Discussion AWS Graviton2 64 vCPU Arm CPU Heightens War of Intel Betrayal

[exquisitechar]

https://www.servethehome.com/aws-graviton2-64-core-arm-cpu-heightens-war-of-intel-betrayal/

Pretty big deal for ARM in servers. Interested in seeing a comparison between this and Rome.

 

[soresu]

Interesting that you have not heard of it. It is -- or was -- a common marketing term.

I often miss the business terms, it just interests me less than the scientific or compute side of things

 

[jpiniero]

Cheap ARM servers using chiplets is an absolute nightmare for both Intel and AMD, because they might be willing to forgo single thread (which is vital to client workloads) to get max MT/mm2/W.

 

[A///]

I often miss the business terms, it just interests me less than the scientific or compute side of things

Loss leaders are everywhere. Especially in big stores like a membership one. Rarely do people go in for one item alone, and even if they did, chances are they'll buy more. It brings in foot traffic, but also pays off in the end for a small loss. Which can be written off.

 

[amrnuke]

Digging this thread up, article link below. While they didn't compare with Rome, the Graviton2 performed really, really well. Excited for a Rome vs Graviton2 matchup.

ARM News, Reviews and Insights | Tom's Hardware

Discover the Tom's Hardware take on the ARM product range, with news, reviews and benchmarking for the hardcore PC enthusiast.

www.anandtech.com

 

[DrMrLordX]

Nice chip. I kinda wish we had some more comparatives like . . . yeah Rome, as well as Huawei's 64c chip and maybe some of the existing ThunderX2 products. And that new 80c Ampere chip though I'm not really sure about availability.

 

[Zstream]

Never understood why Intel never built a cloud. I mean they have all the capabilities..

 

[moinmoin]

Never understood why Intel never built a cloud. I mean they have all the capabilities..

They are a manufacturer, not a service company.

 

[lobz]

Does Amazon have enough volume to justify developing+fabricating this chip? Or do you think they will sell it as a product as well as use it in house?

nah, they're just fooling around /s

 

[CHADBOGA]

Never understood why Intel never built a cloud. I mean they have all the capabilities..

Because Intel are into squandering multiple opportunities that should be easy peasy for them.

Why didn't they become a leading foundry and amortize their own manufacturing costs and do what TSMC has since done as a foundry?

 

[Zstream]

They are a manufacturer, not a service company.

And Amazon just sold books 😉

 

[JasonLD]

And Amazon just sold books 😉

Well...I am sure competing against your biggest customers is such a great idea.

 

[Nothingness]

Why didn't they become a leading foundry and amortize their own manufacturing costs and do what TSMC has since done as a foundry?

This one is easy: they tried and failed miserably because their design rules are so tied to their process that porting any existing design to their process is a royal pain. Add to that their first serious trial (with LG) was supposed to be on 10nm. Boom.

 

[KompuKare]

Why didn't they become a leading foundry and amortize their own manufacturing costs and do what TSMC has since done as a foundry?

As well as what Nothingness said, don't forget that back before mobiles took off, Intel supposedly refused to make a SoC for Apple fearing it was too low margin.
Plus having deliberately crippled Atom almost since inception to segment the market.
These and other factors eventually gave TSMC enough volume to be able to invest massively in their process and fabs.

 

[Tuna-Fish]

It is a big deal, but unless Amazon starts selling Graviton2 to other parties, it's more a story of Intel losing ODM market share than anything else. Intel has bent over backwards to make Amazon, Google, and (presumably) Microsoft happy with custom hardware and early releases. Apparently Amazon isn't happy with IceLake-SP or Cooper Lake.

Amazon would be trying to do this regardless. I've pointed it out before, but the fundamental way Amazon grows to new segments is by looking at what they are spending a lot of money on, and if that something has acceptable margins, they expand to that segment, either by acquisitions or starting projects to do that in-house. They are spending a lot of CPUs -> they hired a cpu design team. That's really all there is to it.

 

[mikegg]

Because Intel are into squandering multiple opportunities that should be easy peasy for them.

Why didn't they become a leading foundry and amortize their own manufacturing costs and do what TSMC has since done as a foundry?

Yes agreed that it made more business sense for Intel to offer chip manufacturing as a service than cloud as a service. Manufacturing is much closer to what Intel does than cloud computing.

