DGX Spark USB ports are USB 4 (40Gbps) :O So why?

[Neurfer]

So First my question:

What is the official spec? Any recommand for a USB4/Thunderbolt4 storage/enclosure/NVMe?

Even though no USA site describes DGX Spark’s USB spec (except one in Sweden), they are USB 4 (40Gbps) as I found out.

Why does’t NVIDIA’s marketing dept make big deal of it? It is a big deal. Theoretically, that mean it can get R/W speed of ~2,800 MB/s to ~3,800 MB/s from a USB attached NVMe SSD. That is a big deal for AI engineers/Data Scientist on a machine like DGX Spark which is not upgradable.

But I got R/W 1.8GB on Samsung NVMe 990 EVO Plus 2TB PCIe 4 in a USB4 NVMe enclosure (Cable powered) and connected directly to one of USBs on DGX Spark. This great, but if I can double the speed, absolutely.

!! IMPORTANT !! If you get only 40M/s, flip the plug. YEP! That works. Also when they heats up and yes they sure do, it drops back to 40M/s. Then no flipping plug works.

---

This one is just FYI for those who wants to find out by themselves.

# Feed the ouput of below to Gemini/ChatGPT for human friendly info
$ sudo dmesg | grep -i "thunderbolt\|usb4"
...
[    0.054132] ACPI: USB4 _OSC: OS supports USB3+ DisplayPort+ PCIe+ XDomain+ 
...

# When it is slow, you will see it under 480M, when it is fast,  20000M/x2
$ lsusb -t
...
/:  Bus 005.Port 001: Dev 001, Class=root_hub, Driver=xhci-hcd/1p, 480M
/:  Bus 006.Port 001: Dev 001, Class=root_hub, Driver=xhci-hcd/1p, 20000M/x2
    |__ Port 001: Dev 005, If 0, Class=Mass Storage, Driver=uas, 20000M/x2
...

Gemini’s response to the output of dmesg above.

Here is what this new dmesg output means:

1. You DO Have a USB4 Controller: The lines ACPI: USB4 _OSC and usb usb4 are definitive proof. Your operating system is detecting and loading a USB4 host controller.
2. It’s Not an Intel Thunderbolt Controller: The lspci command was empty for “Thunderbolt” because your controller is from NVIDIA (as shown by Manufacturer: Linux 6.11.0-1016-nvidia xhci-hcd). This is why boltctl doesn’t see it; boltctl is primarily designed for Intel’s Thunderbolt controllers.
3. Your Problem Remains the Same: Your core issue is that your enclosure is connecting as /dev/sda (USB Mass Storage mode) and not as /dev/nvme0n1 (PCIe Tunneling mode). The “flipping the cable” test confirms this is a physical layer problem.

[maiia]

Thanks for this @Neurfer
It gives me peace of mind seeing your post this morning since I just ordered the kit below last night:
UGREEN 40Gbps M.2 NVMe SSD Enclosure with Dual Chips, External SSD Drive with Cooling Fan
SAMSUNG 990 PRO 2TB SSD NVMe M.2 PCIe Gen4, M.2 2280 SSD Seq. Read Speeds 7,450 MB/s

[Neurfer]

The reason I didn’t mentioned the brand of the enclosure here (I know it showed in the screenshot) is that it still back and force between 1.x G/s and 40M/s. I tested two brands and the other one is even worse result. If I can find an enclosure that supports PCIe 4 tunneling that would be best. If it does, the USB4 storage will show as /dev/nvmeXnY not /dev/sdZ as it does now.

That’s why I posted this to get some recommendation from NVIDIA team.

Otherwise I gonna have to use Amazon’s return policy to try out until I find one or get blocklisted whichever comes first.

Maybe that’s what we should do here. I will list ones I tested that didn’t work, and everyone test at least two other ones and post that result. We will find out soon enough which one works and CHEAPEST 😁

[eggman]

It is USBC 3.2x2. If you’re seeing PCIE support it may be that it is Thunderbolt [3] compatible but not Thunderbolt certified, though I have not tested.

