
This is all assuming Nvidia can even get the ARM CPU developers to follow their roadmap better than IBM could, or if the ARM CPU companies even want to.
NVIDIA was supposedly out in the cold ever since Intel declared war on them and AMD bought ATI 14? years ago. AMD or Intel are going to have to improve their GPU ecosystem a whole lot to convince people to use their GPUs for commercial compute. It's very easy to attach an NVIDIA GPU with an Intel processor or an NVIDIA GPU with an AMD processor. In fact, if one is interested in GPU compute, it is the GPU and its ecosystem that is what's important. One can use an AMD or an Intel platform with the NVIDIA GPU.
NVIDIA expanding their software to ARM is not desperate, it's just a natural expansion of their ecosystem. Now it is on x86, OpenPower, and ARM. If RISC-V comes along they will support that, too.
Oh, and you seem to be under the impression that NVIDIA's GPUs are primarily linked to IBM's CPUs. IBM CPUs account for a very small number of CPUs hosting NVIDIA's data center GPUs. Almost all are Intel processors. NVIDIA will have access to the host bus with CXL just like they do with IBM now. Intel tried to keep NVIDIA off their bus, but once Intel went to GPUs they could no longer do that. NVIDIA's partnership with IBM has mostly to do with supercomputer contracts, which is high profile but both a very small part of the market and also a part of the market that is much, much easier to address. In supercomputing they use compiler directives to target accelerators. In commercial applications they use a whole lot of CUDA code.
"n supercomputing they use compiler directives to target accelerators. "
it's called PRAGMA, and why do folks make such a freaking big deal about different cpu. it's a solved problem for a really long time. now, actually finding all the bits and pieces that need be flagged well could be. but the tools are there.
>> The other ARM laptop attempts by MSFT/Qualcomm has been largely misguided whiffs. I have Galaxy Book S and it is awesome. A76 is enough to emulate x86 apps without a hassle. Don't yet have any native apps except benchmarks. Visual studio and dev software works fine.
They actually ported Visual Studio to it? Nice. I'll have to take a closer look at future ARM laptops. I do personally like the idea, since I don't really do laptop gaming (even though I wish I did, for whatever reason). However, I still have too many tools (Altium Designer in particular) that don't really play nice with ARM. Maybe if I move to a more purely software focus, I can make the jump :D
1. You need to reread the whole thread and not take it out of context. 2. Amazon is moving all of their SaaS and Tools over to ARM AWS Instances. 3. He did said unless there are ARM on PC or Mac....
I feel that these arguments are getting less and less relevant, because most devices we are using on a day-to-day basis are already ARM-based. Billions are invested in the development of ARM-based servers and the software ecosystem, and eventually this will manifest itself in numbers.
"Any benefit that ARM provides is negated by the fact that you need to change everything that is already developed, so there is no cost benefit. "
Problem is that this is nonsense. For example, all the code I write can run on ARM out of the box - I don't even think I'd have to change a compiler flag (though i could be wrong on that).
If this stuff could get picked up by a major cloud provider, and provides actual cost benefit, then I could see a lot of workloads moving over one at a time. Heck, just imagine if Amazon decided to move all thier loadbalancing across to ARM - a single, simple workload, but would probably really help establish ARM as a viable server solution.
I think the tables are turning. A lot of cheap thunderx2 CPUs on sites like eBay, but no motherboards to hold them. I guess companies are hot swapping the CPUs. Once China (ASRock?) Would make mobos for the end user (CPU, 2 channel DDR4, 1x pcie x16 slot, USB, lan), people will start buying that hardware.
TX3: 16DP FLOPS@3GHz * 96 cores = 4.608 TFLOPS A64FX: 2*16DP FLOPS @ 2.2GHz * 48 cores = 3.3792 TFLOPS Also ThunderX3 can access a lot of RAM, not just 32GB
FLOPS are great until you don't have the BW to feed them. And judging by the amount of GPU-based entries in top500, access to massive amounts of memory (per node) does not seem to bother HPC much. Let's wait and see though.
