OCuLink vs Thunderbolt 5 vs 4: eGPU Speeds for Local AI

Published April 23, 2026 · Updated August 23, 2026 · 12 min read

Short answer: OCuLink and Thunderbolt 5 land in the same place - both give an external GPU a PCIe 4.0 x4 link, roughly 7.9 GB/s of PCIe payload. Thunderbolt 4 and 40 Gbps USB4 cap at PCIe 3.0 x4, roughly 3.9 GB/s, so OCuLink is about twice a TB4 link and level with a TB5 one. The difference that actually decides your build is not bandwidth: OCuLink carries raw PCIe lanes with no controller in the path, so it is cheap and low-latency but not hot-pluggable and needs a spare M.2 or PCIe slot. Thunderbolt tunnels PCIe through a controller, so it costs more and adds latency but works on a port your laptop already has.

And for local AI specifically, the link is only in the way twice: while weights load into VRAM, and when you split one model across two GPUs. Everything else happens on-card.

Table of Contents

1. eGPU link speeds at a glance
2. Where those GB/s numbers come from
3. Is there a Thunderbolt to OCuLink adapter?
4. Does the extra bandwidth change tokens per second?
5. What OCuLink costs you in convenience
6. Which link should you pick?
7. What breaks in practice
8. Common questions

---

eGPU link speeds at a glance

Every figure in the "PCIe payload" column is computed from the PCIe specification, not measured. The arithmetic is shown in the next section so you can check it.

Three things fall out of that table that people get wrong constantly:

1. OCuLink has no speed of its own. It is a connector and a cable spec (SFF-8611). Whatever PCIe generation and lane count the host slot and the adapter negotiate is the speed you get. An OCuLink cable hanging off a Gen3 x4 M.2 slot is a Gen3 x4 link, no faster than Thunderbolt 4.
2. Thunderbolt 5 is not slower than OCuLink. Both are PCIe 4.0 x4. If somebody tells you OCuLink beats TB5 on bandwidth, they are comparing a TB5 port against nothing.
3. The 40/80 Gbps headline is not the PCIe number. On Thunderbolt and USB4 that budget is the whole link, shared with DisplayPort streams and USB traffic. The PCIe tunnel is a slice of it. OCuLink's number is the PCIe link, because there is nothing else on the cable - which is why 64 Gbps of OCuLink and 80 Gbps of Thunderbolt 5 arrive at the same PCIe payload.

Primary sources for the link rates: Intel's Thunderbolt overview and the USB-IF USB4 page, which is also where the "PCIe tunnelling is optional in USB4" caveat comes from.

---

Where those GB/s numbers come from

No benchmark needed - PCIe payload rate is arithmetic. Gen 3, 4 and 5 all use 128b/130b line coding, so 2 bits in every 130 are overhead:

Per-lane payload = transfer rate x (128 / 130) / 8 bits

PCIe 3.0:  8 GT/s  x 128/130 =  7.88 Gbit/s = 0.98 GB/s per lane
PCIe 4.0: 16 GT/s  x 128/130 = 15.75 Gbit/s = 1.97 GB/s per lane
PCIe 5.0: 32 GT/s  x 128/130 = 31.51 Gbit/s = 3.94 GB/s per lane

x4  Gen3 = 3.94 GB/s     <- Thunderbolt 3, Thunderbolt 4, USB4 40Gbps
x4  Gen4 = 7.88 GB/s     <- Thunderbolt 5, USB4 v2, OCuLink x4
x8  Gen4 = 15.75 GB/s    <- OCuLink x8
x16 Gen4 = 31.5 GB/s     <- internal slot, reference
x16 Gen5 = 63.0 GB/s     <- internal slot, reference

Real transfers never hit the theoretical payload rate - TLP headers, flow control credits, and the host's own memory subsystem all take a cut. Treat these as ceilings. The ratios between the links, which is what you are actually shopping on, hold regardless.

---

Is there a Thunderbolt to OCuLink adapter?

This is the question the comparison articles skip, so here is the direct answer: a passive Thunderbolt-to-OCuLink cable cannot exist, and neither can an OCuLink-to-Thunderbolt-5 one. They are not two flavours of the same signal.

• OCuLink is raw PCIe. SFF-8611 carries PCIe differential pairs plus sideband signals straight from the host's root complex to the device. No encoding change, no packet re-framing, no chip.
• Thunderbolt and USB4 are tunnelling protocols. PCIe is encapsulated inside Thunderbolt transport, routed over a packet-switched fabric, and un-tunnelled at the far end. Both ends need a Thunderbolt/USB4 controller to do that work.

