Sparkle Radeon RX6600 Thunderbolt eGPU review and teardown

Introduction

Several months ago, my Thunderbolt 4 docking station article caught the attention of a few folks at Sparkle Computer based in Taiwan. As I mentioned in my last Sparkle article, this company has an interesting history. When I worked at Microsoft in the DirectX graphics division, we always got the very latest ATI/AMD and NVIDIA GPUs to test and I remembered the Sparkle name for their NVIDIA GPUs. I hadn’t heard about Sparkle in over 10 years and just assumed they got absorbed by some other Taiwan brand like Asus or MSI. But they are still around and owned by TUL Corp which also owns PowerColor, another GPU add-in-board brand. I was excited to learn that not only does Sparkle make Thunderbolt docking stations, they also make compact Thunderbolt external GPUs (eGPU) based on recent AMD and NVIDIA chips. Their eGPUs look similar to the older Sonnet Breakaway Puck models from 2020. In this article, I’ll be reviewing an engineering sample of the Radeon Mini GameStation TCX-230DA (RX6600) that Sparkle provided.

Full disclosure: While this is not a sponsored post, the product(s) reviewed were provided at no cost for evaluation purposes. Products received in this capacity are destined for teardowns, future device interoperability testing, and/or charitable donations.

Feature summary

• AMD Radeon RX 6600 (Navi 23 XL) GPU with 8GB GDDR6 VRAM
• 4x DisplayPort 1.4a
• 2x USB 3.2 A-style ports (10Gb/s / 4.5W)
• 1Gb/s Ethernet
• Thunderbolt 3 host port with 45W PD charging
• 230W PSU (19.5V @ 11.79A) with detachable IEC13 AC/Mains cable

Unboxing & Physical Characteristics

The packaging is simple with an outer cardboard box adorned with a sticker indicating the specific GPU model – in my case the Radeon RX 6600. The eGPU can also ship with NVIDIA GPUs such as a GTX 1660, Quadro T1000, or Quadro RTX3000. The inner packaging includes an unbranded cardboard box for the PSU, a setup guide, and foam blocks.

Separating the two foam blocks revealed the eGPU with a black anodized aluminum construction and 4 triangular rubber feet on the bottom. On top is a large steel mesh air intake with a 92mm fan underneath. All four sides have additional mesh panels to provide an exit path for hot air. At ~155mm x ~132mm x ~55mm, the volume of the eGPU is only 1.125 liters which I think is the smallest on the market. The PSU is quite bulky, however, and adds another 0.5 liters. It is by FSP group and is model FSP230-AJAN3 with a 4-conductor DIN connector.

BoxSimple packagingeGPU and PSU812g772g155mm wide132mm deep55mm thick177mm wide

85mm deep

33mm thick

Feature testing

After plugging the eGPU into my Surface Laptop Studio Core i5 and installing the latest AMD Radeon drivers, I was off to the races. The GPU was recognized immediately and I was able to use all 4 DisplayPort connectors in addition to the two USB ports and Ethernet port. Any two of the DisplayPort connectors can be used in dual-mode to run HDMI or DVI monitors via cheap “passive” adapters, but to use 4 HDMI/DVI monitors you need more expensive “active” adapters. All the connections are easy to use but I would prefer to have the two USB connectors on the rear next to the DisplayPort connectors to simplify cable routing.

[Surface Laptop Studio with Sparkle RX 6600 eGPU and vertical dock]

Laptop charging is limited to 45W, so my Laptop Studio complained about slow charging with a little yellow flag over the battery icon since it expects 60W+. Under load, the Laptop Studio uses less than 45W when its built-in GPU is unused so there is little risk of battery drain. But charging from a depleted battery state is slow compared to the original Surface charger.

I tried plugging and unplugging the eGPU during OS bootup, after boot, during laptop sleep, and with the laptop fully shutdown and each time everything came alive and just worked the first time.

I have an RX 6600 in my desktop PC and performance in games and 3D applications wasn’t quite as good as on desktop, but close enough and easily 3-5X faster than the built-in Intel Xe graphics.

Under default settings, the whole rig sipped power – rarely did the PSU draw more than ~150 watts – quite a bit of headroom with the 230W PSU. So I tried to push it further…

Performance benchmarking

For performance testing, I used the latest production Windows 11 OS, latest Surface firmware, and latest AMD GPU drivers. Out of the box, performance was similar to an underclocked RX 6600 desktop model, a well-performing RX 6600M notebook, an NVIDIA GTX 1660Ti, or an underclocked RTX 3060. Power tuning from the factory was quite conservative at only 75W TDP. According to Sparkle, running at up to 100W is OK, but behavior is undefined beyond that. (In theory, the MXM connector is capable of up to 220W with 20V @ 10A on the main power rail with 5V @ 2.5A and 3.3V @ 2A on supplemental rails – but realistically the form factor can’t support more than 150W.)

