Sonnet Breakaway Puck RX5700 Thunderbolt eGPU review and teardown

---

This article, by Leaves, has been adapted for English-language audiences and re-posted with permission.

[Original article link]

[Sonnet eGPU purchase link]

---

Introduction

While other teardown posts of the Breakaway Puck series can be found, none reveal the detailed structure. The reason for my interest in this Thunderbolt 3 eGPU is not its MXM eGPU feature, but the unique Intel JHL7540 controller it uses.

The JHL7440 is the 7000 series Thunderbolt 3 external device controller designated by Intel, so you can’t find information about the JHL7540 chip which is intended to be included on PC mainboards. In the follow-up dismantling process, we will see features unique to this product.

Let’s get through the product specs first:

Sonnet official introduction for eGPU Breakaway Puck RX5700：

This is already the 2nd generation of the Breakaway Puck equipped with AMD Radeon RX5500XT or RX5700 GPUs. The previous gen is equipped with RX550 or RX570 GPUs, a single Thunderbolt 3 port, video, but no USB:

The first-gen eGPU Breakaway Puck is equipped with Intel DSL6540 (see teardown), and it can be clearly seen that some design features were abandoned on the board. It is estimated that another Thunderbolt port of the 6540 could be connected to a USB Hub to provide dual USB-A and a Type C port on the side which is different from the current Sparkle/Taidu MXM eGPUs:

Predicted Design

Starting from the I/O configuration of the current eGPU breakaway puck, two Thunderbolt ports and USB3.0 coexist. It seems that the configuration of JHL7540 should be more similar to JHL7440 than the previous gen JHL6540.

The video ports should all be exported from the MXM slot, so the structure of the entire eGPU should be very simple.

Based on the Thunderbolt infrastructure of JHL7540 + PD controller (likely to be CYPD5235, less likely to be a TI controller), PCIe Gen3 x4 is supplied to the downstream MXM slot, and there will be a USB3.0 Hub connected to the downstream USB of JHL7540 that provides two 5Gbps USB ports.

Since the two Thunderbolt ports are exposed, it will not be a cascaded structure of dual Thunderbolt controllers like the Razer Core X Chroma (in fact, it is not impossible to combine two dual-port Thunderbolt controller, but the probability is low).

Unboxing

Not many contents, just the Dock itself, original Thunderbolt cable, and power supply.

Teardown

Remove foot pads to expose screw locations.

One of the screws at four corners is divided into two sections, so you need to change the tool to take it out.

No major components can be seen on the back, only that two screws have been sealed with glue.

The top shell can be removed by shaking, and there is a cable connected to the LED.

The GPU part cannot be taken out as a whole, the heatsink & MXM GPU part are being stuck, so we have to remove the fixing screws from the gap of the fan blade

Then the bottom Thunderbolt board falls out of the middle frame, two PCBs sandwiched the middle frame, and then two plastic upper and lower shells were sandwiched by the middle frame too. The way of assembly is really stupid.

The 4 long screws both position fasten the boards. In fact, the entire case cannot be fastened in place exactly as there is always room for shaking making it feel cheap. In summary, it is not a solid chassis design.

Closer look at TB3 Board

(click photos to zoom in)

The controller itself is just above the two Thunderbolt ports, and the silk screen is indeed JHL7540. The controller uses two discrete TPS65983BA. It is very rare to use a single-port TI controller in the 7000 series product. The only similar case in my impression is the Corsair tbt100 dock.

Closer look at MXM GPU

[editor’s note: stay tuned for a separate article outlining the GPU design]

This heatsink is indeed much more beefy than the previous gen, but personally, I still prefer the MXM heatsink style that comes with the CLEVO models, so that the thickness of the eGPU can be further reduced at the expense of more fan noise.

TB3 board IC tracking

TB3 PCB Details

According to some open source MXM designs, the pinout map of MXM slots is roughly as follows:

At the corresponding position of the slot, you can see the related wiring of PCIe. Although 16 PCIe lanes are available in the MXM interface, the Thunderbolt controller can only provide 4.

4 Transmitter pairs + 4 Receiver pairs + 1 CLK pair, a total of nine differential pairs can be seen.

WAKE# and PERST# and other PCIe signals will not be investigated here.

On the front side, you can see the wiring in the corresponding position leading directly to the JHL7540 As a full-duplex protocol, the PCIe Rx/Tx differential pair needs to be equipped with differential capacitors, they are obvious next to the JHL7540

Then the problem comes. This part of the wiring is also connected to the Thunderbolt controller, but it is obviously not the route for the Thunderbolt port and its not PCIe. So what is it?

In addition to the PCIe wiring on the back, the two video output ports of DP and HDMI are clear, but this position obviously does not correspond to the mysterious wiring in the previous picture.

Instead, in the area adjacent to DP and HDMI, there are two sets of differential pairs starting from the MXM slot, which are the wiring corresponding to the front side.

The correspondence between the top and bottom layers is as follows:

According to the basic interface provided by JHL7540 and the pinout defined by the MXM standard…

These two sets of high-speed differential pairs must be DP1.4 links output from the MXM to the JHL7540. Each set includes 6 differential pairs (4 main Tx Link pairs + AUX differential pair, and an extra pair for Configuration Channel or HPD)

This is quite unexpected and completely beyond conventional eGPU design.

The JHL7540 with two DP inputs enabled will support the MXM graphics card to directly output dual video streams through the downstream Thunderbolt port – something never seen in previous Thunderbolt products.

Other features of the PCB are less interesting.

Video outputs and USB are equipped with TVS protection, and the USB is also equipped with power filtering and port protection.

