8 ms·
Cloud VM benchmarks 2026
- nixgeek 7mo agoGenoa was a big leap from Milan. Turin is a huge leap again. AMD really is doing spectacularly well at the moment. Kudos to Lisa Su and the team.
- jiggawatts 7mo ago> Kudos to Lisa Su and the team. They're a typical hardware maker unable to focus on software, which is why NVIDIA is now a multi-trillion dollar corporation and AMD is "just" a few hundred billion. They've focused too much on CPUs and completely dropped the ball on AI and compute accelerators. It's especially sad considering that the MI300 and related accelerators on paper are competitive with NVIDIA hardware, it's just that they have nowhere near the same software stack, so nobody cares.
- dijit 7mo agoDon’t really care. We were stuck with Intel, its nice that we have better CPUs.
- a012 7mo agoYeah, remember when 4 core 8 threads were the high-end CPUs until AMD Ryzen came out? If AMD didn't do their best job then we're still stuck with the norm of 4 cores for more years as I can imagine.
- close04 7mo agoAMD has to fight both Intel and Nvidia in the market. It chose to take on Intel and clearly it was a wise decision. You can’t win if you fight every battle at once against much stronger opponents. And don’t get me started on the valuation of companies riding the AI bubble.
- theandrewbailey 7mo ago> completely dropped the ball on AI and compute accelerators AMD produces AI chips, and they seem to be doing quite well.[0] If they didn't, AMD wouldn't be worth anywhere near what it is. [0] https://openai.com/index/openai-amd-strategic-partnership/ https://openai.com/index/openai-amd-strategic-partnership/
- LKummer 7mo agoNvidia datacenter GPUs have awful software. If they focus on it they're not doing a good job.
- rr808 7mo agoStill having your own hardware seems so much cheaper. Maybe even just for dev/uat environments? Every big corporate I have worked at has lower cost of capital than Amazon, and yet they want to move to AWS. I just dont understand it.
- doener 7mo agoThere are so many articles like these: "We Moved from AWS to Hetzner. Cut Costs 89%. Here’s the Catch." https://medium.com/lets-code-future/we-moved-from-aws-to-hetzner-cut-costs-89-heres-the-catch-961ea2b9237c https://medium.com/lets-code-future/we-moved-from-aws-to-het...
- selectively 7mo agoYou just linked AI slop.
- doener 7mo agoYou mean the image? The text does not sound like AI at all IMHO.
- dkechag 7mo ago> No theory. No fluff. Just production. ChatGPT tells me "no theory, no fluff" all the time :D
- BeastMachine 7mo agoWhere do you think it learned that phrasing from?
- deaux 7mo agoNot from people using that same phrasing twice within a few sentences. """ No warning. No traffic spike. Just… more money gone. That’s when I finally looked at Hetzner. I’ve seen too many backend systems fail for the same reasons — and too many teams learn the hard way. So I turned those incidents into a practical field manual: real failures, root causes, fixes, and prevention systems. No theory. No fluff. Just production. """ It's clearly slop, they immediately use effectively the same one again: """ That last line isn’t a joke. There were charges I genuinely couldn’t explain. Elastic IPs we forgot to release. Snapshots from instances that no longer existed. CloudFront distributions someone set up for testing. """ No, human writers don't repeat this pattern every single paragraph. They use it at most across in a whole article.
- doener 7mo ago"Unfortunately, the [Hetzner] CPX22is available only in eu-central and ap-southeast, but if that’s OK with you it is the best value and fastest overall."
- himata4113 7mo agoI am still running ROME epyc cpus that I picked up for couple hundred and they're doing great. Power usage is not the best and singlethread is awful (-50%), but multithreaded performance kicks 9950x in the ass at around 90k vs 70k.
- boberoni 7mo agoAnyone have experience with Oracle Cloud and ease of moving away? This benchmark seems to recommend Oracle Cloud, but I’ve heard that Oracle has historically used aggressive licenses and legal terms to keep customers locked-in.
- acedTrex 7mo agoEverytime i have to login to oracle cloud i strongly wish i was not logging into oracle cloud...
