[SOLVED] Mega-expensive mega-fast CPU?

[mujmuj]

https://www.amazon.com/gp/product/B072KTSCCS/ref=ppx_yo_dt_b_search_asin_title?ie=UTF8&psc=1

This is the processor that I currently have. It's already quite high-end. But I am a data scientist who uses a really huge data. I am willing to pay even $5000 or more if I can boost speed just for CPU. Is there a better thing for this situation?

 

[kanewolf]

Start by using some of your budget on AWS instances. Benchmark what kind of host improves your performance. More cores? Higher single threaded performance? More memory bandwidth? GPU accelleration? There are LOTS of options. Using different AWS instances for tens of dollars is a much better way to determine WHAT to buy than just asking a board.

 

[logainofhades]

Easy option, would be to drop a 7980xe, into your current board. You would go from a 10c/20t to an 18c/36t. Though if your software uses don't make use of the extra cores and threads, it's pointless to go that route. So without knowing exactly what your software needs, any upgrade suggestion is just speculation.

 

[kanewolf]

Easy option, would be to drop a 7980xe, into your current board. You would go from a 10c/20t to an 18c/36t. Though if your software uses don't make use of the extra cores and threads, it's pointless to go that route. So without knowing exactly what your software needs, any upgrade suggestion is just speculation.

The base clock speed also drops from 3.3Ghz on his current CPU to 2.6Ghz with the 7980xe. That xe CPU could definitely be slower than his current. The xe CPU does have more cache than the current CPU so some performance could be gotten with that.

This is why I recommend the "try before you buy" approach with AWS. You can get a 2, 10, 20 or 60 core VM for very little money. Benchmark your problem before buying anything.

 

[logainofhades]

The 7980xe is easily overclockable, with proper cooling. Even a 4.0 should be doable without any real effort. Granted the same level of cooling would probably be needed even for what they have now. That family of CPU's are quite the space heaters.

Personally, with the 3950x coming, next month, I would sit tight, and see what the results are. Might see a shakeup in prices, also, as Intel already has lowered the prices, on some of their chips, in response to the current Ryzen 3000 chips.

 

[kanewolf]

The 7980xe is easily overclockable, with proper cooling. Even a 4.0 should be doable without any real effort. Granted the same level of cooling would probably be needed even for what they have now. That family of CPU's are quite the space heaters.

Personally, with the 3950x coming, next month, I would sit tight, and see what the results are. Might see a shakeup in prices, also, as Intel already has lowered the prices, on some of their chips, in response to the current Ryzen 3000 chips.

Overclocking and "data scientist" don't usually go together. Most "commercial" users want plug-and-play with 100% reliability. Not the usual teen or 20something on this board that is willing to do anything to get higher performance.

 

[Alabalcho]

I wonder why noone asked whether the particular application can be run on several CPUs in parallel. This way, CPU performance becomes irrelevant.

 

[stinkycreature1]

Threadripper 3rd Gen: Allow me to introduce myself

 

[geofelt]

What is the nature of the work you do?
Do you run just one app? If so, what is it.
From your post, I might assume that you have cpu limitations.
But, you also allude to "huge data" Are you possibly I/O constrained in what you do.

How well multithreaded is the work you do?
If you are largely single threaded, the current 9th gen intel K processors will run at 5.0
compared to the 4.3 of your current processor.
If you can multitask well, there are some very strong processors available with many more threads than you have.

Then, is it possible to use graphics cards to accelerate calculations.

 

[CountMike]

A detailed look at AMD’s new Epyc “Rome” 7nm server CPUs

Epyc "Rome" continues its little sibling the Ryzen 3000's mad dash forward.

arstechnica.com