Fastest CPU for single-thread applications

[olkipukki]

I’m looking to replace my primary Windows workstation that built 3+ years ago and based on AMD AM4 Ryzen 3000 series.

My initial thoughts to update CPU with 5800X-5950X (motherboard supports that), but seems not much gains for single-thread apps - I’m after at least 30-40%  raw performance increase.
Alternative, completely new build AMD AM5 Ryzen 7n00X (n=7,8,9) with or without 3D … and I found myself in the same situation as years ago -  MB’s only for gamers and nothing yet released for a workstation class.

What about other options?
I have ruled out EPYCs and Threadrippers - these lagging behind for single-thread tasks.

Is everything wrong with Intel?
I don’t mind to come back to Intel world if its better than AMD today.

Any suggestions?

[brucehoult]

You need to specify the workload.

The latest Intel CPUs are damn fast for anything with a working set that fits into  the 2 MB L2 cache on the i7-13700 models -- especially when they turbo to 5.4 GHz.

Note that the 12700s have 1.25 MB L2 cache, the 11700s 512 KB, and the 10700s 256 KB. This can make a huge difference for certain tasks -- much more than the raw CPU core differences.

The increase of L3 cache from 16 MB to 25 MB to 30 MB also makes a big difference to another class of tasks. But if your code has this size working set -- or bigger, and using RAM significantly -- then you're better off with an Apple M1/M2 even though the CPU core is a little slower.

This is also where Threadrippers, with up to 256 MB of L3 cache, are also pretty good. The Apple M1 Ultra tops out at 96 MB of L3 (Apple calls it "system level") cache.

But the Apple chips have incredible bandwidth to RAM if your task really uses a lot of data: up to 800GB/s. Compare that to 76.8 GB/s maximum for the i7-12700.

Note that if your task has a very large working set then the Apple M1/M2 may be the fastest EVEN RUNNING x86 code, because the Rosetta JIT is very good. Or, if you want or need Windows but don't need x86, Windows 11 for ARM runs at native speed (way faster than any other ARM machine) in a VM on an M1/M2. You can run x86 code inside that, but you get Microsoft's sucky x86 emulator, not Apple's great one.

[beanflying]

Me too sort of so I started catching up from when I built my 3700X 32Gb 5700xt around release time for the CPU.

Single thread performance is still sort of ruled by Intel but at a $ cost and way more Watts to get there. So if you can drop a few percent in speed there is real $ to be saved initially and ongoing by staying with AMD.

Given you own a MB that will handle the AM4 CPU's and there is some killer deals out there on some in the range just consider a new CPU and definately take a look at your memory timings (I got a good boost on mine initially doing that)

The AMD 3D vcache thing is Gamer use only as it drops the clock speeds and the testing I have seen on it shows the CPU's to SUCK at productivity tasks.

The more I have looked at it over the last month I guess sees me looking out for a 5900 ot 5950X then sit tight until the AM5 MB's and DDR5 get to a more sensible price. Just had a quick trawl locally 5900X from a local retailer $300USD+tax equivalent, the remaining 5950X's are a lot more.

[SiliconWizard]

That will give you at least something based on benchmarks:

https://www.cpubenchmark.net/singleThread.html

[Berni]

Yep Intel seams to be leading in terms of single threaded performance for just raw compute power.

So the way for getting maximum single threaded performance is to get the highest clock speed 12th gen Intel CPU, slap water cooling on it and overclock it for even more clock speed. The i9 13900k for example will comfortably do 6.1 GHz even on all cores if enough cooling is available. If you really want to go for single threaded you can also disable some cores in BIOS so that the few cores enabled can get a bigger cut of the power budget pie.

This is actually used as a strategy in servers used for high frequency stock trading, they even sell such watercooled overclocked rack mount servers.(Might even come with cherry picked CPUs so that you don't have to play the silicon lottery)

[Haenk]

If you are creating your own software, Apple M1/M2 might be worth a try as they have extremely high memory bandwidth. That will certainly help with processing large data quantities.

[50ShadesOfDirt]

OP didn't reply (yet) with details of requirements ... so I'll add my solution, to help fill the vacuum

Virtualization ... yep ...

The problem with many single-threaded applications, or trying to get to the best single-threaded performance, is the way in which we try to solve that problem with today's technology. If this were DOS, and you had a truly single-threaded application, then a faster cpu would be the common fix, followed by any of the other typical bottleneck fixes.

Today, we have windows (insert your choice of OS insanity), doing all kinds of things in the background ... you don't get to do things in a single-threaded approach any longer. Unless ... you go back to the point where you could give a single-threaded app the undivided attention of the processor. There are many ways to do this today, and lacking details of the application in question, you can also consider some form of virtualization.

Given a multi-core cpu, just assign one or two cores to the host OS (and that overhead stays there for the most part), and get your application into a Virtual Machine state where it has the best chance of running full out. Gobs of whatever it needs ... cpu, ram, ssd, etc. Perhaps even a docker vm, to reduce all other overhead to something approaching zero.

Again, don't know the exact application, or the exact bottlenecks you faced with the old(er) platform ... could be many ways to skin this critter ...

Hope this helps ...

[beanflying]

Much as the OP went MIA in early May (hope they are ok  )

Just installed a 5900X and a Peerless 120SE cooler to replace my 3700X on an otherwise unchanged B450 platform. Very happy with the pickup in performance and no fuss other than having to tweak the BIOS settings again as they got reset. This is without any Boost to the CPU other than what Ryzen does itself so there is some more there I guess but it will likely come at some costs elsewhere and likely only 3-5% gain if that.

Quick Geekbench run, this is the second to get everything warmed up a little first so no real issues with Temps either on the cooler.

[Halcyon]

Much as the OP went MIA in early May (hope they are ok  )

Must be away on summer vacation at the mökki.