TSMC has a slightly bigger market cap than Intel right now. Imagine if Intel saw TSMC's business opportunity. Intel would be worth twice as much right now.

 

[beginner99]

Cheap ARM servers using chiplets is an absolute nightmare for both Intel and AMD, because they might be willing to forgo single thread (which is vital to client workloads) to get max MT/mm2/W.

While I see where you are coming from, mere throughput isn't everything. Latency also matters and if you have a somehwat "complex" workload beyond just serving a web site, single-threaded does matter. If your workload completes in 1 sec or 500ms is huge for the effect it has on the end user.
I say this because I work on rather complex Apps but with few users. Eg throughput doesn't mattter, your phone could probably handle it. But if the users has the "instant" load feeling vs waiting a bit, that is huge. Don't know exact value anymore but if load time exceeds 2-3 seconds it doesn't matter if its 5 or 10s. It's just slow and crappy. Latency is soooo underrated for good User experience.

 

[Andrei.]

This one is easy: they tried and failed miserably because their design rules are so tied to their process that porting any existing design to their process is a royal pain. Add to that their first serious trial (with LG) was supposed to be on 10nm. Boom.

Not only was their foundry 10nm absolutely non-functional in hilarious ways, they also burned down bridges on 14nm, denying their broken process was at fault for issues for over half a year to customers.

 

[Nothingness]

TSMC has a slightly bigger market cap than Intel right now. Imagine if Intel saw TSMC's business opportunity. Intel would be worth twice as much right now.

As previously written they saw it and even tried. But they are so full of themselves they completely failed at delivering usable foundering service.

And their web page is still up with many funny statements: https://www.intel.com/content/www/us/en/foundry/overview.html

Intel has long been known as the champion of Moore’s Law, the engine of constant innovation in silicon manufacturing that is at the heart of modern technology. Through the Intel Custom Foundry, customers can harness Moore’s Law in their own products. Our foundry is the world’s most advanced and complete contract integrated circuit (IC) manufacturer. Leverage the process development leadership, manufacturing prowess, and packaging resources of the world's largest semiconductor company, Intel, to benefit your products.

Intel enables Moore’s Law through the world’s most advanced processes, wafer fabs, and packaging solutions.

Hilarious.

 

[Topweasel]

Cheap ARM servers using chiplets is an absolute nightmare for both Intel and AMD, because they might be willing to forgo single thread (which is vital to client workloads) to get max MT/mm2/W.

I am not sure its that big of an issue and will affect Intel waaaay more than it does AMD.

Designing and producing a CPU is pretty damn expensive. If you aren't Apple/Samsung/Microsoft/Amazon/Tesla(well for different reasons). Designing an in house CPU and converting your tool chain to is an incredibly long winded, expensive, and dangerous solution. Even when getting others off the market CPU, the whole infrastructure move is expensive enough. There was a really good reason not tied just to the fact Mercer was sooooooo slow that people abandoned Itanium near instantly when AMD created X86-64. As long as you have lets say an AMD x86 solution that's within 20% of the performance and still cheap, they would be much more likely to move to that solution. Just look at AMD they might have been 6 years too early, but they bought a Arm server company and just found the sales market of the device less than good.

Amazon isn't really going to sell them. Not without them basically offering a full tool chain for the companies they are selling to, that match up perfectly and it will always be realitively small companies. What this does mean is as a major ODM buyer (remember they were buying SL-SP almost a year before it was available). They have the funding to get a CPU design, negotiate with TMSC, and develop the tools sets (since they have from the very beginning of AWS) to use these CPU's. It still won't replace everything, but it certainly lowers their reliance on over priced, insecure, and slowly developing Intel CPU's. For AMD this endeavor by Amazon is so expensive that their requirements on production were probably numbers AMD would struggle to keep up with. Which probably played another part in Amazon going this route.

 

[Nothingness]

Designing and producing a CPU is pretty damn expensive. If you aren't Apple/Samsung/Microsoft/Amazon/Tesla(well for different reasons).

You're mixing things. In your list only Apple and Samsung are designing their own CPUs. The others rely on ARM Ltd CPU designs. This doesn't mean it's not expensive of course.