You are quoting SPEC performance yet there are precisely ZERO submitted SPEC2017 results on Arm architectures at spec.org. Until you submit such results they remain useless!
Doesn't seem odd to me. Andrei is certainly right -- submission of results does not change the results. And in general, if those results can be reproduced, why should official submission matter?
Zen2's excellent prefetchers and large caches help mitigate that chiplet disadvantage. This seems overly optimistic to me. Be interested in seeing how it competes with Milan.
what I expect is that someone, may haps ARM, will take to SCM and build a registerless, cacheless, bufferless cpu (in just a few million transistors) that looks a lot like the olde TI-990. everything is done directly in memory. could (not sure, "will") blow the doors off everything.
No word on die size it seems, but it should be *quite* big. SMT-4 and these 4 SIMD blocks require extra die area, and this is a freaking 96-core CPU. They might have needed to use less L2 and L3 cache to keep the die size at a sane level.
Relevant story: I work for an enterprise SAAS shop. Our SAAS is java. For the first time, it is looking like non-x86 is the best price/performance for our software (on Graviton2). So ... we're in test with this now, and it seems likely we will go ARM and never go back assuming their future performance claims hold up.
But the point of this is: there is a huge, really huge, amount of java software out there. It has been the most popular platform for more than a decade now. And it is trivial to move that software to ARM.
Basically anything that abstracts away* from the "quirks" of a platform makes cross-movement easier. Good to see because it prevents Wintel like situations.
There are numerous gauges of the mailing framework, the five star mail ($27.3 million) and the standard mail ($18 million) are the most mainstream ones from the parcel. The all out number of <a href="https://www.savonlinnafestivals.com/liteblue/"... workers in the association are 7.5 million and checking. Out of these around 500,000 are profession laborers and around 150,000 are the individuals who are non-vocation workers.
I have to LOL at all the commenters who can't wrap their head around the fact that Amazon, Microsoft, Google, Facebook and more are spending 10's of billions a year on hardware, power and cooling, and think that a billion dollar outlay to get this kind of competitive advantage deployed in their DCs and on workloads where it saves money over the long run wouldnt happen. Not a clue. This is not a technology development meant to replace your PC my friends. The market is far larger than that. They don't give a **** about your PC, phone or tablet, they care about how efficiently they can get you what you want.
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
Facebook
Twitter
RSS
44 Comments
Back to Article
RallJ - Monday, March 16, 2020 - link
Just marketing fluff no real info. Far behind the N1. This is DOA.eek2121 - Monday, March 16, 2020 - link
My favorite part is they claim x86 has “low memory bandwidth” yet they have the exact same bandwidth as EPYC Rome. I will continue to be skeptical until they offer something concrete and testable.ProDigit - Sunday, April 26, 2020 - link
AMD Ryzen series, does have quite some latency, due to their auto frequency settings with xmp timings and infinity fabric and core auto overboost; much more than Intel, which is why AMD absolutely NEEDS fast memory to perform.This latency can lead up to several seconds, before auto config has stabilized, and the system runs pretty optimal and lag-free.
It doesn't affect compute by much, save for the beginning load times, but it does affect VMs, and adds to the latency of cloud interactions.
I guess they're saying that this will not be the case with thunderx3.
dianajmclean6 - Monday, March 23, 2020 - link
Six months ago I lost my job and after that I was fortunate enough to stumble upon a great website which literally saved me• I started working for them online and in a short time after I've started averaging 15k a month••• iｃash68.ｃｏMDug - Monday, March 16, 2020 - link
Interesting read by Linus Torvalds. https://www.realworldtech.com/forum/?threadid=1834...\Basically why it hasn't taken off is nobody is developing on those systems. Any benefit that ARM provides is negated by the fact that you need to change everything that is already developed, so there is no cost benefit. People will pay more for an x86 box simply because it's what they developed their load on. He points to example of why RISC vendors died off.
Same on software side. There was no cross-development. It's too costly and relatively painful. And developers go to where the hardware and software already exit and is easily develop on.
