News AMD Threadripper Pro Listed at $6,000 for 64 Cores, $3,000 for 32 Cores

[Admin]

Retail pricing for AMD's new Threadripper Pro processors has been posted at retailers.

AMD Threadripper Pro Listed at $6,000 for 64 Cores, $3,000 for 32 Cores : Read more

 

[mdd1963]

$6k! Is that all! ?

Take my money!

 

[samopa]

128 lane of PCIE 4.0, great for mining, not so great for workstation ...

 

[Jim90]

"While Threadripper Pro pricing is eye-watering " --> keep this in context with what you're getting in comparison with the competition, and the intended consumer.

 

[mikewinddale]

This makes me extremely happy. I tend to perform tasks that are more memory intensive than compute intensive. My most recent project took 3 days to run because my Ryzen 7 2700X + 64 GB RAM ran out of RAM. I've been frustrated by the fact that EPYC clock speeds are so much lower than Ryzen and ThreadRipper speeds, so that if I wanted registered memory (and the attendant higher capacity), I'd have to sacrifice 20% of my clock speed.

The price on the 3955X Pro with 16 cores is actually relatively affordable, compared to what I expected it to be. I wish we could buy the 3945X with 12 cores, but still, the 3955X with 16 cores is relatively affordable. (Again, RELATIVELY.) So I'm happy.

The EPYC 7F52 - with 16 cores - has a clock speed of only 3.5/3.9 GHz, yet it costs $3,100. So the ThreadRipper Pro 3955X has a higher clock speed yet costs only about 1/3 as much. I'm happy.

The only thing I wish is that AMD would simplify the product stack by merging ThreadRipper Pro with either ThreadRipper or with EPYC. They could either make EPYCs with higher clock speeds to replace the ThreadRipper Pro, or they could make the ThreadRipper support registered memory, with motherboard support optional. I'm not sure why AMD felt the need to segment their products this way. But whatever.

 

[CerianK]

Unless someone has suggestions (e.g. rent dedicated server space on each first, or ask ServeTheHome?), it is unfortunate that I cannot be sure what the performance difference would be between the 3990X and 3995WX in my specific workload (TestU01 BigCrush) with 128 threads loaded. Each process thread uses anywhere from 150MB-850MB, so I only need a maximum of 128GB of RAM.

To side-step the question, I was tempted to buy a pair of Epyc 7742 engineering/qualification samples to save money while increasing the performance further, but that opens up a whole other can of worms.

Certainly others are in this same type of situation, but perhaps most are in a more 'cost is no object' situation.

 

[lazyabum]

Well we can always count the Intel Xeon Platinum 10992 to come through with half the performance for twice the price.

 

[msroadkill612]

This makes me extremely happy. I tend to perform tasks that are more memory intensive than compute intensive. My most recent project took 3 days to run because my Ryzen 7 2700X + 64 GB RAM ran out of RAM. I've been frustrated by the fact that EPYC clock speeds are so much lower than Ryzen and

I inexpertly puzzle about the your surprisingly common predicament - the ~effective 64GB ceiling for super cost effective AM4 - yet it seems that with the advent of nvme, there is an opportunity to simulate vast amounts of memory, or even better, code around the problem.

an am4 x570 has 16 lanes of pcie 4 bandwidth available for nvme if using an 8 lane gpu, & 24 lanes if headless.

4 of those available lanes are allocated to chipset, but multiple nvme can share this chipset bandwidth.

thats a lot of bandwidth for nvme (2 GB/s per pcie 4 lane) - 64GB/s theoretical total e.g for a realistic 8GB gpu rig as above.

Cleverly utilised, this could be used as a ram extender for many of these ram intensive tasks - it seems to this newb.

I can see how OS memory swapping to raid arrays as a ram extender could be fraught, but with an array of up to ; 3x pcie 4 cpu linked nvme, & 5x chipset nvme (2x pcie 4 x4 & 3x pcie 4 x1),,,, cleverly distributing the working data around such an array could preclude the need for many operations that now use dram. Its still slower of course, but days could become hours very cheaply.

Note also, that each nvme can affordably have 8x+ processors & substantial dram cache

Perhaps some simple calculations could even be performed independently by the nvme

In short - code should consider if data really needs to reside in ram during problem execution - an array of nvme, often w/ prefetched data from nand to its dram cache, can be a productive substitute - and is hugely capacity scalable.

From what i hear, even the big players like to save a buck when they can, but for some jobs, anything but the fastest is a false economy.

once the task becomes routine, a cost effective farm is preferred to batch process big routine tasks. A source told me the 3900x was the sweet spot cpu for his such big math clusters