[maiia]

Thank you both for sharing!
I added mine below:

Hopefully the enclosure arriving tomorrow wasn’t one of the ones that didn’t work else I’ll be returning mine too

About NVIDIA DGX Spark Reservation Countdown

They’ve finally arrived! I’m a 35### US reservation made 7 weeks ago. Grateful to have my bundle & while a bit of a hurdle can’t really complain. Also expecting a few additional support pieces for my set up arriving this week: TP-Link high-speed Ethernet Switch –> for direct Ethernet distribution) Obram Cat 8 high-speed Ethernet Cables –> to minimize any cable speed impact Furman PST-8 digital Power Station –> for voltage regulation and power surge protection UGreen 40Gbps SSD Enclosure wit…

[FlossingEnthusiast]

@Neurfer

Even though no USA site describes DGX Spark’s USB spec (except one in Sweden), they are USB 4 (40Gbps) as I found out.

They’re not USB 4. They’re USB 3.2 Gen2.

From the output of lsusb -tvv:

/: Bus 002.Port 001: Dev 001, Class=root_hub, Driver=xhci-hcd/1p, 20000M/x2
/: Bus 004.Port 001: Dev 001, Class=root_hub, Driver=xhci-hcd/1p, 20000M/x2
/: Bus 006.Port 001: Dev 001, Class=root_hub, Driver=xhci-hcd/1p, 20000M/x2
/: Bus 008.Port 001: Dev 001, Class=root_hub, Driver=xhci-hcd/1p, 20000M/x2
/: Bus 010.Port 001: Dev 001, Class=root_hub, Driver=xhci-hcd/1p, 20000M/x2
/: Bus 012.Port 001: Dev 001, Class=root_hub, Driver=xhci-hcd/1p, 20000M/x2

That’s 20Gbit/s over two lanes. Getting 1.8GB/s speeds is about right; theoretical max would be 2.5GB/sec if we removed all protocol overheard.

[Neurfer]

Fast SSD Enclosures on DGX Spark:

@maiia, thanks for the UGreen fan enclosure recommendation!

After drowning in NVMe benchmarks, here’s the brutally simple truth about getting max speed out of our 40Gbps enclosures on the DGX Spark:

1. 🛑 Stop Using JHL7440 (Thunderbolt 3)

• JHL7440 (e.g., in ACASIS 40Gbps enclosures) is the wrong chip. It’s only TB3 and it connects as a slow device, maxing out at 950 MB/s read.
• ASM2464PD (e.g., in Orico, UGreen) is the right chip (USB4/TB4 compatible). This one hits the fast speeds we want: 1.9 GB/s read.

2. 🔥 The Real Problem is the SSD

The chip isn’t the main enemy though; NVMe drive is too hot for its own good.

• Samsung 990 Evo Plus have hyper-aggressive thermal guards.
• They get hot instantly, throttle the speed down to 40 MB/s (slower than a free conference USB drive!), and become unusable.

I saw a speed drop from 1.9GB/s to 40M/s on a 100GB file copy. Unacceptable.

3. ✅ The Fix: Stable NVMe + Active Cooling

You need a drive that plays nice and an enclosure that does the work:

• Enclosure: ASM2464PD chip with a fan (passive cooling fails) like UGreen fan enclosure.
• SSD: Use a thermally stable drive like PNY, WD_Black, or Crucial. The 2230 size is great for ventilation.

---

@maiia: If you will, pls Try This on the UGreen Enclosure Samsung 990 pro to check if it has same problem

1. Connect the enclosure to the USB-C port by the HDMI.
2. Check lsusb -t. Make sure it shows up as 20000Mx2 on Bus 002.
3. Run a large copy benchmark 2GB repeatedly. It should stay at 1.9 GB/s for awhile (few mins).
4. Unplug, plug back in, or reboot. If it ever connects as a slow 480M device (Bus 001), it confirms the Samsung thermal guard kicked in and flipped the connection. Return the Samsung drive.

I’m ditching all the slow, hot components and keeping only the PNY NVMe. That ACASIS chassis with a fan was beautiful, but the JHL7440 chip was a total flop.

