Precision 7550 & 7750 Owners' Thread

It would be great if Dell could offer a special variant of the Precision with the PCIe topology different:

• CPU 8x to DGFF
• CPU 4x to Primary NVMe
• CPU 4x to TB3

There are very few situations with the current arrangement (CPU 16x to DGFF) where the highest end Mobile Quadro can completely saturate the connection. I am not sure if the Dell rep has ever commented about why it is not possible, but I think a few people could benefit from it if there was an option.

 

To add a bit to this great answer, not only do the NVMe drives need to share the PCIe 4x chipset bandwidth (also called the direct media interface or DMI bus), but so does every other non-GPU device. That means you are sharing it with all USB devices, networking, and in the case of the precision series (at least the 7740 and earlier), all thunderbolt devices. I haven't seen a schematic for the 7750 yet so it's possible this changed, but given the motherboard layout and design of the DGFF form factor, I would be exceptionally surprised if this were the case. In the 7x40 and earlier devices, only the DGFF gets dedicated CPU PCIe lanes. (Schematic attached)

I also suspect any speed benefit you see from using RAID 0 on this platform, especially to random I/O, would be offset by increased latency as already proposed. One reproducible exception to this rule is the case of using some QLC drives that have a small SLC cache. Once many of those drives exhaust the SLC cache, they slow down considerably and using RAID 0 with two such drives can effectively double the size of that much faster cache.

For what it's worth, this condition isn't unique to Dell or the Precision line. Almost all Intel laptops share this PCIe design. There are a few exceptions to this rule, but they are niche cases of devices either coming from the factory with a fast RAID 0 SSD (Asus, MSI, etc.) or Alienware routing PCIe lanes to their graphics amplifier/TB3 controller for fast external GPUs.

Edit: Ninja'd .

 

I've raised this point (and my desire for an Alienware AGA or similar on a Precision) with my Dell rep every year. This past month he finally connected me with an engineer who explained that yes, the PCIe topology would ideally allow for 8x lane assignments to TB3, or NVMe, but to do so significantly increases the complexity of the motherboard's design. PCIe lanes would need to be routed from multiple locations on the board back to the CPU rather than to the chipset. Doing so requires either a larger board or sacrifices must be made elsewhere.

When I raised the example of the 51m accommodating a different PCIe layout, he commented saying yes but not only is the 51m motherboard larger, it sacrifices a lot in terms of EMI shielding (something I have long suspected) and using a memory topology that limits the speed the RAM is able to reliably operate at in the 51m. The recent memory redesign of the 51m R2 (from 4 dimms down to 2) is a result of faster ram being desired on an already crowded board without opting for soldered RAM. On a smaller precision board, he said they are unwilling to compromise on EMI/memory trace quality or reduce the number of NVMe drives / SO-DIMMs. While he would be fine with a larger machine, customers are not, and his team is already struggling to accommodate PCIe gen 4 with their current design.

Food for thought.

 

Agreed here... Given the architectural similarities, and the fact that the DGFF connector is the same even if the card is shaped differently (should be 16 PCIe lanes to the dGPU), I strongly doubt that the 7750 is any different than previous systems in this regard.

Moving to PCIe 4 might help out here somewhat (doubling the bandwidth of all of the PCIe links), and it looks like this might happen with the next-generation (11th gen / 2021) systems. Other than that, seems like this is the way that the architecture is going to be. (I dunno... maybe at some point, Intel could add a few more PCIe lanes to the CPU and increase the PCH connection to 8 lanes?)

 

Yes, and here is the diagram for DMI 3.0 for intel 10gen CPUs used on 7*50 precision models. So, all these connections to PCH will share the unique DMI 3.0 which is connected to CPU only with 4 lanes, providing maximum 3.94GBps bandwidth.

Trying to do RAID 0, to get more than 3.7GBps which is the limit of my NVMe single drive is not possible.

I wanted to order a tb3 to rj45 10Gbs adapter, to take advantage of my 10Gbs network, so it will not be possible to use all the limits of the network and the ssd drive, because while downloading over network and writing to the NVMe drive, all of this is capped by the ~ 4GB DMI limit.

In the past I was connecting my laptop to my external display through the tb3 port, so now I am going to connect directly to the display port, as it looks like it is connected to the CPU directly without passing by DMI while tb3 is sharing DMI…

 

Do keep in mind orders of magnitude here (32 Gbps = 4 GB/s). You should be able to use a 10 Gbps network adapter fully and write the data to a SSD drive using only 20 Gbps of bandwidth, which fits inside of the 32 Gbps limit for the PCH. If you are writing other stuff besides what you are downloading to the SSD then you could hit a limit, sure. But I imagine in most cases you won't be stressing everything at once.

I'm not sure if external displays connected to the USB-C ports count against the PCH bandwidth limit. If you have a dGPU and you either have "discrete graphics output" enabled or graphics switching disabled in the BIOS, then displays connected to the USB-C ports (using DisplayPort mode) are connected directly to the dGPU. A 4K display running at 60 Hz would use about 12 Gbps of bandwidth constantly (if I did my math right?) so it would be a noticeable hit if it did use PCH bandwidth.

 

No worries about connecting a displayport adapter to Thunderbolt 3. The display outputs from the CPU and dGPU both go through a mux switch before going to TB3. Display signals are completely separate from the DMI. Any USB or PCIe connections from the Thunderbolt 3 will go through the chipset.

 

You are right! So, I can read/write in parallel all 10Gbps network data using 20Gbps and have 12Gbps (1.5GBps) available on DMI for other stuff. Thanks also for GPU part clarifications.

Well noted thank you!

 

Another question concerns the remote management of our devices.
There are times when I prefer to travel with lighter, cheaper device that just serves to remotely connect to my precision and even if it has been stolen or damaged, I am safe (heavy network tasks can also be done regardless my actual "bad" connection) …

I connect via RDP, but I want even more control, I want to be able to access the BIOS of the device, power off and power on it remotely if necessary.

I have 2 options:

1) Connect a KVM device (keyboard-video-mouse over IP) to my laptop. Expensive option and I am not sure if this kind of device exists for laptops (for desktops there are devices like latronix kvm over IP one).

2) Use Intel vPro AMT and try to enable the KVM feature

So the question is, has anyone ever been able to use the KVM of Intel's AMT technology on the latest precision mobile models?

To support KVM, the CPU must have integrated graphics. This is why, for example, a precision tower model like the 5820 with a xeon w-22*5 processor cannot have KVM enabled, even if the processor supports vPro.

So in case of precision 7*50 models with xeon w-108*5m CPU and vPro activated by DELL, normally there is no reason not to be able to use the KVM because this CPU also has built-in graphics.

Before I dive in to do some testing, I'll appreciate if anyone can share their experience on this (I'll have access to my machine by next week).

There is no much information on the web about this. I found some information from Dell, it concerns Skylake processors. Here is the information about the supported processors but this information is not updated recently, and the list of CPUs provided does not include i9 or xeon-w108*5m CPUs ...

 

I haven't tried Intel vPro AMT. But a standard hardware remote KVM should be fine. If the laptop is left powered on but with the lid shut, then the BIOS is fully accessible via USB keyboard/mouse and an external display.