[Default 75W power limit in MorePowerTool]

[Default fan PWM settings]

The fan tuning is a little odd with zero RPM enabled all the way to the hysteresis range of 55-70°C. Then the fan kicks in at 30% PWM, or 950 rpm. The fan is capable of running at 15% PWM for ~350RPM, so you can optimize the fan curve quite a bit to maximize performance and minimize fan noise. At 100% PWM, the fan maxes out at ~3700RPM, not the 6000 listed. It is loud at full blast!

[Optimized settings in Radeon app]

Setting a modest undervolt to 1100mV, raising VRAM clock to 1900MHz and adjusting the fan curve brings more performance but uses less or the same power.

AMD Radeon App & MorePowerTool Tuning	TimeSpy Scores	FurMark Scores (1080p)	Performance increase
Default – VRAM: 1750MHz – GPU: 500-2044MHz – Voltage: 1150mV – Power limit: 75W	6313 (6225 GPU / 4987 CPU) 1948MHz @66°C	5306 score @75°C	–
AMD Radeon App Tuning – VRAM: 1900MHz – GPU: 2200-2700MHz – Voltage: 1100mV (-50mV) – Power limit: +20% / 90W – manual fan curve	6852 (7338 GPU / 4985 CPU) 2246MHz @69°C	5862 score @70°C	10-18%
AMD Radeon App Tuning with MorePowerTool tweak same settings as above except – Power limit: +60% / 120W via morepowertool	7296 (7908 GPU / 5074 CPU) 2539MHz @71°C	7715 score @79°	30-45%

Sparkle TCX-230DA (RX6600) GPU performance scaling

[Timespy performance at 75W]

[Timespy performance at 120W]

For this particular board, I reckon the 90W setting with an optimized fan curve is probably the best bet since GPU hotspot temperatures rise quite a bit beyond that power level and the fan gets loud. But when running at 120W, it rivals some of the low-end RX 6600 desktop boards.

Game testing

Teardown

Teardown of the eGPU was simple. First, I carefully pulled off the 4 rubber feet from the bottom and placed the sticky sides on wax paper so I could re-use them. I loosened 4 Philips screws which held the bottom and top chassis halves together. Then I jiggled the bottom back and forth to get it free. The GPU was attached to the bottom plate and it seemed the copper heat pipes were jamming against the side of the chassis when trying to pull everything apart. I finally got it apart but had to un-plug an LED indicator wire to fully separate the two halves.

[Disassembly with LED cable still attached to the chassis]

[Bottom plate and LED cable separated]

Inside is an MXM form factor GPU plugged into a Thunderbolt host controller board. I loosened two more Philips (+) screws to free the MXM board and carefully popped it out. I didn’t have experience with MXM, but it worked similar to an M.2 SSD connector – just ~5x wider.

[Bottom of MXM RX6600 PCB showing compact tantalum capacitors for power filtering]

4 spring-loaded screws held the heatsink assembly to the GPU PCB. The heatsink is wonderfully compact and manages to squeeze 3x 5mm heatpipes and a copper slug for efficient heat transfer away from the GPU die. The main heatsink body and fins are aluminum. You can see grey thermal pads adhered to the heatsink for the 4x GDDR6 memory chips and various power delivery chips.

[GPU Heatsink with thermal pads]

Three more screws held the fan to the heatsink assembly. This is a standard low-profile 92mm fan found on many desktop GPUs. Rather than plugging into the GPU board, the cooling fan plugs directly into the Thunderbolt host controller board. The MXM interface pin# 24 TH_PWM allows the GPU to tell the host controller board how to manage fan speed.

[FirstD FD9015U12D (12V @ 0.55A)]

The GPU board is labelled EM23G V1.1. EM probably means “embedded” while 23 probably corresponds to AMD Navi 23. Full analysis below.

[RX 6600 board model EM23G V1.1 in MXM 3.1 “B” form factor]

The Thunderbolt host controller board is less interesting with a design we’ve seen in other Thunderbolt docks on egpu.io. There are two Intel JHL controllers, an ASMedia USB controller, and an Intel Ethernet controller. Most of the board real-estate seems to be related to power delivery for the MXM connector and laptop charging with larger inductors and capacitors spread out. With 4 DisplayPort receptacles, there isn’t much room for much else and the two USB ports are forced over to the side of the PCB. I think it would have been better if the DisplayPort receptacles were rotated 90° and stacked rather than arranged in a row- this would provide enough room for the USB ports to be at the rear.

[JHL6540DA V2.0 PCB front]

[PCB rear]

For this engineering sample, the bottom of the PCB had a defect highlighted by Sparkle QC with a red arrow-shaped piece of tape.