Good luck! Pls Let us know your results. So I can decide whether to buy UGreen or something else.

[Lucid]

Thanks for your research into this, just checked my Adata SE920 USB4 (with FAN) and it has the ASM2464PD, so its nice to know it should play well with my sparks.

Just getting them up and running, will follow your steps later

[maiia]

Thanks for sharing @Neurfer

I dove back into my Comet Perplexity Enterprise Max Research Assistant that originally recommended the Samsung 990 PRO and shared your feedback.

It confirmed even before setting up my DGX Spark Bundle and running a test (which I’m happy to do anyway) that heat will definitely throttle it’s speed even with the UGREEN enclosure and recommended I go with the WD Black as the best choice.

Drive	Interface	Price (2TB)	Native Speed	USB4 Enclosure Speed	Thermals in Enclosure	Value for External Use
WD Black SN850X	PCIe 4.0 x4	$158	7,300 MB/s	~3.0–3.3 GB/s	Excellent, cool	⭐⭐⭐⭐⭐ Best
Solidigm P44 Pro	PCIe 4.0 x4	$252	7,000 MB/s	~3.0–3.3 GB/s	Excellent, stable	⭐⭐⭐⭐ Premium reliability
Samsung 990 Pro	PCIe 4.0 x4	$179	7,450 MB/s	~1.9–3.0 GB/s*	Poor, throttles hard	⭐⭐ Risky
PNY CS3250	PCIe 5.0 x4	$220	14,900 MB/s	~3.0–3.3 GB/s (capped)	Unknown in externals	⭐⭐ Overkill, pricey
WD Black SN8100	PCIe 5.0 x4	$280	14,900 MB/s	~3.0–3.3 GB/s (capped)	Better than old Gen5, but still hotter than Gen4	⭐⭐ Overkill, expensive
Crucial T710	PCIe 5.0 x4	$252	14,900 MB/s	~3.0–3.3 GB/s (capped)	Good for Gen5, still needs cooling	⭐⭐ Overkill

When I challenged whether it was better to invest in the P44 Pro and hedge for long term efficiency for my use case. The recommendation is still the WD Black SN8 100 + UGREEN but for your use case you may want to consider the P44 Pro so I’m sharing the comparison here for reference.

Feature	WD Black SN850X	Solidigm P44 Pro	Winner for External USB4 Use
Price (2TB)	$158	$252	SN850X(37% cheaper)
Price (4TB)	$280	~$450–500 (rare)	SN850X(much cheaper)
Sequential Read	7,300 MB/s	7,000 MB/s	SN850X (slightly faster)
Sequential Write	6,600 MB/s	6,500 MB/s	SN850X (slightly faster)
Random Read IOPS	1,200K	1,400K	P44 Pro (better random reads)
Random Write IOPS	1,100K	1,300K	P44 Pro (better random writes)
Endurance (TBW) - 2TB	1,200 TBW	1,200 TBW	Tie
Endurance (TBW) - 4TB	2,400 TBW	Not available (2TB max)	SN850X (4TB exists)
MTBF	1.75 million hours	1.6 million hours	SN850X (slightly higher)
Power Efficiency	Good (5.5–6W active)	Excellent (5.3W active, best-in-class)	P44 Pro (more efficient)
Thermal Performance	Very good; runs cool, minimal throttle	Excellent; minimal throttle, optimized	P44 Pro (slightly better, but close)
USB4 Enclosure Speed	~2.8–3.3 GB/s sustained	~2.8–3.3 GB/s sustained	Tie(both max out enclosure)
Warranty	5 years	5 years	Tie
Availability (4TB)	Widely available	2TB max; no 4TB option	SN850X
Real-world reliability	Proven; widely used, strong reviews	Strong; newer, some SLC cache bugs reported	SN850X (more established track record)

Key differences and why they matter (or don’t) for your use case

1. Price: SN850X is 37% cheaper

• 2TB: SN850X $158 vs P44 Pro $252 = $94 savings

• 4TB: SN850X $280 vs P44 Pro (unavailable; 2TB max)

• Verdict: SN850X offers far better value, especially at 4TB.​

2. Sequential speeds: SN850X is slightly faster (but both max out USB4)

• SN850X: 7,300/6,600 MB/s native

• P44 Pro: 7,000/6,500 MB/s native

• In your USB4 enclosure, both are capped at ~2.8–3.3 GB/s sustained, so native speed differences don’t matter.​

• Verdict: Tie in real-world external use.

