It always strikes me that the best place of information for a cloud provider is not from that provider but a third party website. This is not a good comment for the cloud provider.
While the 5 variant isn't yet available outside of the preview, you can of course spin a 4 up and run geekbench yourself. Plenty of people have and you can find them in the GB DB. And of course most people spin up their specific workload to see how it compares.
Core per core it pales compared to Apple's superlative processors, and falls behind AMD as well.
But...that doesn't matter. You buy cloud resources generally for $/perf, and the Graviton's are far and away ahead on that metric.
In Amazon's Graviton 5 PR they note that over half of all new compute capacity added to AWS over the past three years has been Graviton-based. That's an amazing stat.
It really is incredible how ARM basically commoditized processors (in a good way).
Inversely, I think it's siloed things in somewhat unhealthy ways. We now have a number of vendors that sell/rent you machines that are not generally purchasable. I don't think we've seen too many negative consequences yet, but if things continue in this direction then choosing a cloud provider for a high performance application (eg, something you'll want to compile to machine code and is therefore architecture specific in some way as opposed to a python flask app or something), one may have to make decisions that lock one into a particular cloud vendor. Or at least, it will further increase the cost of changing vendors if you have to significantly tweak your application for some oddities between diff arm implementations at different hosting providers, etc.
I would much rather see some kind of mandatory open market sale of all cpu lines so that in theory you can run graviton procs in rackspace, apple m5 servers in azure, etc.
Graviton CPUs are just Neoverse cores (V3 in this case). While it’s true that you can’t just buy a box with the same cores, the cores are basically the same as what you’ll get on a Google or Azure cloud instance (eventually… the latter two have yet to make available anything with Neoverse V3 yet).
More specifically, the CPU cores in AWS Graviton5 are Neoverse V3 cores, which implement the Armv9.2-A ISA specification.
Neoverse V3 is the server version of the Cortex-X4 core which has been used in a large number of smartphones.
The Neoverse V3 and Cortex-X4 cores are very similar in size and performance with the Intel E-cores Skymont and Darkmont (the E-cores of Arrow Lake and of the future Panther Lake).
Intel will launch next year a server CPU with Darkmont cores (Clearwater Forest), which will have cores similar to this AWS Graviton5, but for now Intel only has the Sierra Forest server CPUs with E-cores (belonging to the Xeon 6 series), which use much weaker CPU cores than those of the new Graviton5 (i.e. cores equivalent with the Crestmont E-cores of the old Meteor Lake).
AMD Zen 5 CPUs are significantly better for computationally-intensive workloads, but for general-purpose applications without great computational demands the cores of Graviton5 and also Intel Skymont/Darkmont have greater performance per die area and power consumption, therefore lower cost.
>The Neoverse V3 and Cortex-X4 cores are very similar in size and performance with the Intel E-cores Skymont and Darkmont (the E-cores of Arrow Lake and of the future Panther Lake).
That is not entirely accurate. X4 is big core design. All of its predecessor and successor has always had >1mm2 die space design. X4 is already on the smaller scale, it was the last ARM design before they went all in chasing Apple's A Series IPC. IRRC it was about 1.5mm2 depending on L2 cache. E-Core for Intel has always been below 1mm2. And again IRRC that die size has always been Intel's design guidelines and limits for E-Core design.
More recent X5 / X925 and X6 / X930 / C1 Ultra?? ( I can no longer remember those names ) are double the size of X4. With X930 / C1 Ultra very close to A19 Pro Performance. Within ~5%.
I assume they stick with X4 is simply because it offers best Performance / Die Space, but it is still a 2-3 years old design. On the other hand I am eagerly waiting for Zen 6c with 256 Core. I cant wait to see the Oxide team using Zen 6c, forget about the cloud. 90%+ of companies could fit their IT resources in a few racks.
Nope. Cortex-X4 is not a big core design, though you are right that at the time of its launch in 2023 the Arm company was not offering bigger cores yet.
The cores designed now by the Arm company for non-embedded applications are distributed into 4 sizes, of which the smaller 2 sizes correspond to what were the original "big and little sizes", but what was originally the big size has been continued into what are now medium-to-small cores, and the last such core before the rebranding was Cortex-A725.
Cortex-X4 is of the second size, medium-to-large. Cortex-X925 was the last big core design before Arm changed the branding this year, so several recent smartphones use Cortex-X925 as the big core, Cortex-X4 as the medium-sized core and Cortex-A725 as the small cores, omitting the smallest Cortex-A520 cores.
