I’ve been curious for a while: Has anyone done the envelope math to estimate how much of the Graviton price advantage offered by AWS is a reflection of true price/performance vs. AWS taking a short term margin hit to encourage growth of a CPU platform they control?

> by AWS is a reflection of true price/performance

I assume you mean cost / performance as in Cost to amazon and not price as in price paid by consumers? Because Graviton is both faster and cheaper per vCPU. ( Amazon unit for a single thread )

>AWS taking a short term margin hit

It is hard because even if you can reliably extrapolate the cost of Graviton chip by wafer price. You will never know how they put the R&D into accounting. Do you have NRE cost in every single one those chips? ( which is what most people assume but not necessarily true ) Or do you move the whole R&D of Graviton into whole AWS R&D budget? And then you will have to calculate the power usage per workload to work out the power saving which gives you the TCO. Since 30%+ of Operating cost are going to energy usage. Being power efficient is also extremely important at the scale of AWS.

All in all I do not believe that Amazon is takin a short term margin hit ( and I dont understand where this narrative came from ), it is cheaper to produce, it uses less energy, and Amazon is pricing it as such, slightly cheaper than Intel and AMD's counterpart. Just like they did with the first generation of ARM. Graviton actually being faster per core just happens to be a bonus.

One has to remember. Moving to ARM / Graviton instances isn't free. You have to do your own testing and benchmarks to make sure it is indeed faster for your Apps. And that isn't always the case. You also have to make sure every part of your stack compile and works 100% on ARM. It just happens the cost saving is so gigantic lots of companies are spending time to move, which create a positive feedback loop in both open source and proprietary software to move towards ARM.

I'm not sure is it real but they obviously have incentive to sell Graviton instances cheaper than it actually achieved. By doing discount, they can encourage customers to use more Graviton instance (it reduces development cost per chip) and they'll get discount from Intel/AMD.

>they'll get discount from Intel/AMD.

This would only make sense if Intel and AMD are still the dominant platform for Amazon. But in reality nearly 50% of all new EC2 instances built on are Graviton, and they are only limited by TSMC capacity. ( Or more like Amazon inability to predict the Graviton demand ) in another year or two this could close down to 70-80%. Last time I heard this pace is two years ahead of their internal estimate and schedule. And most of these deals were signed years before, i.e Amazon are still enjoying discounts while producing Graviton. So I dont think discount from Intel and AMD factor much into those decisions. If Intel no longer give discount AWS just purchase more from AMD. Amazon has all the cards in these deals. Not Intel or AMD.

It's the Amazon playbook. They charge what they think will be profitable 5 years in the future and pay their way through until they can profitably support that price.

> and I dont understand where this narrative came from

The paranoid might suggest it is coming from Intel's PR team, who might want AWS customers to believe that the savings might not be sustainable.

Intel charges a much larger markup per chip than AWS can get from their own silicon + licensing ARM IP; the power consumption of Intel chips is also terrible (it's been cited that it's 60% lower power consumption for c7g instances compared to c6i instances, for instance, in the latest reinvent talk...) -- and power is what predominates operating cost day to day once the chip is fabbed/installed.

There's no way to know what price AWS pays for any processor. Having said that, it looks like ARM server processors do have an area efficiency advantage over Intel/AMD.

Currently the cheapest spot instances with >30gb are

a1/32gb/16vcpu (early graviton) @ 8c and

M5zn/48gb/12vcpu (late Intel) @ 12c

I think the intel is better value. What are those graviton instances with better value (given that they are probably much slower than Intel)?

> I think the intel is better value. What are those graviton instances with better value (given that they are probably much slower than Intel)?

Intel instances are significantly more expensive when considering the cost for performance. In fact it's not even close [0].

[0] https://www.anandtech.com/show/15578/cloud-clash-amazon-grav...

The parent seemed to be specifically calling out Intel spot instances.

The key part being the cheapest available Graviton and Intel spot instances:

a1/32gb/16vcpu (early graviton) @ 8c and

versus:

M5zn/48gb/12vcpu (late Intel) @ 12c

Perhaps that will change as Graviton ages and availability changes.

> https://aws.amazon.com/about-aws/whats-new/2021/11/amazon-ec...

Fair enough - I was trying to provide some data showing that Graviton is on par and exceeds cost/performance value in some cases.

> Perhaps that will change as Graviton ages and availability changes.