3. Random IOPS: P44 Pro is better

• P44 Pro: 1,400K/1,300K random read/write IOPS

• SN850X: 1,200K/1,100K

• Random IOPS matter for database workloads, VM hosting, and small-file operations, not for large sequential backups, system images, or dataset staging.​

• Verdict: P44 Pro wins.

4. Power efficiency: P44 Pro is slightly better (but not critical externally)

• P44 Pro draws 5.3W active; SN850X draws ~5.5–6W.​

• Power efficiency is more important for laptops (battery life) than for external enclosures, which are bus-powered and supply enough current.​

• The P44 Pro runs slightly cooler under sustained load, but the SN850X also has excellent thermals and doesn’t throttle in USB4 enclosures with fan cooling.​

• Verdict: P44 Pro has a slight edge, but not worth the premium for external use.

5. Sustained write performance: Both are excellent, P44 Pro slightly more consistent

• P44 Pro is praised for “consistent peak performance with minimal throttling” and excellent sustained writes.​

• SN850X also has very strong sustained writes and is widely used in gaming/workstation contexts without throttling complaints.​

• In USB4 enclosures with active cooling (like your UGREEN), both maintain 2.8–3.3 GB/s for long sequential writes.​

• Verdict: P44 Pro is marginally better, but SN850X is already excellent for your needs.

6. Endurance and warranty: Tie

• Both: 1,200 TBW (2TB), 5-year warranty, and >1.6 million hours MTBF.​

• SN850X has a slightly higher MTBF (1.75M vs 1.6M hours), but both are far more than sufficient.​

• Verdict: Tie; both are highly durable.

7. 4TB availability: SN850X only

• SN850X is available in 1TB, 2TB, and 4TB.​

• P44 Pro maxes out at 2TB; no 4TB option exists.​

• If you want 4TB for full system images + multiple snapshots, SN850X is your only Gen4 choice in this tier.​

• Verdict: SN850X wins if you need 4TB.

8. Real-world reliability and track record

• SN850X: Widely adopted since 2022; strong community feedback; proven in gaming, workstation, and external enclosure use.​

• P44 Pro: Launched late 2022; strong reviews but newer; some reports of SLC cache bugs causing write speeds to degrade after months of heavy use (firmware issue, likely fixed).​

• Verdict: SN850X has a more established, proven track record for reliability.

---

Why the P44 Pro’s advantages don’t justify the premium for you

The P44 Pro is an excellent drive, but its strengths are most valuable in internal desktop/workstation use where:

• Random IOPS matter (databases, VMs, small-file workloads).

• Power efficiency matters (laptops, always-on systems).

• Absolute lowest thermals are critical (passive cooling, tight M.2 slots).

For your external USB4 backup/restore/staging use case:

• You’re doing large sequential writes (system images, datasets), not random I/O → SN850X’s sequential speed advantage and already-excellent thermals are sufficient.

• The enclosure has active cooling → both drives will stay cool and sustain max speeds.

• USB4 caps both drives at ~3 GB/s → you don’t benefit from P44 Pro’s efficiency or random IOPS.

• You’re paying $94 extra (2TB) for marginal gains you won’t notice in practice.