The_Assimilator - Monday, March 16, 2020 - link
Yup. People keep going on and on about how Arm hardware is so much cheaper, when hardware is only a portion - often small - of the TCO. x86 is entrenched and the inertia to overcome that entrenchment is massive. Hence why the only companies that are actually interested in Arm servers are the companies that don't have to pay that massive software debt - i.e. hosting providers like Amazon.webdoctors - Monday, March 16, 2020 - link
This has been known for more than 10 yrs.I think the idea was the platform provider would port the entire toolchain, back when AMD bought SuperMICRO. The platform providers need to port the entire platform to ARM. Like the OS, the database software, the entire software ecosystem so when folks are selling SAAS, it doesnt matter what the CPU type is, because the customer doesn't care.
Look at Android SW development, you dont know what the base CPU Type is. The linker deals with it.
rahvin - Monday, March 16, 2020 - link
AMD never purchased supermicro. AMD did buy a company called SeaMicro that develops large mainframe style computers.questionlp - Monday, March 16, 2020 - link
SeaMicro had built ultra-dense micro servers, not mainframe-style computers. They were shut down by AMD a couple of years ago, I think.rahvin - Tuesday, March 17, 2020 - link
Ultra dense with high IO is mainframe IMO. The mainframe space has blurred a lot over the last decade with many new companies entering the market with these ultra dense x86 servers that behave very similar to mainframes of the past. Maybe the IO isn't quite high enough to qualify as mainframe but they are close enough I wouldn't qualify it.senttoschool - Monday, March 16, 2020 - link
That’s fine. I don’t think ARM makers expect total domination upon release.I think this is a 5-10 year shift.
You first need significantly better hardware to convince some software makers to switch. That’s how it gets started. You also have an 800 gorilla in Amazon pushing this.
mdriftmeyer - Tuesday, March 17, 2020 - link
This is a never will be shift. It's will have its targeted areas and no it will not replace the advancements of and the massive ones to come from AMD, never mind Intel.eek2121 - Monday, March 16, 2020 - link
Linux runs just fine on ARM.eastcoast_pete - Monday, March 16, 2020 - link
If ARM and it's licensees would want to jumpstart development, make a thousand or more free workstation-type systems with your newest and hottest server chips available longer-term free of charge for the top developers in that space. Even the big boys will start playing with a free toy. It would mean even more for smaller, independent developers, and don't forget to give special consideration to those individuals involved in open-source software, starting with Linux. I know such a "giveaway/free loan" would cost tens of millions of dollars, but it'd give them buzz and likely help getting that software out there.rahvin - Monday, March 16, 2020 - link
Unlikely. There aren't just a handful of software companies. Even if they did get working with the top 10 commercial applications they'd still be handicapped by the broad support x86 offers.jeremyshaw - Monday, March 16, 2020 - link
For Nvidia, ARM is rapidly becoming do or die. AMD CPUs are no longer a joke, and Intel is pushing its dGPU ambitions to survive as well. IBM is faltering without a good foundry, and they chose their old partner Samsung, who is still largely unproven at HPC equipment.Nvidia is out in the cold, here, and are desperate enough to have ported their CUDA development stack to ARM and are now trying to roadmap the Jetson series. None of that matters, IMO, without a laptop somewhere, and I don't see Nvidia entering that market, ever. The other ARM laptop attempts by MSFT/Qualcomm has been largely misguided whiffs.
This is all assuming Nvidia can even get the ARM CPU developers to follow their roadmap better than IBM could, or if the ARM CPU companies even want to.
Yojimbo - Monday, March 16, 2020 - link
NVIDIA was supposedly out in the cold ever since Intel declared war on them and AMD bought ATI 14? years ago. AMD or Intel are going to have to improve their GPU ecosystem a whole lot to convince people to use their GPUs for commercial compute. It's very easy to attach an NVIDIA GPU with an Intel processor or an NVIDIA GPU with an AMD processor. In fact, if one is interested in GPU compute, it is the GPU and its ecosystem that is what's important. One can use an AMD or an Intel platform with the NVIDIA GPU.NVIDIA expanding their software to ARM is not desperate, it's just a natural expansion of their ecosystem. Now it is on x86, OpenPower, and ARM. If RISC-V comes along they will support that, too.