Cortex-X4 and Intel Skymont have exactly the same size, 1.7 square millimeter with 1 MB L2 cache memory (in Dimensity 9400 and Lunar Lake).
Moreover, Neoverse V3/Cortex-X4 and Intel Skymont/Darkmont have approximately the same number of execution units of each kind in their backends. Only their frontends are very different, which is caused by the different ISAs that must be decoded, Aarch64 vs. x86-64.
The last Arm big core before rebranding, Cortex-X925, was totally unsuitable as a server core, as it had very poor performance per area, having a double area in comparison with Cortex-X4, but a performance greater by only a few tens percent at most. Therefore the performance per socket of a server CPU would have been much lower than that of a Graviton5, had it been implemented with Cortex-X925, due to the much lower number of cores per socket that could have been achieved.
Cortex-X4 was launched in 2023 and it was the big core of the 2024 flagship smartphones, then it has become the medium core of the 2025 flagship smartphones. Its server variant, Neoverse V3, has been launched in 2024 and it has been deployed in products only this year, first by NVIDIA (in Orin) and now by AWS.
It is not at all an obsolete core. As I have said, Intel will have only next year a server CPU with E-cores as good as Cortex-X4. We do not know yet any real numbers about the newly announced Arm cores that have replaced Cortex-A520, Cortex-A725, Cortex-X4 and Cortex-X925, so we do not know if they are really significantly better. The numbers used by Arm in presentations cannot be verified independently and usually when the performance is measured much later in actual products it does not match the optimistic predictions.
The new generation of cores might be measurably better only for computational applications, because they now include matrix execution units, but their first implementation may be not optimal yet, as it happened in the past with the first implementation of SVE, when the new cores had worse energy efficiency than the previous generation (which was corrected by improved implementations later).
Well, also no licensing costs to AMD/intel. So even if at slightly worse performance per chip it’ll end up being cheaper still. AWS doesn’t need to make money on their chips, as they already have the Ec2 margin.
Good question! I read two different Amazon press releases on this but still had to come here for the answer. It seems strange they don't want to advertise the ISA of a compute product - does marketing think it might scare people away?
It seems they don't document the ISA for any instance types. This could be deliberate (and unrelated to marketing) in case they decide to pull features from the instance types in a microcode update. Without any ISA specifics, previous customer commitments towards instance types would still apply.
Awhile back I was researching cloud instances for performance, And I noticed that AWS didn't have the latest generations of AMD/Intel. Which are far superior to Graviton 4.
It seems obvious to me that AWS using their market dominance to shift workloads to Graviton.
This sort of makes sense. If there is no competitive advantage in buying the latest AMD or Intel CPUs, why buy them when you can just deploy a generic (ARM licensed) CPU at cheaper prices.
The competitive advantage right now is in NVIDIA chips and I guess AWS needs all their free cash to buy those instead of non-competitive advantage CPUs.
I imagine it can take time to actually validate and build out that new infrastructure at scale after AMD/Intel announces these products to the market. It wouldn't surprise me if hyperscalers like AWS, Google, Microsoft, et. al. get a little bit of early previews of this hardware, but it still takes time to negotiate sales, buy the chips, and then actually receive the new chips and make actually useful systems.
Meanwhile, when AWS announces a new chip its probably something they have already been building out in their datacenters.
General purpose not AI specific? I can't believe it.
Pricing when? :(
https://aws.amazon.com/ec2/pricing/on-demand/
If only dedicated game servers could run on aarch64...
I've been experimenting FEX on Ampere A1 with x86 game servers but the performance is not that impressed
Doesn't help that Unity requires forking over a pile of cash just to build for Linux ARM ("Embedded Linux") and everything else is free.
discussed a couple days ago: https://news.ycombinator.com/item?id=46191993
AWS introduces Graviton5–the company's most powerful and efficient CPU (14 comments)
Is there a list of Geekbench performance metrics for the various Graviton CPUs?
I need a reference point so I can compare it to Intel/AMD and Apple's ARM cpus.
Otherwise it is buzzwords and superlatives. I need numbers so I can understand.
https://instances.vantage.sh/ shows coremark scores for each EC2 instance type.
It always strikes me that the best place of information for a cloud provider is not from that provider but a third party website. This is not a good comment for the cloud provider.
Funny story: When I was at AWS, I found that the easiest way automate instance data collection was by using the Vantage website code (it's on GitHub).