---

Why SN850X is the right choice:

• 37% cheaper than P44 Pro at 2TB.​

• Slightly faster sequential speeds (though both max out USB4).​

• Proven thermal stability in external enclosures.​

• Widely available in 4TB for larger backups.​

• Established reliability track record since 2022.​

---

Bottom line

Return the Samsung 990 Pro and buy the WD Black SN850X 2TB or 4TB. The Solidigm P44 Pro is a fantastic drive, but its advantages—slightly better random IOPS and power efficiency—don’t matter for your external USB4 sequential backup/restore workload, and you’d be paying $94+ extra for performance you won’t see. The SN850X delivers the same real-world USB4 speeds, has better value, runs cool, and is proven reliable in external enclosures.​

Final cart:

• UGREEN 40Gbps USB4 NVMe enclosure (ASM2464PD, fan) – $125 ✅

• WD Black SN850X 2TB – $158 ✅ (or 4TB for $280 if you need more capacity)

• Total: $283 (2TB) or $405 (4TB)

You’ll have a rock-solid, fast, cool-running external backup solution for your DGX Sparks without overpaying.

---

I also looked at the PNY vs. the WD Black as well … sharing to clarify why I went with the pivot to WD vs. PNY which is looks pretty good too

Factor	WD Black SN850X	PNY XLR8 CS3140	Winner for USB4 External Use
Price (2TB)	$158	~$180–$200	SN850X(11–21% cheaper)
Price (4TB)	$280	~$350–$400	SN850X(20–30% cheaper)
Native Sequential Speed	7,300/6,600 MB/s	7,500/6,850 MB/s	CS3140 (slightly faster on paper)
USB4 Enclosure Speed	~2.8–3.3 GB/s sustained	~2.8–3.3 GB/s sustained	Tie(both max out USB4/TB4)
Controller	WD in-house (G2)	Phison PS5018-E18 (premium)	Tie (both top-tier Gen4)
NAND	BiCS5 TLC (Kioxia/WD partnership)	176L Micron TLC (older batches) or 112L Kioxia (newer batches)	SN850X (more consistent)
Thermals in USB4 Enclosure	Excellent; runs cool, no throttling	Good; runs hot (~67°C peak in tests), but doesn’t throttle significantly	SN850X(cooler, more headroom)
Power Draw	~5.5–6W active	~6–7W active (less efficient than SN850X)	SN850X(lower power = less heat)
Track Record in Externals	Proven since 2022; widely used in USB4/TB4 enclosures	Some USB4 use, but less documented than SN850X	SN850X(more established)
Real-World Performance	Consistently hits advertised speeds	Slightly below advertised in real-world tests (7,000 MB/s vs 7,500 claimed)	SN850X(more reliable)
Endurance (TBW) - 2TB	1,200 TBW	1,400 TBW	CS3140 (slightly better)
Endurance (TBW) - 4TB	2,400 TBW	3,000 TBW	CS3140 (better)
MTBF	1.75 million hours	2.0 million hours	CS3140 (slightly higher)
Warranty	5 years	5 years (with “warranty void if removed” heatsink issue)	SN850X(cleaner warranty policy)
Consistency	Same NAND/controller across all batches	NAND changed mid-production (Micron→Kioxia), performance varies by batch	SN850X(more consistent)
Value for USB4 External Use	Excellent	Good, but worse price/performance	SN850X(better value)

I’ll test the Samsung as well as the WD when it arrives and share outcomes when I get a chance in my setup process.

[x4080]

Hi, is DGX spark have USB4 or USB3.2 ? Anybody tested external ssd with it yet ?

[Neurfer]

TLDR;

1. USB ports: ~= USB 3.2 Gen 2x2 20Gbps
2. Max R/W: 2GB/s from USB 4/3 40Gbps enclosures with NVMe PCIe 4
3. Spec: USB 3.2 Gen 2 (10Gbps) <> USB 3.2 Gen 2x2 (20Gbps)
4. Enclosure Chips: JHL 7440 bad, ASM2464 good.
5. Real R/W: <40MB/s because of the heat.