- dkechag 7mo agoI wrote the article. I would NEVER tell anyone to use Oracle, as the vendor lock in and strong-arming and pricing is ridiculous. That said, I am hosting small projects on Oracle Cloud, due to the super low cost. I can just move them whenever they decide to be naughty, I am not using an oracle DB or anything proprietary, just linux VMs with my own mysql setup.
- olalonde 7mo agoI was confused by your comment. I assume you meant to say "I would NEVER tell anyone to use Oracle DB" (as opposed to Oracle Cloud)?
- esseph 7mo agoI would never let a friend use Oracle anything. Oracle is a predator.
- speed_spread 7mo agoBut is a lawnmower a predator to the grass?
- dkechag 7mo agoI meant pretty much anything Oracle... except their free/cheap VMs :)
- lostmsu 7mo agoGCP (near the top) 3 years reserved 16 ARM cores + 120 GB costs + 1TB local SSD costs > $1k/month. It does not even match the specs of a Ryzen AI 395 Framework miniPC that goes for ~$4k AFAIK.
- _pdp_ 7mo agoYou can extract a lot of value from bare-metal servers from Hetzner but you need to put some effort initially to get them going. That being said, it is not really that difficult. And, frankly, it a lot more fun.
- navigate8310 7mo agoCould you explain the "some effort initially" part?
- esafak 7mo agoThe usual bare metal lack of amenities, that's all.
- jurgenburgen 7mo agoYou take care of disaster recovery, failing machines, the works. Getting a new (or replacement) machine from Hetzner takes longer than spinning one up on a cloud provider so you need to be able to survive some machines failing.
- speedgoose 7mo agoAlso no encryption at rest. You have to setup LUKS yourself.
- Thaxll 7mo agoYou can't compare VPS with VMs from major cloud provider, VPS don't offer anything beside basic compute. Also virtualization from cloud provider is way better because they have custom hardware and software so you don't suffer from noisy neighbours for example.
- kev009 7mo agoIt's ironic to see Oracle as a value play, but you couldn't bank on that indefinitely.
- esseph 7mo agoThey're desperate for marketshare
- zackify 7mo agoI just ran some massive tests on our own CI. I use AMD Turin for this on gcp, which was noted as one of the fastest ones in the article. The most insane part here is that the AMD EPYC 4565p can beat the turin's used on the cloud providers, by as much as 2x in the single core. Our tests took 2 minutes on GCP, 1 minute flat on the 4565p with its boost to 5.1ghz holding steady vs only 4.1ghz on the gcp ones. GCP charges $130 a month for 8vcpus. ALSO this is for SPOT that can be killed at any moment. My 4565p is a $500 cpu... 32 vcpus... racked in a datacenter. The machine cost under 2k. i am trying hard to convince more people to rack themselves especially for CI actions. The cloud provider charging $130 / mo for 3x less vcpus you break even in a couple months, it doesn't matter if it dies a few months later. On top of that you're getting full dedicated and 2x the perf. Anyways... glad to see I chose the right cpu type for gcloud even though nothing comes close to the cost / perf of self racking
- Aurornis 7mo ago> My 4565p is a $500 cpu... 32 vcpus... racked in a datacenter. The machine cost under 2k. > The cloud provider charging $140 / mo for 3x less vcpus you break even in a couple months, it doesn't matter if it dies a few months later How do you calculate break even in a couple months if the machine costs $2,000 and you still have to pay colo fees? If your colo fees were $100 month you wouldn’t break even for over 4 years. You could try to find cheaper colocation but even with free colocation your example doesn’t break even for over a year.
- zackify 7mo agothe 140/mo is for 3x less vcpu, so $420/mo savings if you use all those same cores. sorry for the poor comparison wording there. in a few months already up to $1300+ by 6 months already paid the machine. colo fees are cheap if you need more than just 1u. even with a 50-100 fee you easily get way more performance and come ahead within a year
- Aurornis 7mo ago> by 6 months already paid the machine. You originally said “a couple months” but now it’s 6 months and assumption of $0 collocation fees which isn’t realistic In my experience situations rarely call for precisely 32 cores for a fixed period of 3 years to support calculations like this anyway. We start with a small set of cloud servers and scale them up as traffic grows. Today’s tooling makes it easy to auto scale throughout the day, even. When trying to rack a server everyone aims higher because it sucks to start running into limits unexpectedly and be stuck on a server that wasn’t big enough to handle the load. Then you have to start considering having at least two servers in case one starts failing. Racking a single self-built server is great for hobby projects but it’s always more complicated for serving real business workloads.