Designing an in house CPU and converting your tool chain to is an incredibly long winded, expensive, and dangerous solution. Even when getting others off the market CPU, the whole infrastructure move is expensive enough. There was a really good reason not tied just to the fact Mercer was sooooooo slow that people abandoned Itanium near instantly when AMD created X86-64. As long as you have lets say an AMD x86 solution that's within 20% of the performance and still cheap, they would be much more likely to move to that solution.

ARM tool support already is very good. What do you think is missing?

Just look at AMD they might have been 6 years too early, but they bought a Arm server company and just found the sales market of the device less than good.

What ARM server company did AMD buy?

Amazon isn't really going to sell them. Not without them basically offering a full tool chain for the companies they are selling to, that match up perfectly and it will always be realitively small companies. What this does mean is as a major ODM buyer (remember they were buying SL-SP almost a year before it was available). They have the funding to get a CPU design, negotiate with TMSC, and develop the tools sets (since they have from the very beginning of AWS) to use these CPU's. It still won't replace everything, but it certainly lowers their reliance on over priced, insecure, and slowly developing Intel CPU's. For AMD this endeavor by Amazon is so expensive that their requirements on production were probably numbers AMD would struggle to keep up with. Which probably played another part in Amazon going this route.

You're again talking about tools. What do you think AWS had to develop as far as SW tools go? I'm wondering what you call tools. For me it's compilers, JVM and such SW for which ARM support is rather good.

 

[jpiniero]

I am not sure its that big of an issue and will affect Intel waaaay more than it does AMD.

There's a lot of the market that will stick with Intel even with them being so noncompetitive. AMD not only has to deal with the inertia of ditching Intel, they have to deal with any ARM competitors like Gravitron.

 

[Topweasel]

You're mixing things. In your list only Apple and Samsung are designing their own CPUs. The others rely on ARM Ltd CPU designs. This doesn't mean it's not expensive of course.

No these guys license core designs from ARM. They still actually have to develop and have the actual CPU built. Samsung and Apple just have a license that allows them to make changes to the core design to make derivative products.

ARM tool support already is very good. What do you think is missing?

General ARM support is good. Server infrastructure tool chains are almost non existent.

What ARM server company did AMD buy?

SeaMicro, it was still a good buy as AMD uses a lot of their Interconnect technology when working on IF.

You're again talking about tools. What do you think AWS had to develop as far as SW tools go? I'm wondering what you call tools. For me it's compilers, JVM and such SW for which ARM support is rather good.

Again this a micro, I see arm used everywhere look. You need to look at what the type of product it is. It's a datacenter server CPU, not a cellphone CPU. This isn't going to be used by software development houses. It's going to in theory be used by people looking for Non-intel/cheaper/less power hungry/and a few things possible better performance fit, server jobs. Those kind of tools don't exist in the wild and for the most part have to be developed by hand. The move over for a company not buying on the scale of Amazon or as largely affected by supply shortages, as Amazon. This move would cost more time and money, and money, than any of the savings.

 

[Dribble]

Now they've got the performance ARM is going to dominate this market unless you need windows. They will be so much cheaper - nothing to stop various Chinese companies pumping out their own variants for peanuts, there's no x86 lockdown here. In addition people like Amazon can customize it, something you can't do to an Intel/AMD x86 cpu - they could put special io, ai, encription, whatever they like on the chip.

 

[CHADBOGA]

As well as what Nothingness said, don't forget that back before mobiles took off, Intel supposedly refused to make a SoC for Apple fearing it was too low margin.
Plus having deliberately crippled Atom almost since inception to segment the market.
These and other factors eventually gave TSMC enough volume to be able to invest massively in their process and fabs.

Sure, but that just goes to show what poor strategic thinkers there are at Intel.

I'll also add that I never understood why they didn't make a far larger effort to make their own discrete GPU's from 20+ years ago.

Yes I know they had a famous failure with the i740, but rather than learn from this, they just tucked their tails between their legs and ran away.

Now they have to start all over again from scratch and in the interim, think of how many billions Nvidia has made over the last 20+ years.

For a company Intel's size, I expect a lot more than they deliver.

 

[CHADBOGA]

This one is easy: they tried and failed miserably because their design rules are so tied to their process that porting any existing design to their process is a royal pain.

Surely it could not have been that difficult to make adjustments.