Long answer:

1. NVIDIA’s own spec of DGX SPARK don’t show USB spec. But ASUS Ascent GX10’s datasheet does. It’s USB 3.2 Gen 2x2 20Gbps. Yes, it’s disappointing that this is a state-of-art and what a beautiful art it is!, then they put USB 3.2 not USB 4 NOT Thunderbolt 4 or 5 :( But also I can understand this could be a heat management issue. DGX Spark do get hot as is already. I can’t imagine how hot it would be if there are USB 4 and Thunderbolt 4+ chips in them.
2. Any USB 4/3 enclosure with ASM2464 chip with any decent NVMe 2TB delivered R/W 2GB/s. But not the ones with JHL 7440 chip, even though it says 40Gbps. It delivered R 980MB/s, W 650MB/s only.
3. Be very careful of the spec of the cases. To get maximum speed possible, it must say USB 3.2 Gen 2x2. If it say USB 3.2 Gen 2 only NO. If it say USB 4 only, NO.
4. Because I only tested on one case with JHL 7440, I can’t say for sure, but its R/W speed was less than half of all other cases with ASM2464 chip. One thing though, that case did no other case did: it stayed cool when it wasn’t being accessed actively. And never went down to 40MB/s throttled speed. Someone please verify if you will.
5. Sooo… none of above almost don’t matter because the thermal throttling meaning almost immediately within just few minutes from the DGX is turned on, it switched from the read speed of 2GB/s to the miserable <40MB/s. My cheap USB drive run 10x faster than that and never goes down to that low. It’s disgrace. The passive heat management case make no sense here either. But even all the cases with fan, the fans were too small. I am guessing because it has to be powered by the USB-C it can’t afford to draw too much amps.

So I came to the conclusion that what @maiia is trying to do using the external fans might be the only way, but go a step further. Have a brilliantly elegant way to manage the heat from USB enclosure that complements this beautiful DGX Spark’s design.

That’s an indoor bug trap in case anyone wondering. Killing two birds with… two bugs??, yea whatever. So far after hours being up and few GBs being copied, and it HASN’T fall down from the grace to <40MB/s hell.

• GRAUGEAR USB 4.0 PCIe Gen4x4 40Gbps: they have two versions: one with JHL 7440 and another ASM2464. Guess what I got.
• Crucial T500 1TB PCIe Gen4 NVMe M.2 SSD. I chose this over WD_BLACK and Samsungbecause of its cache helps in AI use-cases compared to gaming. In short, few big file copy vs lots of small files. Also 1TB generates less heat than bigger ones and I figured I will not have to deal with files larger than 100GB most of time. So instead of paying for 2TB one, just get few 1TBs with other cases.
• Indoor Insect Trap with quite (relatively speaking) fan.

BEST PART: You might get a whip of smell, see a tiny smoke, or even occasionally hear a scream of a little bug. You are welcome! I know, It’s a perfect revenge for all that itches from their bites while keeping your precious NVMe cool.

===== FYI

[maiia]

I’ve been having trouble getting the DGX Spark to see the ext disk as an NVMe not sda

I tried:

• Flipping the UGREEN cable
• Testing different USB ports
• Updating the firmware update
• Updating grub
• Tried updating BIOS but then neither one of the Sparks saw my keyboard so I can go into BIOS Advanced setting and see if it’s blocked by security.

I did reboot with my keyboard wired to see if it wifi to wired tripping it up and I tried both of my logi keyboards with no luck. I might need a basic no wifi wired keyboard if it’s being THAT picky.

NAME MAJ:MIN RM SIZE RO TYPE MOUNTPOINTS
loop0 7:0 0 4K 1 loop /snap/bare/5
loop1 7:1 0 68.9M 1 loop /snap/core22/2140
loop2 7:2 0 233.3M 1 loop /snap/firefox/7077
loop3 7:3 0 232.8M 1 loop /snap/firefox/7182
loop4 7:4 0 493.6M 1 loop /snap/gnome-42-2204/228
loop5 7:5 0 91.7M 1 loop /snap/gtk-common-themes/1535
loop6 7:6 0 44.3M 1 loop /snap/snapd/25585
loop7 7:7 0 218.9M 1 loop /snap/thunderbird/826
sda 8:0 0 1.8T 0 disk
nvme0n1 259:0 0 3.7T 0 disk
├─nvme0n1p1 259:1 0 298M 0 part /boot/efi
└─nvme0n1p2 259:2 0 3.7T 0 part /