Yojimbo - Monday, March 16, 2020 - link
Oh, and you seem to be under the impression that NVIDIA's GPUs are primarily linked to IBM's CPUs. IBM CPUs account for a very small number of CPUs hosting NVIDIA's data center GPUs. Almost all are Intel processors. NVIDIA will have access to the host bus with CXL just like they do with IBM now. Intel tried to keep NVIDIA off their bus, but once Intel went to GPUs they could no longer do that. NVIDIA's partnership with IBM has mostly to do with supercomputer contracts, which is high profile but both a very small part of the market and also a part of the market that is much, much easier to address. In supercomputing they use compiler directives to target accelerators. In commercial applications they use a whole lot of CUDA code.FunBunny2 - Monday, March 16, 2020 - link
"n supercomputing they use compiler directives to target accelerators. "it's called PRAGMA, and why do folks make such a freaking big deal about different cpu. it's a solved problem for a really long time. now, actually finding all the bits and pieces that need be flagged well could be. but the tools are there.
vvid - Tuesday, March 17, 2020 - link
>> The other ARM laptop attempts by MSFT/Qualcomm has been largely misguided whiffs.I have Galaxy Book S and it is awesome. A76 is enough to emulate x86 apps without a hassle. Don't yet have any native apps except benchmarks. Visual studio and dev software works fine.
jeremyshaw - Monday, March 23, 2020 - link
They actually ported Visual Studio to it? Nice. I'll have to take a closer look at future ARM laptops. I do personally like the idea, since I don't really do laptop gaming (even though I wish I did, for whatever reason). However, I still have too many tools (Altium Designer in particular) that don't really play nice with ARM. Maybe if I move to a more purely software focus, I can make the jump :Dksec - Monday, March 16, 2020 - link
1. You need to reread the whole thread and not take it out of context.2. Amazon is moving all of their SaaS and Tools over to ARM AWS Instances.
3. He did said unless there are ARM on PC or Mac....
OreoCookie - Monday, March 16, 2020 - link
I feel that these arguments are getting less and less relevant, because most devices we are using on a day-to-day basis are already ARM-based. Billions are invested in the development of ARM-based servers and the software ecosystem, and eventually this will manifest itself in numbers.NPSF3000 - Sunday, March 29, 2020 - link
"Any benefit that ARM provides is negated by the fact that you need to change everything that is already developed, so there is no cost benefit. "Problem is that this is nonsense. For example, all the code I write can run on ARM out of the box - I don't even think I'd have to change a compiler flag (though i could be wrong on that).
If this stuff could get picked up by a major cloud provider, and provides actual cost benefit, then I could see a lot of workloads moving over one at a time. Heck, just imagine if Amazon decided to move all thier loadbalancing across to ARM - a single, simple workload, but would probably really help establish ARM as a viable server solution.
ProDigit - Sunday, April 26, 2020 - link
I think the tables are turning. A lot of cheap thunderx2 CPUs on sites like eBay, but no motherboards to hold them. I guess companies are hot swapping the CPUs. Once China (ASRock?) Would make mobos for the end user (CPU, 2 channel DDR4, 1x pcie x16 slot, USB, lan), people will start buying that hardware.blu42 - Monday, March 16, 2020 - link
4x 128-bit SIMD @ 204GB/socket/s? Seems like the current green500 leader A64FX has not much to worry about ; )vvid - Tuesday, March 17, 2020 - link
TX3: 16DP FLOPS@3GHz * 96 cores = 4.608 TFLOPSA64FX: 2*16DP FLOPS @ 2.2GHz * 48 cores = 3.3792 TFLOPS
Also ThunderX3 can access a lot of RAM, not just 32GB
blu42 - Wednesday, March 18, 2020 - link
FLOPS are great until you don't have the BW to feed them. And judging by the amount of GPU-based entries in top500, access to massive amounts of memory (per node) does not seem to bother HPC much. Let's wait and see though.Pedant01 - Monday, March 16, 2020 - link
You are quoting SPEC performance yet there are precisely ZERO submitted SPEC2017 results on Arm architectures at spec.org.Until you submit such results they remain useless!
