We do.
Small sites that have really bad or intermittent internet connectivity in remote locations and just need some small SQL + file/print + authentication services.
Cloud doesn’t work everywhere 
Sorry, I’m blind aparently. 
I see specification of tests 2700X only.
Can you, write text for me. Thanks.
he meant this slide
2 samples of tr2 are shown running with x390, x399 mobo’s (24 core, and 32Core) both state quad memory channels
Gigabyte x399 Aorus Gaming 7 Mobo, and x390 Witchaven Dev CR8 board
Price up boards and get back to us. That’s where you will see the major cost difference.
TR is binned for clock speed, EPYC is less aggressive…
EPYC low core count parts are also distributing their cores across 4 dies, so for example an 8 core EPYC can be very cheap - they can make use of a die with 3/4 broken cores in every CCX. i.e., the absolute trash good for nothing else in the lineup.
So long as the fabric and memory bus is good - it’s an Epyc 8 core part. That part wouldn’t cut the mustard for Threadripper (v1 at least) which needs 2 cores per CCX to work to reach the core counts…
TR CPUs will probably remain more expensive than low core count epyc for the above reasons…
edit:
the more i learn about Ryzen / Threadripper / Epyc, the more impressed I am at the genius AMD have displayed here. they have so many options for partially broken dies… having a defective die good for at least one of their product lines must be very close to 100% of the time…
1 Like
well intel hasn’t picked itself up since, if amd provides blades for epyc line - our company is certainly going to jump. We would be able to upgrade our memory capabilities and storage speeds by 2-3x
you can grab some SP3 socket for same similar price as TR4 socket. (ex. asus rog, and Supermicro H11SSL-NC ATX - some go for $300 with 8 channel memory) - sure you can get some really expensive SP3 socket board, but thats up to buyer.
Yeah but that would mean major tooling changes, supply chain logistic work, etc. On a lower volume part like threadripper, it probably wouldn’t save enough to outweigh the economy of scale by just placing a bigger order for the 4094 pin sockets, CPU packages and manufacturing plant to solder the dies to the package, etc.
Far easier to just use the same parts and just not hook some bits up.
they designed it well infini fabric and all, they don’t loose money - make massive profit.
As a datacenter admin, even if i had special need for ‘desktop’ like server - i wouldn’t go with desktop parts… but i could see how starting company could have. (you simply do not bother with building them manually… it would be too much work for servers - and then cooling jzuz christ - a shit-show with TR.)
OK. They used quad-channel for tests (memory configuration).
They can do it on octa-channel too!
AMD made sockets to last AM4, TR4.
Now making a socket and supporting it for 4 years or more is WISE and NICE. We can engineer out secrets.
This is opposed to Intel and a new motherboard for reasons and $$$
The engineer in me want’s to build something that lasts. Not some trumped up sub zero creep.
I think this should be the new thread name:
1 Like
This is some next level denial @anon5205053
It’s quad channel, it’s X399.
If there is an X499 platform it’s more likely to come from Intel.
When I said this I didn’t just pull it out of thin air.
The only thing that came from speculation is the X499 name.
Why do you think gamers nexus said the same thing three days after I did?
1 Like
Intel needs to roll out some power hungry garbage to back up claims, This year now.
How long before they ask forgiveness and and glue together intel 10nm chips. Next year I guess.
2 Likes
Here’s some benchmarks from the Epyc 7601(server version of the 32C TR) vs. the Xeon 8176 for those interested.
SPEC CPU2006 single thread integer
SPEC CPU2006 multicore.
The sleeping dragon is Epyc & 7nm now over last years shit.
Not sure if those benchmarks are useful given it says dual epyc cpu, unless that’s a dual 16 core epyc cpu which might give us some idea of how TR32core will work.
My apologies if this has already been brought up (there’s a lot to read through in this thread), but what performance difference is there for a workstation workload that 8 channel memory is perfeable to 4 channel? I thought it only mattered if you had a ton of I/O and not so much for raw compute power itself.
So I can with certainty say that TR2 has a 4Chan Memory Controller.
The CCX is split into near and far dies with regards to memory access and if you run latency sensitive software it has to be tweaked to take this fact into account depending on which core your thread resides in.
And the CPU even tries to do some of that itself. 
Something something testing
2 Likes
Those benchmarks are plenty useful if you have a lick of CPU knowledge.
If two of the die don’t have direct memory access they will need to use infinity fabric. It probably won’t affect general performance too much, but memory heavy tasks will probably take a big hit.
Could be one per die to eliminate the infinity fabric as bottleneck.
Edit:
Welp, there goes that hope.