- OsrsNeedsf2P 7mo agotl;dr Hetzner has best performance for the price. But also Hetzner just bumped prices like 30%
- metadat 7mo agoWhat about performance per dollar? Did I miss this part.
- dkechag 7mo agoThere's a table of contents near the top. Use it to jump to the performance per dollar.
- PaulKeeble 7mo agoA few comparisons just from a gaming PC with a 9800X3D (8C16T 5.2 Ghz Boost). 7 Zip benchmark 9800X3D 130 GIPs compression, 134 GIPs decompress. C8A 21577 MIPs (21.5GIPs) compression, 9868 MIPS decompression (9.9GIPs). Geekbench 5 9800X3D 16975 multithread, 2474 single thread C8A 4049 multithread, 2240 single thread A desktop class CPU is definitely quicker single threaded and multithreaded, no surprises there most of these are dual core. The single threaded performance of the C8A is actually pretty good but its also the best of the bunch by a wide margin most of the CPUs are far behind. Memory performance appears to be attrocious all around.
- dijit 7mo agoThanks for the benchmarks! I’d only add that its very common for gaming computers to be screaming fast- moreso than workstation machines or servers which are a bit more conservative with performance and emphasise correctness (slower cores, slower ram, more ECC). Its not a lot, but it can feel annoying when you sit on a company issued workstation that cost €10,000 but get worse performance than a €2,500 gaming computer.
- jeffbee 7mo agoThe only conclusion you can draw from this is nobody wants to use Oracle and they are trying to buy customers.
- alex_dev42 7mo ago[dead]
- paulsutter 7mo ago> I've seen too many "we'll just rack a few servers" projects turn into full-time infrastructure jobs Really? How many?
- supriyo-biswas 7mo agoI’d really appreciate if you would avoid posting LLM generated comments here. Thanks.
- preserves 7mo agoDisclosure: I work on VMs at Google Compute Engine :) This was a really, really good write-up. I appreciated the breadth of VMs tested and the spread of benchmarks. A few random observations: 1. Turin is a beast. 2. The data on price-performance makes Hetzner look really fantastic, especially for small scale projects where region placement doesn’t matter much and big bursty scaling isn’t required. 3. I think the first ever cloud VM I ever provisioned was on DigitalOcean. I was surprised at how old their fleet was, but I guess they have some limited Emerald Rapids offerings now: https://www.digitalocean.com/blog/introducing-5th-gen-xeon-processors https://www.digitalocean.com/blog/introducing-5th-gen-xeon-p...
- AtlasBarfed 7mo agoNeeds: network performance and costs, desperately. I wouldn't mind SAN/non-local storage performance and costs too. VM cost isn't where AWS gets you. It's ALLLLLLLL the other nickel and diming that they kill you on, especially since outbound data transfer costs a log of money to get off of the platform.
- nodesocket 7mo agoAwesome write up. I use EC2 t4g instances extensively as I tend to scale workloads horizontal (using Kubernetes, Docker) and admittedly the workloads aren’t cpu constrained. Would be interesting to see how t4g’s compare to other low end comparable VMs on GCP, DigitalOcean, Hetzner. Though I get the difficulty testing as it requires maxing out CPU credits (waiting) for accurate results before starting testing for each VM. Looking forward to t5g’s whenever (if ever) they release.