Graviton isn't standing still - AWS bought Annapura labs in 2015, so while older versions of Graviton will obviously shift left in terms of it's value, AWS is going to continue to push in that direction. Graviton 3 was recently announced [0].

[0] https://aws.amazon.com/about-aws/whats-new/2021/11/amazon-ec...

Mmh. Based on that Anandtech article, I guess the graviton2 cpus are actually similar in performance to intel ones. But still, my spot instance searches (especially also taking into account interruption frequencies), prices seem a bit of a wash. It doesn't look like Graviton is 40% cheaper. This may be different for on-demand or reserved pricing.

I wrote a framework that does regexes of the Common Crawl. The Common Crawl is a series of captures of billions of pages of the WWW. Each member in the series has ~60 TB captures.

My framework runs a fairly simple pipeline: for each 1GB file, stream it from S3, decompress it, search it with a regex library, stream results back to S3.

A1 spot instances have consistently been cheaper for this task when measured on a $/GB scanned basis. Part of this might be that the framework doesn't need lots of RAM, so paying for more than 2GB per core on the non-A1 instances is just wasted expense for me.

Wouldn’t a lambda be far cheaper?

In my experience, no. Quite the opposite.

Lambdas can be cost effective for bursty loads, but if you can push sustained CPU utilization, as this example does, EC2 is cheaper.

For example, an ARM-based Lambda with 2 GB of RAM running for 1 hour is $0.09. An a1.medium is $0.025/hr on-demand or $0.0084/hr spot.

Part of this is that Lambda scales your CPU performance with your memory usage. So even if your code can run in 128 MB of RAM, there's no cost benefit to you if it's CPU-constrained, because it will run 16x slower than if you allocated it 2048 MB of RAM.

It seems unclear, the Intel has fewer cores (6 vs A1's 8) and probably less memory bandwidth per core, but then each core would be faster than the Graviton1. You'd need to benchmark your specific workload to find out.

The a1.4xlarge has 16 cores, not 8. One of the sales pitches of the Graviton is that it has no SMT, so you get more predictable per-core performance.

Oops, yes, it's 16 cores vs 6.

Custom instructions / features that are only available on AWS silicon will improve their vendor lock-in. And thus would warrant short term price dumping. Future cost/performance improvements will allow them to raise their margin. Kind of worried what that will mean for the portability of software long-term.

> Custom instructions / features that are only available on AWS silicon

If there are custom instructions added, AWS customers aren't aware of them. Feature set pretty much unchanged too. It's a non-effect on that front.

Are you 'locked in' to intel because of AVX? The cost of moving is recompiling your binary.

What I had in mind is CUDA vs. openCL

Also the field is not static. Amazon does a lot of lobbying. Let's imagine you can check a certain legal compliance box simply by running on graviton.

Or the next industry trend which might be funded and hyped by AWS requires a specific feature for isolation that only really works on graviton.

The 6i instances are beating all amd and generally-available graviton instances in raw performance and performance-per-dollar on my workloads. I don’t think you’ll be well-served carrying this mental bias around. Better to actually try it and find out.

If you’re going to compare the paper performance of an unreleased machine to a 3-year-old Intel machine then yeah probably the former will win.

> I don’t think you’ll be well-served carrying this mental bias around.

I didn't mean to pick sides, or suggest anyone picks graviton by default. I meant to comment on how the communication around these new architectures by AWS has shifted, and what - I think - it signals about Intel's current and future market position with AWS.

For reference, my stuff runs on Heroku, I'm clearly not that concerned about price/performance ;)

> The 6i instances are beating all amd and generally-available graviton instances in raw performance and performance-per-dollar on my workloads.

I can confirm that:

Build time of a Linux 5.13.11 kernel with `make defconfig all`

c6i.32xlarge: 24.871411s

m6a.48xlarge: 27.622526s

Build time of a Linux 5.13.11 kernel with `make allmodconfig all`

c6i.32xlarge: 183.775255s

m6a.48xlarge: 205.326048s

128-way Intel seems to consistently beat 196-way AMD for software builds.

(Correction: 192-way AMD.)

> only take intel if you need something like SAP

Not even that any more.

From the article: "A key highlight is that M6a instances are SAP certified"

That really surprising. I switch our workloads from m5 to m5a and there was a noticeable drop in performance and increase in latency.

I'm running in-house go services that consume from kinesis streams, so nothing strange like SAP.