Andrei Frumusanu - Monday, March 16, 2020 - link
We also don't submit any of our SPEC numbers here at AT, does that mean our figures are also useless? It's a silly argument.Dug - Wednesday, March 18, 2020 - link
Not really a silly argument. When there are no results it seems kind of odd.blu42 - Wednesday, March 18, 2020 - link
Doesn't seem odd to me. Andrei is certainly right -- submission of results does not change the results. And in general, if those results can be reproduced, why should official submission matter?AshlayW - Monday, March 16, 2020 - link
Zen2's excellent prefetchers and large caches help mitigate that chiplet disadvantage. This seems overly optimistic to me. Be interested in seeing how it competes with Milan.FunBunny2 - Monday, March 16, 2020 - link
what I expect is that someone, may haps ARM, will take to SCM and build a registerless, cacheless, bufferless cpu (in just a few million transistors) that looks a lot like the olde TI-990. everything is done directly in memory. could (not sure, "will") blow the doors off everything.Santoval - Monday, March 16, 2020 - link
No word on die size it seems, but it should be *quite* big. SMT-4 and these 4 SIMD blocks require extra die area, and this is a freaking 96-core CPU. They might have needed to use less L2 and L3 cache to keep the die size at a sane level.mdriftmeyer - Tuesday, March 17, 2020 - link
I look forward to Zen 3 being revealed soon.carollee - Tuesday, March 17, 2020 - link
Thanks, a lot buddy for such a piece of wonderful information.Check this out, mate: https://www.yourteaminindia.com/blog/flutter-app-d...
drexnx - Tuesday, March 17, 2020 - link
ah yes, another ARM WILL TAKE OVER SOON! piece from ARMdreiARM beating x86 in servers is the cold fusion of microprocessors, always "just a few years away"
bcrules82 - Tuesday, March 17, 2020 - link
This article starts all wrong. ThunderX2+ derive from Broadcom Vulcan IP.https://gcc.gnu.org/legacy-ml/gcc-patches/2016-12/...
Original Cavium ThunderX uarch is in Octeon TX/TX2 product line. And the lowest end TX2 line was Marvell Armada.
surt - Wednesday, March 18, 2020 - link
Relevant story:I work for an enterprise SAAS shop. Our SAAS is java. For the first time, it is looking like non-x86 is the best price/performance for our software (on Graviton2). So ... we're in test with this now, and it seems likely we will go ARM and never go back assuming their future performance claims hold up.
But the point of this is: there is a huge, really huge, amount of java software out there. It has been the most popular platform for more than a decade now. And it is trivial to move that software to ARM.
Threska - Wednesday, March 18, 2020 - link
Basically anything that abstracts away* from the "quirks" of a platform makes cross-movement easier. Good to see because it prevents Wintel like situations.*VMs for the win.
savinnfesals - Thursday, March 19, 2020 - link
There are numerous gauges of the mailing framework, the five star mail ($27.3 million) and the standard mail ($18 million) are the most mainstream ones from the parcel. The all out number of <a href="https://www.savonlinnafestivals.com/liteblue/"... workers in the association are 7.5 million and checking. Out of these around 500,000 are profession laborers and around 150,000 are the individuals who are non-vocation workers.wanderer66 - Thursday, April 2, 2020 - link
I have to LOL at all the commenters who can't wrap their head around the fact that Amazon, Microsoft, Google, Facebook and more are spending 10's of billions a year on hardware, power and cooling, and think that a billion dollar outlay to get this kind of competitive advantage deployed in their DCs and on workloads where it saves money over the long run wouldnt happen. Not a clue. This is not a technology development meant to replace your PC my friends. The market is far larger than that. They don't give a **** about your PC, phone or tablet, they care about how efficiently they can get you what you want.