So converting between them is not adapting a connector - it is a full protocol bridge. Anything sold as a "Thunderbolt to OCuLink adapter" has to contain a Thunderbolt device controller, which means:

• it costs roughly what a Thunderbolt eGPU enclosure costs, because it is one, minus the box and the PSU;
• it inherits the Thunderbolt PCIe ceiling - PCIe 3.0 x4 behind a TB4 host, PCIe 4.0 x4 behind a TB5 host. You do not get OCuLink speed by putting an OCuLink connector on the end of a Thunderbolt link;
• you gain nothing over a normal enclosure except cable routing, which is why almost nobody makes them.

Going the other direction - plugging an OCuLink dock into a Thunderbolt 5 port - is worse. That would need a Thunderbolt host controller to re-tunnel raw PCIe, and no such product category exists.

What actually exists, and what to buy

The cheap OCuLink adapters are all on the host side - they take PCIe lanes you already have and route them out of the chassis:

The practical rule: match the dock to the port you already have. There is no cheap converter that turns one into the other, and a machine with no spare M.2 or PCIe slot is a Thunderbolt machine whether you like it or not.

---

Does the extra bandwidth change tokens per second?

Mostly no, and the reason is worth understanding before you spend money on the faster link.

NVIDIA rates the RTX 4090 at 1,008 GB/s of on-card memory bandwidth. A PCIe 3.0 x4 host link runs at 3.94 GB/s. That is 3.94 ÷ 1008 = 0.4% of the GPU's own memory bandwidth. Once every weight is resident in VRAM, generating one token means the GPU reads gigabytes from its own memory and the host link carries a few kilobytes - your prompt in, your tokens out. The cable is not in the loop.

This is exactly why gaming eGPU benchmarks mislead people about AI. A game pushes whole frames and textures across the link on a 16 ms budget; that workload is link-bound. Token generation with a resident model is not.

The two places the link genuinely matters

1. Loading and swapping models. Weights cross the link once, and how long that takes is pure division. A Q4_K_M quant averages about 4.5 bits per weight, so:

8B model:  8e9  x 4.5 bits / 8 = ~4.5 GB of weights
70B model: 70e9 x 4.5 bits / 8 = ~39 GB of weights

These are arithmetic floors - weight bytes divided by link payload rate at 100% efficiency. Your real load times will be longer, because disk reads, page cache misses and VRAM allocation all add to it. But the gap between rows is the part you are buying, and on a 70B model it is roughly five seconds per swap.

Load one model in the morning and chat all day, and you pay that once - it is noise. Run a pipeline that hot-swaps an embedder, a chat model and a reranker per request, and it is the whole ballgame. That is the real dividing line, not tokens per second.

2. Splitting one model across two GPUs. Tensor-parallel inference exchanges activations between cards every layer. Do that through a shared, tunnelled x4 link and it falls apart. If multi-GPU is the plan, you want lanes per card - OCuLink x8, or a desktop board with real slots. Our eGPU local AI benchmarks cover the multi-card side in more depth.

What about partial offload?

When a model does not fit in VRAM and layers get split between CPU and GPU, people assume the cable becomes the bottleneck. It usually does not. What crosses the boundary each token is the hidden state - kilobytes, not weights, which stay put on whichever side owns them. The thing that collapses your throughput is CPU memory bandwidth doing the CPU-side layers, not the PCIe link. The fix is a card with enough VRAM, not a faster cable: see how much VRAM local models actually need.

Latency

OCuLink has a genuine structural advantage here - no packetisation, no protocol switching hops, just PCIe. Thunderbolt and USB4 add both. It is real, and it is the reason OCuLink feels snappier on rapid-fire short requests. We do not publish a millisecond figure because we have not measured one on this hardware, and for a streaming token response GPU compute time dwarfs it anyway.

---

What OCuLink costs you in convenience

Bandwidth is the tie. The tie-breakers are all physical:

• Hot-plug. Thunderbolt and USB4 are designed for surprise removal; the stack expects devices to come and go. OCuLink is a bare PCIe link with no such contract. Pulling the cable with the GPU active will typically hang or crash the host. If the machine ever leaves the desk, this matters more than any GB/s figure.
• Power. Thunderbolt enclosures usually bundle a PSU sized for the box. OCuLink docks generally do not - you supply an ATX or SFX PSU and the cabling. Cheaper, uglier, more of your Saturday.
• Slots. Every OCuLink route consumes an M.2 or PCIe slot you might have wanted for storage. Thunderbolt costs you a port you already had spare.
• Cost. An M.2-to-OCuLink adapter plus cable is dramatically cheaper than a Thunderbolt eGPU enclosure - typically a small fraction of it - because you are buying a connector and a cable rather than a controller chip, a chassis and a power supply. Prices move constantly, so check current listings rather than trusting a number in an article.
• Availability. Thunderbolt is on almost every modern laptop. OCuLink is on very few, and retrofitting it means opening the machine.

---

Which link should you pick?