The cobbler's children have no shoes.
While the 5 variant isn't yet available outside of the preview, you can of course spin a 4 up and run geekbench yourself. Plenty of people have and you can find them in the GB DB. And of course most people spin up their specific workload to see how it compares.
Core per core it pales compared to Apple's superlative processors, and falls behind AMD as well.
But...that doesn't matter. You buy cloud resources generally for $/perf, and the Graviton's are far and away ahead on that metric.
In Amazon's Graviton 5 PR they note that over half of all new compute capacity added to AWS over the past three years has been Graviton-based. That's an amazing stat.
It really is incredible how ARM basically commoditized processors (in a good way).
Inversely, I think it's siloed things in somewhat unhealthy ways. We now have a number of vendors that sell/rent you machines that are not generally purchasable. I don't think we've seen too many negative consequences yet, but if things continue in this direction then choosing a cloud provider for a high performance application (eg, something you'll want to compile to machine code and is therefore architecture specific in some way as opposed to a python flask app or something), one may have to make decisions that lock one into a particular cloud vendor. Or at least, it will further increase the cost of changing vendors if you have to significantly tweak your application for some oddities between diff arm implementations at different hosting providers, etc.
I would much rather see some kind of mandatory open market sale of all cpu lines so that in theory you can run graviton procs in rackspace, apple m5 servers in azure, etc.
Graviton CPUs are just Neoverse cores (V3 in this case). While it’s true that you can’t just buy a box with the same cores, the cores are basically the same as what you’ll get on a Google or Azure cloud instance (eventually… the latter two have yet to make available anything with Neoverse V3 yet).
Didn't M8g just come out? Am I crazy?
Not crazy. They just have a pretty rapid release cadence for Graviton. New chips ~ every two years.
So these are aarch64, right?
More specifically, the CPU cores in AWS Graviton5 are Neoverse V3 cores, which implement the Armv9.2-A ISA specification.
Neoverse V3 is the server version of the Cortex-X4 core which has been used in a large number of smartphones.
The Neoverse V3 and Cortex-X4 cores are very similar in size and performance with the Intel E-cores Skymont and Darkmont (the E-cores of Arrow Lake and of the future Panther Lake).
Intel will launch next year a server CPU with Darkmont cores (Clearwater Forest), which will have cores similar to this AWS Graviton5, but for now Intel only has the Sierra Forest server CPUs with E-cores (belonging to the Xeon 6 series), which use much weaker CPU cores than those of the new Graviton5 (i.e. cores equivalent with the Crestmont E-cores of the old Meteor Lake).
AMD Zen 5 CPUs are significantly better for computationally-intensive workloads, but for general-purpose applications without great computational demands the cores of Graviton5 and also Intel Skymont/Darkmont have greater performance per die area and power consumption, therefore lower cost.
>The Neoverse V3 and Cortex-X4 cores are very similar in size and performance with the Intel E-cores Skymont and Darkmont (the E-cores of Arrow Lake and of the future Panther Lake).
That is not entirely accurate. X4 is big core design. All of its predecessor and successor has always had >1mm2 die space design. X4 is already on the smaller scale, it was the last ARM design before they went all in chasing Apple's A Series IPC. IRRC it was about 1.5mm2 depending on L2 cache. E-Core for Intel has always been below 1mm2. And again IRRC that die size has always been Intel's design guidelines and limits for E-Core design.
More recent X5 / X925 and X6 / X930 / C1 Ultra?? ( I can no longer remember those names ) are double the size of X4. With X930 / C1 Ultra very close to A19 Pro Performance. Within ~5%.
I assume they stick with X4 is simply because it offers best Performance / Die Space, but it is still a 2-3 years old design. On the other hand I am eagerly waiting for Zen 6c with 256 Core. I cant wait to see the Oxide team using Zen 6c, forget about the cloud. 90%+ of companies could fit their IT resources in a few racks.
Nope. Cortex-X4 is not a big core design, though you are right that at the time of its launch in 2023 the Arm company was not offering bigger cores yet.
The cores designed now by the Arm company for non-embedded applications are distributed into 4 sizes, of which the smaller 2 sizes correspond to what were the original "big and little sizes", but what was originally the big size has been continued into what are now medium-to-small cores, and the last such core before the rebranding was Cortex-A725.