/: Bus 001.Port 001: Dev 001, Class=root_hub, Driver=xhci-hcd/1p, 480M
/: Bus 002.Port 001: Dev 001, Class=root_hub, Driver=xhci-hcd/1p, 20000M/x2
|__ Port 001: Dev 002, If 0, Class=Mass Storage, Driver=uas, 10000M
/: Bus 003.Port 001: Dev 001, Class=root_hub, Driver=xhci-hcd/1p, 480M
/: Bus 004.Port 001: Dev 001, Class=root_hub, Driver=xhci-hcd/1p, 20000M/x2
/: Bus 005.Port 001: Dev 001, Class=root_hub, Driver=xhci-hcd/1p, 480M
/: Bus 006.Port 001: Dev 001, Class=root_hub, Driver=xhci-hcd/1p, 20000M/x2
/: Bus 007.Port 001: Dev 001, Class=root_hub, Driver=xhci-hcd/1p, 480M
/: Bus 008.Port 001: Dev 001, Class=root_hub, Driver=xhci-hcd/1p, 20000M/x2
/: Bus 009.Port 001: Dev 001, Class=root_hub, Driver=xhci-hcd/1p, 480M
/: Bus 010.Port 001: Dev 001, Class=root_hub, Driver=xhci-hcd/1p, 20000M/x2
/: Bus 011.Port 001: Dev 001, Class=root_hub, Driver=xhci-hcd/2p, 480M
|__ Port 002: Dev 002, If 0, Class=Wireless, Driver=btusb, 480M
|__ Port 002: Dev 002, If 1, Class=Wireless, Driver=btusb, 480M
|__ Port 002: Dev 002, If 2, Class=Wireless, Driver=btusb, 480M
/: Bus 012.Port 001: Dev 001, Class=root_hub, Driver=xhci-hcd/1p, 20000M/x2

The ran the benchmark on both the WD_Black and Samsung 990 PRO below using same parameters for comparison. Both are negotiating less than half of what’s expected.

[Neurfer]

How did you update BIOS other than one that was pushed at the very 1st installation?

I’ve been running the following code everyday, so far no change.

sudo apt update
sudo apt upgrade
sudo fwupdmgr refresh
sudo fwupdmgr upgrade

This is the BIOS version I see.

I used the wired (really wired NOT USB dongle one) keyboard whenever I need to recover or access BIOS. And instead of holding DEL key, I had to press repeatedly like in 2D shooter game to go into BIOS mode. Even then , sometimes I miss and had to do all over again. I re-imaged about 6 or 7 times so I am getting used to it!

Regarding UGREEN, There are two UGREEN 40Gbps. From your test result, it looks like you got the JHL based one which is really USB 3.2 Gen 2 NOT USB 3.2 Gen 2x2 NOT USB 4 from what AI told me. Also in my case, Disk utilities shows the name of the chip, not NVMe’s name.. I didn’t know that was possible. Did you do something to make that happen?

1. JHL 7440 based (R 1GB/s) -
   Amazon.com: UGREEN 40Gbps M.2 NVMe SSD Enclosure with Dual Chips, External SSD Drive with Cooling Fan, Compatible with Thunderbolt 4/3 USB4/3.2/2.0, Support 1/2/4/8TB M-Key/(M&B) Key 2280 Side SSD Enclsoure : Electronics
2. ASM based (R 2GB/s) -
   Amazon.com: UGREEN 40Gbps M.2 NVMe Enclosure with Cooling Fan, Anti-Drop SSD Enclosure for MacBook, Aluminum M.2 Enclosure Support M/B+M Key Size 2280/2260/2242/2230 Compatible with Thunderbolt 3/4/USB4/3.2/3.1 : Electronics

[maiia]

You’re right! I did get the wrong one! The 40Gbps UGREEN labeled MAC threw me off.

The device name visible on the Benchmark screen may be a UGREEN Firmware feature

I used the same code for the DGX firmware and I have the same BIOS info

I’ll order the right UGREEN or maybe yours since it’s “proven” might be the safer bet.

There’s also an interesting open fan option GRAUGEAR had with the ASM2464 chip.

Amazon.com: GRAUGEAR 40Gbps M.2 NVMe SSD Enclosure, USB4 NVMe SSD Enclosure with...

