Azure Dv6 VM Family Performance & Cost Comparison
A look at the performance and cost of Azure’s new Dv6 VM series and compared to previous Dv5 and Dv4 families.
Artificial Intelligence. Quantum. Cutting edge new tech is exciting. Makes for clickable, readable, shareable content. This isn’t about that. This is about boring, reliable, dependable tech instead. The tech that quietly gets on with it and powers most of what we all do and rely on every day. Like your financial risk calculations.
A few weeks ago Microsoft announced General Availability of the new Dv6 VM family. Powered by Intel’s latest Emerald Rapids Xeon CPU, its an upgrade to your tired Dv5 or even Dv4 machines. And I have at various times promised to provide COREx benchmark data to see how well it stacks up and if it makes sense to migrate your workload. It might have been a little longer ago than I realised! How time flies.
Well just as Intel announces its new CEO I’ve finally gotten around to writing this post. I’ve rambled on long enough, so the TLDR is just do it. Move to Dv6.
In pure performance terms, the new Dv6 family is around 34% faster than the previous generation Dv5 and a whopping 68% faster than Dv4.
That’s all well and good but it they’re priced too high your workload costs will trend in the wrong direction, and we want them dropping like the US stock market.
Luckily Microsoft have you covered on that front too.
Assuming you’re hammering the CPUs and fully utilising the machine (as you should be for HPC!) then performance adjusted costs are lower too. If you’re using spot capacity then Dv6 is about 27% cheaper than Dv5 and 42% cheaper than Dv4. Even using on demand prices its 21% and 38% cheaper respectively (prices as of March 2025).
If you want a little more, some alternative benchmarks perhaps, then take read of the following by Dr Darko Mocelj. Need to make sure I get on that preview list for future VM families 😁
damocelj
Test Methodology
For the sticklers among you the benchmarks were run and paid for by HMx Labs in our own Azure subscription in the UK South region. Each benchmark was run a minimum of three times, spinning up a brand new VM for each run. The results presented above are the average (mean) values. All tests used the 64 CPU variants.