Cortex-X4 is of the second size, medium-to-large. Cortex-X925 was the last big core design before Arm changed the branding this year, so several recent smartphones use Cortex-X925 as the big core, Cortex-X4 as the medium-sized core and Cortex-A725 as the small cores, omitting the smallest Cortex-A520 cores.
Cortex-X4 and Intel Skymont have exactly the same size, 1.7 square millimeter with 1 MB L2 cache memory (in Dimensity 9400 and Lunar Lake).
Moreover, Neoverse V3/Cortex-X4 and Intel Skymont/Darkmont have approximately the same number of execution units of each kind in their backends. Only their frontends are very different, which is caused by the different ISAs that must be decoded, Aarch64 vs. x86-64.
The last Arm big core before rebranding, Cortex-X925, was totally unsuitable as a server core, as it had very poor performance per area, having a double area in comparison with Cortex-X4, but a performance greater by only a few tens percent at most. Therefore the performance per socket of a server CPU would have been much lower than that of a Graviton5, had it been implemented with Cortex-X925, due to the much lower number of cores per socket that could have been achieved.
Cortex-X4 was launched in 2023 and it was the big core of the 2024 flagship smartphones, then it has become the medium core of the 2025 flagship smartphones. Its server variant, Neoverse V3, has been launched in 2024 and it has been deployed in products only this year, first by NVIDIA (in Orin) and now by AWS.
It is not at all an obsolete core. As I have said, Intel will have only next year a server CPU with E-cores as good as Cortex-X4. We do not know yet any real numbers about the newly announced Arm cores that have replaced Cortex-A520, Cortex-A725, Cortex-X4 and Cortex-X925, so we do not know if they are really significantly better. The numbers used by Arm in presentations cannot be verified independently and usually when the performance is measured much later in actual products it does not match the optimistic predictions.
The new generation of cores might be measurably better only for computational applications, because they now include matrix execution units, but their first implementation may be not optimal yet, as it happened in the past with the first implementation of SVE, when the new cores had worse energy efficiency than the previous generation (which was corrected by improved implementations later).
Well, also no licensing costs to AMD/intel. So even if at slightly worse performance per chip it’ll end up being cheaper still. AWS doesn’t need to make money on their chips, as they already have the Ec2 margin.
Do you have any insight on when these will be generally available?
Amazon says "Sign up for the preview today".
I have no connection with them, so I have no idea when these instances will be generally available.
Privileged big customers appear to be already testing them.
Good question! I read two different Amazon press releases on this but still had to come here for the answer. It seems strange they don't want to advertise the ISA of a compute product - does marketing think it might scare people away?
It seems they don't document the ISA for any instance types. This could be deliberate (and unrelated to marketing) in case they decide to pull features from the instance types in a microcode update. Without any ISA specifics, previous customer commitments towards instance types would still apply.
They list what specific cpus you get for each instance type, see eg https://docs.aws.amazon.com/ec2/latest/instancetypes/gp.html
At this point I think they just assume that everyone who cares already know that graviton=arm
Yes, Graviton chips are aarch64.
Awhile back I was researching cloud instances for performance, And I noticed that AWS didn't have the latest generations of AMD/Intel. Which are far superior to Graviton 4.
It seems obvious to me that AWS using their market dominance to shift workloads to Graviton.
This sort of makes sense. If there is no competitive advantage in buying the latest AMD or Intel CPUs, why buy them when you can just deploy a generic (ARM licensed) CPU at cheaper prices.
The competitive advantage right now is in NVIDIA chips and I guess AWS needs all their free cash to buy those instead of non-competitive advantage CPUs.
I think Graviton would still be much more energy efficient though? (I'm not sure)
I believe the main motivator for AWS is efficiency, not performance. $ of income per watt spend is much better for them on Graviton.
At what point was that true? For example right now ec2 has granite rapids cpus available which are very much the latest and greatest from intel.
I imagine it can take time to actually validate and build out that new infrastructure at scale after AMD/Intel announces these products to the market. It wouldn't surprise me if hyperscalers like AWS, Google, Microsoft, et. al. get a little bit of early previews of this hardware, but it still takes time to negotiate sales, buy the chips, and then actually receive the new chips and make actually useful systems.
Meanwhile, when AWS announces a new chip its probably something they have already been building out in their datacenters.
>Which are far superior to Graviton 4.
Not if you are looking at price/performance. AWS could be taking a loss to elevate the product though, no way to know for sure.
If they were taking a loss, they wouldn't run a crapton of internal workloads on Graviton.