[Dislodged capacitor on engineering sample PCB]

PCB analysis

The Thunderbolt host controller board has the following components:

• Thunderbolt 3 receptacle with 45W power delivery

• various mosfets and regulators for MXM power delivery
• 4-pin DIN DC power input receptacle
• 4-pin PWM fan header
• 2-pin status LED header
• 1x 2-pin, 2x 3-pin, and 2x 5-pin mystery headers

According to Intel’s guidance on Thunderbolt eGFX designs, USB hub chipsets are should not be used – presumably for performance and reliability reasons. So an eGFX box with 2 or more USB ports needs to have a PCIe->USB host controller and a secondary Thunderbolt controller to supply the necessary PCIe lanes. I find this requirement a little silly since folks that need more USB ports will just plug their own hub in anyway. Also, for products marketed as general PCIe enclosures, there is no such requirement from Intel.

The GPU was more interesting and had the following major components:

• AMD Navi 23 XL (RX 6600 desktop GPU)
• 4x D9ZPP MT61K512M32KPA 16Gb Micron GDDR6 memory modules (8GB total)
• Power delivery components
  • 2x OnSemi NCP81022 4+1 voltage controllers
  • 10x OnSemi NCP302155 55A integrated power stages (so 10 phases available)
  • 4x Onsemi NCP81162 current balancing phase doublers
  • 1x OnSemi NCP59744 linear regulator
  • 2x OnSemi NCP3235 buck converters
  • 6x 0.15mH inductors (for vCore power phases)
  • 4x 0.33mH inductors (for GDDR6 / SoC power phases)
  • 2x 1.5mH inductors
  • 25x Tantalum capacitors
• MXM connector

OnSemi NCP302155

OnSemi NCP81022

Micron MT61K512M32KPA-16 GDDR6

OnSemi NCP3235

OnSemi NCP3235

OnSemi NCP302155

4x OnSemi NCP81162 and 1x NCP59744

It is noteworthy that two separate NCP81022 (4+1) voltage controllers are used. I believe this is because none of the heat generating components are allowed on the underside of the PCB so size and available layer count probably necessitates distributing the voltage controllers closer to the power stages. Based on the presence of 4x NCP81162 phase-doublers, it appears only 6 of the available 10 discrete phases from the voltage controllers are actually in use. With the compact PCB, it is difficult to follow signal traces, but it looks like 3 of these phases are for the GPU vCore (doubled to 6 phases), 1 phase is for SoC (doubled to 2 phases) and the last 2 phases are for GDDR6 and VDDCI. I’ve outlined my guess below ala Buildzoid-style.

[Power delivery on EM23G V1.1]

Critique & Conclusion

The Sparkle TCX-230DA provides a good amount of GPU performance in a small form factor. It is basically a modernized version of the Sonnet Breakaway Puck RX5700 which offers a little extra performance, better power efficiency, and support for hardware accelerated ray-tracing. It has 4 built-in DisplayPort 1.4 outputs which is convenient for professional settings but only has 45W laptop charging. The older Sonnet model provided 60W laptop charging power but only 2 built-in video outputs. In my testing, I found that 45W charging was OK since the laptop’s built-in GPU is mostly idle once connected, so it is rare that more than 45W would ever be needed just for the system. Still it would have been better for Sparkle to provide 60W+. The PCB already uses the TI TPS65983 PD controller which is 100W capable so I’m not sure why power is limited to 45W especially with the massive headroom provided by the 230W PSU. I suspect a firmware update could augment power to reach at least 60W but I don’t know how to make this change.

I also would have liked more built-in USB ports and some ports located on the same side as the DisplayPort receptacles. Rather than the ASMedia 2-port PCIe->USB controller, perhaps a 4-port controller model like the Via VL805 or TI TUSB7340 could be used? It seems the Intel Thunderbolt certification program steers vendors towards the higher tier ASMedia chipsets even if they have fewer ports than the competition.

I don’t know the retail price of these units, but if the Sonnet compact eGPU prices are anything to go by, these will easily be in the $500-$999 (USD) price tier depending on the specific GPU model inside. This is definitely a niche market given you can get a budget gaming laptop or a mid-tier desktop PC for that much. But if you need a compact form factor eGPU, this is by far the most performance you can get in a small package today.

Additional notes

The unit I had died after I completed the bulk of the review but before I could do extensive game testing. It was an engineering sample and had the pcb defect noted above so that could have been the issue. So I sent the unit back to Sparkle. When I get an updated production unit, I will update the post with more information.

In the meantime, I saved the VBIOS and GPU-Z “validation” results here:

Although the VBIOS is developed by TUL and includes this text string “TUL_NAVI23_EM23GV111027_GDDR6_8GB”, the PCI ID is PCI\VEN_1002&DEV_73FF&SUBSYS_24131787. 1787 corresponds to the Hightech (HIS) board manufacturer for which the last publicly announced product was the Radeon 5700XT 3 years ago. Perhaps Hightech is a contract manufacturer for other brands now?

References