If you have not chosen the GPU yet, the card matters far more than the cable - start with the budget local AI machine guide or, for the laptop side of this decision, laptop GPU VRAM for local AI.

---

What breaks in practice

1. Not every USB-C port tunnels PCIe. PCIe tunnelling is optional in the USB4 specification and mandatory only for Thunderbolt certification. A 40 Gbps USB4 port with no PCIe tunnel will charge your laptop and drive a monitor and will never see an eGPU. Check for the Thunderbolt lightning mark or the vendor's explicit "USB4 with PCIe tunnelling" wording before buying an enclosure.

2. The link trains at Gen3 when you expected Gen4. Long OCuLink cables and cheap M.2 adapters frequently negotiate down for signal-integrity reasons, which silently halves your bandwidth. Verify:

lspci -vvv | grep -i LnkSta

Speed 16GT/s is Gen4. Speed 8GT/s is Gen3 - try a shorter cable.

3. OCuLink hot-unplug. It is electrically PCIe. Shut down or detach the device through the kernel first; do not yank it.

4. Mixed Thunderbolt generations run at the slower end. A TB5 host with a TB4 enclosure is a TB4 link, PCIe 3.0 x4. Both ends have to be TB5 to get PCIe 4.0 x4.

5. Hybrid-GPU laptops pick the wrong card. With an internal NVIDIA dGPU plus an external one, runtimes sometimes bind to the internal card. Set CUDA_VISIBLE_DEVICES explicitly and confirm with nvidia-smi which device the model actually landed on.

6. Suspend and resume. GPU passthrough across S3 sleep is unreliable with eGPUs on Linux. Disable suspend on a host that drives an eGPU full-time.

7. Enclosure PSU headroom. Older Thunderbolt enclosures were specified for cards that drew far less than a current flagship. Check the PSU rating against the card's board power, or cap the card with nvidia-smi -pl.

---

Common questions

Is OCuLink faster than Thunderbolt 4?

For PCIe bandwidth, yes - about twice as fast, assuming your OCuLink link trains at PCIe 4.0 x4 (7.88 GB/s) against Thunderbolt 4's PCIe 3.0 x4 (3.94 GB/s). Whether you can feel that depends entirely on how often you load models.

Is OCuLink faster than Thunderbolt 5?

No. Both are PCIe 4.0 x4, so the payload ceiling is the same ~7.88 GB/s. OCuLink wins on price and on latency, because there is no controller tunnelling the traffic. Thunderbolt 5 wins on hot-plug and on working with the port your laptop already has.

What is OCuLink's speed?

It does not have one. OCuLink is a connector and cable standard that carries whatever PCIe link the host provides. Off a PCIe 4.0 x4 source it is ~7.88 GB/s; off a Gen3 x4 M.2 slot it is ~3.94 GB/s; x8 variants double it again. Always check what your host slot actually offers before assuming the higher number.

Is there a Thunderbolt to OCuLink adapter?

Not as a passive cable - the two are different things electrically, and bridging them requires a Thunderbolt controller. Anything sold as one is effectively an eGPU enclosure with an OCuLink connector on it, and it is still capped at the Thunderbolt PCIe tunnel. Full explanation above.

Can I run a 70B model over Thunderbolt 4?

Yes, provided the card has the VRAM to hold it. Once loaded, the link is nearly idle. The cost you pay is load time - roughly ten seconds of pure transfer floor for a ~39 GB Q4 quant on PCIe 3.0 x4, versus about five on a PCIe 4.0 x4 link.

Does USB4 always work for an eGPU?

No. PCIe tunnelling is optional in USB4. Thunderbolt-certified ports must have it; generic USB4 ports may not. This is the single most common reason a new eGPU enclosure does nothing when plugged in.

Does OCuLink work with any laptop?

No. You need a native OCuLink port or a free internal M.2 / PCIe slot to adapt out of. Most thin-and-lights have neither, which makes Thunderbolt the only route regardless of cost.

Can I daisy-chain enclosures for multi-GPU?

You can chain Thunderbolt devices, but they share one PCIe tunnel, so tensor-parallel inference across them performs badly. For multi-GPU, put lanes behind each card - OCuLink x8 adapters or a desktop board with real slots.

---

Bottom line

The bandwidth argument is smaller than the internet makes it. OCuLink and Thunderbolt 5 both deliver PCIe 4.0 x4; Thunderbolt 4 and 40 Gbps USB4 deliver half of that; and for a model that fits in VRAM, none of it touches your tokens per second because the GPU is reading from its own memory at over a hundred times the speed of any of these cables.

Pick on the physical constraints instead. Free M.2 slot and the machine never moves? OCuLink, cheaply. Laptop that gets unplugged daily, or no spare slot? Thunderbolt, and do not feel bad about it - the only number you are giving up is model load time.