Each TPS65983B is equipped with the same independent power supply design. So in fact, it should not be a problem for both Thunderbolt ports to output 60W. Although I don’t like limited power supply of only 60W output, but this is normal.

The three power supply designs with 2R2 power inductors also look the same; but placing two identical modules so close to the right edge? Maybe it’s a USB-specific power supply design…

On the left and right of the QR code sticker, you can see a row of pads and two pin headers exposed, which should be the LED for indication and the interface for Debug.

Topology

Although the analysis process is not complicated, and there are only a few ICs, the conclusion is still unexpected. According to the current information, it can be summarized as follows:

Functional & Performance Verification

Performance Benchmarks

Benchmark specs:

• 3D Mark Time Spy v1.2
• AIDA64
• RX5700 with default settings (no overclocking)

Host #1 Notebook platform specs:

• Intel Core i5-8265U
• 8GB RAM
• RX5700 GPU score fluctuates between 4000~4300

Host #2 Desktop platform specs:

• Intel Core i9-7980XE
• 32GB dual-channel RAM
• GC Titan Ridge add-in-board provides TB3 ports for test; display is set on 2K resolution connected to 2080ti, not eGPU
• RX5700 GPU score ~5500

With a Thunderbolt 3 SSD enclosure cascaded from the secondary Thunderbolt port, we can achieve ~3000MB/s read speeds.

Charging performance test

Indeed as indicated in the manual, only 60W laptop charging is supported which is disappointing. Charging power is limited to 15W on the second port.

Dual-port Thunderbolt function test (video output observation)

Lets confirm what differences there are, if any, between the two Thunderbolt ports on the Breakaway Puck.

After swapping the host cable back and forth between the two ports, there seems to be no functional difference. Except for the power supply config (15W vs 60W) and the host->eGPU connection sometimes failing when connected to the secondary port requiring the host to be reset.

When connected and running, all features work normally with daisy chaining more Thunderbolt devices, USB devices, and even DP Alt Mode devices, even after swapping the cable between the two ports.

A monitor connected to the second Thunderbolt port will only be recognized by the RX5700 GPU driver, not the Host internal GPU.

DP Alt mode display, only supports up to HBR2 (monitor support issue), so Control Center shows 5.4Gbps x4, of course with DP IN design previously discovered, this is expected.

For conventional eGPUs based on JHL7440 designed according to Intel guidance, the downstream Thunderbolt port is not supposed to be exposed externally. And even if it is exposed, it can only output the video stream from the Host internal GPU, not the eGPU.

After disconnecting the external GPU via the AMD XConnect™ software tool, the video output function of DFP TB port stops. Can this second TB port do anything except output the eGPU video stream?

The test after removing the MXM graphics confirmed this guess. Without loading of the AMD GPU and software driver, the Thunderbolt port still cannot output the Host video stream, but USB, charging, and TB3 daisy-chaining all work as expected.

Case Removed observation

The fans will work only when the Thunderbolt connection is established and the AMD driver is loaded. It seems that the fan controller uses MXM signals.

USB Mode Compatibility Test

The device cannot be recognized in USB/DP Alt Mode, and the USB Hub function is also invalid. So the device is only compatible with Thunderbolt hosts.

Conclusion

Lets summarize the drawbacks first:

• As noted before, the top and bottom covers are loose and the assembly is cumbersome due to its design.
• I don’t understand why side-mounted port interface packages are used. Using dual-port stacked packages common on desktop mainboards should further increase the number of interfaces. As for the wiring difficulty, it should not be a problem for this kind eGPU board; GL3523 can provide 4 downstream USB, all of them should be exposed here.
• 5Gbps USB again, 10Gbps please!!!
• Only 60W charging power, too bad. Although the power input of most iGPU notebooks only reaches ~50w? Maybe “Design for Mac” will only allow designers to consider 13-inch MacBook devices.
• DC power interface uses a 4-pin aviation plug style, and it must be aligned when plugging in. Dell’s 7.4mm round hole plug has no problem for 300W; Caldigit, Belkin and other third parties are using it. So why this more complicated design?
• Sonnet’s official marketing claims its the smallest eGPU. Most people have never seen the Lenovo Thunderbolt 3 Graphics Dock which is smaller but has a non-replaceable GPU since it doesn’t use MXM.

• Only manuals and FAQs on Sonnet website, no driver/firmware downloads or updates

When I got this eGPU, I expected it was just a small update to place a JHL7540 in the previous generation design to gain a downstream TB3 port and 2 USB while having fewer monitor outputs. But after examining the PCB, I was pleasantly surprised. In the official marketing, the support for Apple XDR 6K display was briefly mentioned but I didn’t pay too much attention. I thought it was just a marketing promotion and Sonnet specially bragging about it.

After dismantling, the XDR support and “designed for Mac” refer to such a strange solution in the actual product. In fact all 4 DP outputs of the GPU are fully utilized, but in a way that no previous eGPU has ever used.

Although Sparkle/Taidu has recently launched similar MXM eGPU products (JHL6540 + JHL6240 solution), the actual PCB structure is quite different.

This RX5700 eGPU brought me far more joy than I expected. What’s most interesting is that as a Thunderbolt device, it does not completely rely on the data flow output by the host, but can replace the host video stream with its own in the daisy chain. Its a pity that this design is not needed in normal scenarios. [Editors note: having these capabilities be the norm could drive more innovation.]

In addition, we’ve learned that the JHL7540 can indeed be used in devices and is not solely for host PC mainboards. So what about the Thunderbolt 4 JHL8540? I’m looking forward to that possibility.

References