- Imustaskforhelp 7mo agoThere are definitely ways to get some really good hardware from lowend shops as well. Hetzner is pretty decent too actually but I actually think that OVH might be even cheaper while still being competitive, but in the case of OVH, I think one minor issue people can have is the setup fee at times but time to time like during black fridays or special deals, there are ways to get 0 setup fees. If someone wants something stable, I think that OVH is pretty great as well and is comparable to Hetzner in terms of pure price but I have heard that given their scale, their support can be 50/50 but I recommend joining their discord and maybe even using (twitter oof) to message them as thsi was something which worked for someone on hackernews the last time something like this was discussed. If you are okay with some more steal factor, use netcup. You should also probably look at gaming type servers too. I know a person who is a one man shop who was more passionate about these stuff and did have high-end hardware (200k$) in investment in his provider. Going more into the specifics of finance from provider side, usually the idea is that they recoup the costs in 5 years if running at decent capacity/having sold quite a bit but the first few years so as in example of the 200k$, they currently make IIRC 40k-60k$ per year, you have to somehow find your way to customers but that is being messed up because of AI ram prices as their costs to fix any broken hardware has absolutely skyrocketed eating much of their profits. It's an extremely competitive market at times so if you are looking for something specific. You can actually ask about it in forums like lowendtalk, lowendspirit and these people/providers can respond matching what you are looking for price/performance in for and they can also provide test servers if need be and there is a unified form of benchmarking within this space called yabs (yet-another-benchmark-script) which you can ask for a provider. But in all regards, if someone doesn't want to go through this hassle, I will just say that the same provider that I mentioned earlier had also in public mentioned how hetzner prices are pretty competitive and in all honesty I agree with that too. If someone doesn't want to go through some/any part of this hassle, then hetzner/OVH are my go-to safe options. They are big enough that their downtime shouldn't be blamed on you for picking them while being really good in themselves. (Something which I have actually heard quite often mentioned on hackernews) And within this space, if you don't absolutely know what you want, there is no generic winner as there are niches that are occupied. So if you ever want to find any alternative to hetzner for even more cheaper, its best that you first decide what are the things exactly that you want and then ask rather than see for pre-existing if possible. Also if possible, make deals during blackfriday/cyber-monday. Not sure about Hetzner but in OVH/other providers, you can get recurring deals and the sever setup costs etc. removed sometimes.
- puskavi 7mo agoim glad to see that the 36€ im paying for 250/250 fiber is still competitive in grand scheme of internet
- deaux 7mo agoVultr and HostHatch are also worth considering.
- npn 7mo agoDon't use Vultr. I still have 60 dollars credit there that I can never spend, because they will limit your account if you do not use it for sometime. And the pricing is laughable expensive comparr to OVH.
- esseph 7mo agoVultr is great, have been using them for many years.
- andai 7mo agoVultr topped the benchies a few years ago. Not sure how they fare now, would be good to see some more recent benchmarks. Was sad to see them missing from this one.
- irenetusuq 7mo ago[dead]
- lobofta 7mo agoWould love a GPU benchmark too especially for training and inference workloads
- s_dev 7mo agoWould have been nice to see Scaleway included, I recently migrated from Digital Ocean to them and found them to be very similar in pricing and performance.
- secondcoming 7mo agoWhy is the n4d-2 spot price nearly double that of the n2d-2?
- grzes 7mo agothe only thing that keeps me from migrating from DO to Hetzner is the lack of monitoring agent on the latter. its really nice to set up some hdd & cpu alerts from within cloud provider control panel directly.
- pothamk 7mo ago[flagged]
- e-master 7mo ago> The Azure pricing is at least as complex as AWS/GCP, plus the pricing tool seems worse. They also lag behind the other two major providers in CPU releases - Turin and Granite Rapids are still in closed preview at the time of writing this. Small nitpick, but Turin-based VM's have been available for more than a month now [1], and they are a beast. [1] - https://techcommunity.microsoft.com/blog/azurecompute/announcing-general-availability-of-azure-daeafasv7-series-vms-based-on-amd-%E2%80%98turi/4488627 https://techcommunity.microsoft.com/blog/azurecompute/announ...
- dkechag 7mo agoThe testing was from mid-September to mid-January, Turin was released on Azure end of January. They simply did not make the cut, unfortunately I could not wait forever for them, they always seem to lag behind at least half a year :( I might do an addendum just for Azure...
- chrisweekly 7mo agoYIL (Yesterday I Learned) about Blue Dot https://tui.bluedot.ink/ https://tui.bluedot.ink/ a CLI for managing cloud resources that lets you compare prices, apples:apples, AFAICT. I haven't tried it yet but its featureset looks pretty great.