Is this HP z840 a good deal?

[Jeff Kilgroe]

Short: In single CPU mode, some PCIe slots operate at slower speed.. and some are not available.

This is true of all dual Xeon systems. Adding the second CPU adds more PCIe lanes. 40 lanes on a single Xeon (or i7 CPU) just isn't that much and often creates issues. This is another argument in favor of the dual Xeon over a single Xeon/i7 configuration.

 

[M.Halsell]

HP Turbo Drive

HP Turbo Drive

Jeff, is the HP Turbo Drive latency lower because of being connected directly the PCI bus? Unfortunately, it takes a much needed slot.

Also, how far away is the ver 3 Thunderbolt controller? And will that new spec diminish the need to rely on expensive video cards to drive 4K displays? In general, will Thunderbolt ever trickle down to consumer class machines?

 

[Jeff Kilgroe]

The Turbo Drive is just a PCIe SSD. In fact, it's a PCIe interface controller for a M.2 /NVM SSD -- similar to what is built into a lot of current motherboards and what is being used in the Mac Pro cylinder systems. It's faster than other SSD units you can install in the system or drive bays because you're not limited by the relatively slow SATA interface. It's PCIe v3.0 X4 direct attached flash storage. Lots of other products on the market that are the same. There are also other SSDs out there that are much bigger and faster (like the ones from Fusion IO) -- and currently a lot more expensive.

Thunderbolt 3 should arrive early in 2016. It offers a few small enhancements over Thunderbolt 2 in terms of ability or functionality. But the main difference is it doubles the data rate over the existing channels. A single Thunderbolt 3 cable will transport 40Gbps of data (Thunderbolt 2 is 20Gbps). This doesn't mean a whole lot for GPUs. Newer GPUs are pushing well beyond the abilities of PCIe v3 X8 and even now X16 slots. So while Thunderbolt gets faster, GPUs are getting faster and more capable too. Intel may open up some more options for Thunderbolt GPU support, but IMO it's still going to be quite crippling for any contemporary GPU.

As for Thunderbolt in consumer systems. I think this is becoming increasingly unlikely. Thunderbolt hasn't really offered a compelling reason for consumers to embrace it. The cables with active electronics are expensive and its only true advantage over USB has been the much greater bandwidth and improved latency. Two features that the vast majority of consumers don't need. USB currently operates at 10Gbps with v3.1 and will be offering 40Gbps, the same as Thunderbolt 3, with the v3.5 revision and a projected 100Gbps with v4.0 (that's a few years out though, 40Gbps v3.5 should launch by the end of 2016).

Thunderbolt 3 is also beginning to converge with USB and with the next release, we see a shift from mini-displayport connectors over to USB-C connectors. Ultimately, the two will fully converge, probably with Thunderbolt and USB both hitting v4.0. What Thunderbolt tech will offer is the ability to use active cables to extend range and probably little else. That's already how Thunderbolt 3 will operate -- use Thunderbolt active cables to gain additional range (optical cabling) or to access a few of the Thunderbolt features that USB doesn't currently have in v3.x.

 

[Eric Bowen]

Hmm... More to think about. When I assess the CPUs of an I7 5960X 8 core VS. the Xeon 2620 V3 6 core (source: http://cpuboss.com/cpus/Intel-Xeon-E5-2620-v3-vs-Intel-Core-i7-5960X) the resulting comparison is very close indeed. In fact, I don't usually overclock, but the I7 outperforms the XEON in that metric of performance.

Would it be it just as advantageous to go with a single CPU instead of two?

Dual Xeon systems have each CPU handling half of the ram slots and half the PCI-E slots on Dual Xeon boards. That means only half of each are active when you install 1 chip only. I would only recommend an X99 system for single CPU configurations whether I7 or Xeon since all ram and slots are active. Also the X99 has the Thunderbolt card option where as the dual Xeon boards outside of HP's add in card do not currently. 16GB Unbuffered DDR4 dimms are finally starting to hit the market but the speed is Jedec 2133. It will probably be next year before you start seeing 16GB DDR4 in the 2666+ speed range.

Thunderbolt likely moves to Optical as the standard was originally envisioned probably by TB4 but maybe later. Cable runs and transmission range really open up in the optical environment. USB C will have to raise the Power over bus capability of system boards which overall will help lead to the Thunderbolt Optical transition. The copper benefit is power over bus but likely can be resolved with using USB C connection in the future to power those Thunderbolt devices.

 

[M.Halsell]

Patrick F/Jeff K or anyone that can chime in: When using multi threaded applications, as the CPU's all reach 100% - have you been able to peak the physical RAM usage? So often, the CPUs are working hard, but the ram allocation stays surprisingly low or un-bothered. Obviously if I have texture heavy 3d render or creating a ram play cache, the memory usage will reflect, but just in general I rarely see the ram taxed heavily.

 

[Eric Bowen]

Memory management is application programmed and controlled with caching models in CPU processing based apps. The OS has some overhead management to that but the majority is application side and current caching models. GPU acceleration applications have the same basic memory management application side but not the Video driver effects the caching models. GPU acceleration pipelines through system ram for everything the GPU handles and back again for any CPU processing after. That is why GPU acceleration can show such variable results based on applications, CPU, ram and amount, and GPU in the system.

 

[PatrickFaith]

Patrick F/Jeff K or anyone that can chime in: When using multi threaded applications, as the CPU's all reach 100% - have you been able to peak the physical RAM usage? So often, the CPUs are working hard, but the ram allocation stays surprisingly low or un-bothered. Obviously if I have texture heavy 3d render or creating a ram play cache, the memory usage will reflect, but just in general I rarely see the ram taxed heavily.

I'm using about 2 gig per core, and the os is using around 32 gigto cache disk. i think i have 196 gig, but its not going above 64 gig used except when i am meshing water in maya.