Indeed. And time (cuda, MKL code etc)

It's also for a sprint of few month, to do reasearch and MVPs. Later I might go with bigger things.

Such as 4xTitan RTX etc

I think it basically boils down to your platform not making sense for your budget: you're sacrificing some pretty big gains in return for some pretty small gains. If you're 100% committed to Intel (though I seriously suggest you re-evaluate those concerns), I think you should go with older server gear rather than the current-gen consumer platform, given your budget. AVX-512 often doesn't even offer that much of a performance gain (counterintuitive, I know.)

Similarly, I think you can get way more bang for your buck by getting multiple 1080s/1080 Tis instead of a 1 or 2 2080 Tis.

I'm having also loads of unit tests running as I code, so I need some really reactive CPU. I have a 6700k that is sluggish, and it can't even play a 4k H265 video. And I'll be processing loads of things like that.

I'm not suggesting a 6700K, I'm suggesting server grade gear (probably dual 2011 or 2011-3.)

Anyways, if that's something you're interested, I'm happy to spec something out for you. If not, I don't really see a good option besides spending the cash and going X299.

I own a 6700k on a matx, and I'm looking for something faster and more expandable.

I have also a fair amount of compute on 1 core (the UI thread), hence the 9900k choice.

Okay, if you really want FP16 and lots of memory, maybe it's best to go with a Quadro or Tesla. I know, that's why I'm recommending server gear.

Although since you're using this for just a few months, I would seriously recommend just using AWS or something

Compute with AVX512 can be 4x faster than without, this depends on my code.

I got about 4-5k as budget for the update (new WS), so, no tesla.

There's older Teslas, but I get your point

I would be surprised if you hit that fast with AVX, but okay.

Frankly, it really sounds like you're demanding a feature set that you can only get on X299. I somewhat doubt you're going to need everything, but at the end of the day, it's up to you.

For $5k, you could also get 350 hours on Amazon's top end box

The original question sounds quite simple in the end. And not that much important,

No AVX-512 tho

I'm also using the new gear 100+ h per week... So it will be cheaper at home.

Your training time is going to be a lot lower with AWS, but yes, I see your point

Frankly though, I don't think either Z390 or Z370 makes sense

You've kinda worked yourself into a corner where you don't have the budget for the feature set you're demanding

And I really don't see myself transmitting 10-60GB for each workload (once every few minutes)

Once you figure out what you're willing to sacrifice, I'll be happy to help

But until then, I don't think there's much further I can do.

So, here are the options as I understand them: sacrifice Intel, AVX-512, memory and PCI-e bandwidth, FP16, overall GPU compute performance, or your budget

You challenged me on the whole solution, and I had to explain the requirements. I thank you for that. This helps me to reconsider things.

No worries. Let me know what you decide.

I'm almost certain it will be 9900k + 1-2 2080ti, rather than a 9940x + 0-1 2080

Also probably the Formula.

BTW I'm not sure we can get RTX/Volta GPUs already on second hand market already.

Anyway. Thanks for your time.