AMD ThreadRipper Reviews [Aug 30 - Tom's 1920X]

[LightningZ71]

Having dug through the numbers on the Anand game mode followup, I think I've come to these conclusions:
1) game mode looses the L3 in the partially deactivated die. This absolutely has a negative impact on game performance as less data is close to cores.
2) part of the reason that the 16c/16t performance with UMA is as good as it is because it keeps that L3 active and can take advantage of it.
3) part of the issue with gaming in creative mode with all threads active is contention on the links between the dies between inter core memory synchronization, core to far memory access, and having to reach across to far PCIe devices and hard drives.
4) in SMT disabled mode, you have half of the thread synchronization traffic on those links, leaving more bandwidth available for other tasks.
5) you have twice the L3 per thread, also reducing the hit from memory latency.

Ideally, a gamer would want to run in SMT off mode, UMA mode, and with the highest clocked DRAM they can manage. Also, with SMT disabled, individual cores will generate less heat, and heat will be more evenly distributed through the package. I suggest that this will enable the highest possible overclock without blowing the power budget and may result in the highest possible scores in benchmarks.

I'd like to see that theory tested.
(Edited due to autocorrect...)

 

[LightningZ71]

I also have questions about the PCI-e implementation. What is the impact on gaming performance when in gaming mode when the PCI-e graphics adapter is located on a PCI-e slot that is connected to the inactive CPU die? While I appreciate that there is prodigious bandwidth between the two dies, there is a latency penalty that has to be payed to cross that IF link set between them. It stands to reason that, if game mode keeps game data close to the enabled core, but the GPU is located on the inactive core, that it will have to execute ALL of its memory transfers across those IF links to the active core, which will then have to conduct its own memory snooping to verify that the requested address is indeed in its local memory, and then begin the read. That's a LOT of extra latency for each transaction. It gets worse when the system has to request memory from the far die with a GPU located on the far die as there will be contention for the IF links for memory reads and writes in both directions at that point.

A select group of gaming benchmarks need to be run with Gaming mode enabled and the GPU tested in each PCI-e slot to see if there is an effect there that's measurable. This will also see if AMD's gaming mode software can determine where the primary GPU is and select the core to deactivate based on that information.

The same goes for NVMe hard drives. What's the penalty for a drive being on the inactive die in gaming mode? There's a whole lot of things that should be tested here to figure out what's going on.

And, again, as I suggested in the previous post, any affect from this situation would be diminished in Creative mode, SMT disabled, UMA enabled, as threads would be on both dies, both L3 caches would be active to reduce memory access latency, and threads could be moved closer to the devices that they access.

 

[nathanddrews]

Updated with Tom's Hardware 1920X review:
http://www.tomshardware.com/reviews/amd-ryzen-threadripper-1920x-cpu,5183-13.html

 

[Markfw]

Updated with Tom's Hardware 1920X review:
http://www.tomshardware.com/reviews/amd-ryzen-threadripper-1920x-cpu,5183-13.html

I originally thought the 1920X would be a waste, but I see that higher speed and lower price make it a real contender in its price class depending on the workload.