Amazon.com: GRAUGEAR 40Gbps M.2 NVMe SSD Enclosure Docking Station, USB4...

[Neurfer]

@maiia

Amazon.com: GRAUGEAR 40Gbps M.2 NVMe SSD Enclosure, USB4 NVMe SSD Enclosure with Cooling Fan and Dual-Pipe Heatsink Aluminum, Thunderbolt 4/3, ASM2464PD Chip : Electronics

I am sad to report that it throttled last night. Then when I wake up even though it was absolutely ice cold, it started on 40MB/s. I had to flip the USB plug, what the hack! I don’t know anything any more.

I am now seriously thinking about getting the Docking Station one and get a heatsink attached to NVMe.

FYI, as long as it uses ASM chip, you can get the temperature reading from its SMART data.

sudo smartctl -d sntasmedia -A /dev/sda

[maiia]

Our problem is the DGX Spark potentially heat throttling impacting SSD performance.

I was looking into workarounds in this thread but realized it’s a build flaw that won’t be fixed till the next version which might be larger in size to accommodate for some of the issues and increase RAM. (e.g. MAC mini vs. MAC Studio).

Till then cleaver community workarounds may be our go to for now unless NVIDIA push a firmware fix.

I ordered ASM2464 PD UGREEN and the GRAUGEAR with open fan docking station to test both out on the 4TB WD_Black and the 2TB Samsung PRO 990 and returning the wrong one.

DGX Spark Bundle Cooling Makeshift Rack

Update: I was hoping something like this would work, but seeing that there’s barely any drop in temp makes me think the DGX Spark casing is the problem, trapping the heat in and making cooling difficult, which makes me think that the rack wouldn’t work, at least not as expected. Another Update: TECs are not a good idea even near, not on under or touching, the Spark. The temp differential was about to start causing condensation. [image] [image] [image]

[maiia]

Another thought

If this is a casing issue I wonder if any of the partners fixed it with their casing design

Also, looks like the lower SSD versions in ASUS, HP, and DELL with their modified circulation are performing better for obvious SSD size reasons but even the 4TB ASUS appears to be handling heat better with its 5 heat pipes + dual fans + 57 fins + producing 1.6× better cooling. HP claims +35% cooling with its Molecular beam heatsink and DELL also claim +35% cooling, and +2°C better results, with their 3 fans.

https://marketplace.nvidia.com/en-us/enterprise/personal-ai-supercomputers/

[eugr]

@maiia

I’ve been having trouble getting the DGX Spark to see the ext disk as an NVMe not sda

I don’t think you will, since DGX Spark has USB 3.2gen2x2 ports and doesn’t support Thunderbolt /USB4 (which is required for PCIe over USB). Even if it’s actually USB4, DGX OS uses USB3.2 drivers.

But given that I’m currently running the newest 6.17.1 kernel that NVIDIA has in their repo, with USB4 enabled, and it still sees it as USB 3.2 20Gbps, means one of the following: 1) it is indeed USB 3.2, or Mediatek Linux drivers do not support USB4 on this SOC yet.

[Neurfer]

OMG, WTH! I just started DGX Spark after being off for hours. As you can see in the my sensors reading everything is cool and it even reads the temperature on the USB enclosure as 30C, and yet, it started on read speed 40MB/s. Turning the power off thru “Disk utility” or rebooting DGX don’t makes it go back to 2GB/s. But unplug and replugging the cable makes it go back to 2GB/s. Go figure.

No AI can figure this one out. I am giving up on the USB enclosure and focus on AI stuff. I am sure NVIDIA folks will figure this one out for us sooner than later. HINT HINT!

Returning everything except the bug trap. It works! Just not on NVMes but on bugs.

[maiia]

Thanks, @eugr Perplexity pulled a few resources that mentioned the hardware is indeed USB 3.2 and I was wrong to think it was USB 4 or that it’ll support USB 4/ Thunderbolt.

Is the kernel update to the newest 6.17.1 worth it? Are you finding any improvements or experiencing any instability?