External GPUs were always second best. CopprLink may change that

[Skye Jacobs]

First look: External GPUs have always existed as the lesser alternative to regular GPUs installed as internal cards, trading raw performance for portability or flexibility. That tradeoff may finally be starting to break down. New benchmarks suggest PCI-SIG's CopprLink standard is approaching what once seemed out of reach: external GPU performance that comes remarkably close to a direct motherboard connection.

Developed by PCI-SIG for high-bandwidth enterprise use, CopprLink provides a full PCIe 5.0 x16 interface with throughput of 32 GT/s per lane, or 64 GB/s of total bandwidth. That is four times the throughput of an OCuLink connection using PCIe 4.0 x8, one of the fastest eGPU options currently available alongside Thunderbolt.

The implications could be significant, especially for high-end workstations and AI servers that need a more flexible way to connect external graphics hardware.

PCWorld's test platform paired Nvidia's Founders Edition RTX 5090 with HighPoint's RocketStor 8631D enclosure, an industrial-grade chassis aimed at rack-scale AI workloads rather than gamers. The enclosure alone costs $1,300 and includes a 1,300-watt power supply with high-current connectors built for the latest GPUs. On the host side, the system used HighPoint's Rocket 7634D switch adapter, a $999 PCIe Gen 5 card that introduces CopprLink's new CDFP connector.

CDFP serves as the physical backbone of CopprLink, built for stability in server environments where accidental disconnects are not an option. Once installed, it is not easy to remove by mistake, requiring a release tab to be pulled first.

Together, the enclosure and adapter come to roughly $2,300 before the GPU even enters the equation, pushing the total cost of the setup past $5,000. At least for now, that places CopprLink firmly in the realm of enterprise hardware rather than enthusiast gaming.

Still, the benchmark results may reshape how companies think about scaling external compute. The operating system detected the RTX 5090 over CopprLink immediately, with no extra drivers or manual setup required. As far as the system was concerned, the GPU behaved like a card plugged directly into a conventional motherboard slot.

In testing, the CopprLink connection delivered performance that was nearly indistinguishable from a native PCIe link, trailing by just 2.3%. That is a striking result, and one that effectively blows past the long-standing bandwidth limitations that have defined external GPU designs for years.

That kind of efficiency could matter for data centers and developers looking for more modular ways to deploy GPU power. Consumers are still largely stuck with slower Thunderbolt-based enclosures that run into bandwidth bottlenecks well before PCIe is fully utilized, but enterprise users may soon have a path to scaling GPU-heavy workloads externally without giving up internal-grade performance.

CopprLink looks less like a routine spec bump and more like a break from older eGPU approaches. For now, it remains expensive, narrow in focus, and confined to professional environments, but by matching an RTX 5090's native motherboard performance, CopprLink may have just set the new standard for what external GPU technology can deliver.

Permalink to story:

External GPUs were always second best. CopprLink may change that

 

[yRaz]

Outside of people replying to this post, no one is actually buying 90 class GPUs besides streamers or AI bros. I get that a 5090 is cheaper than a b6000 or whatever workstation card equivalent, but you can just rent time on a server farm rather than buy a workstation and that's a considerably cheaper option right now. Since most datacenters are going completely unused right now(can't wait for the AI bros to reply to that statement) renting compute in a data center is stupid cheap. I just want to thank the AI bros for investing in data centers making compute cheap, they're subsidizing it right now. You could build a $100,000US work station that takes weeks to process something or rent time in a data center to have something done in an couple hours for $1000.

 

[jonny888]

Outside of people replying to this post, no one is actually buying 90 class GPUs besides streamers or AI bros. I get that a 5090 is cheaper than a b6000 or whatever workstation card equivalent, but you can just rent time on a server farm rather than buy a workstation and that's a considerably cheaper option right now. Since most datacenters are going completely unused right now(can't wait for the AI bros to reply to that statement) renting compute in a data center is stupid cheap. I just want to thank the AI bros for investing in data centers making compute cheap, they're subsidizing it right now. You could build a $100,000US work station that takes weeks to process something or rent time in a data center to have something done in an couple hours for $1000.

Considering that datacentres are one of the intended/main targets for this, I'm not really sure what kind of argument you're trying to make.

Also, you don't think there are any people who bought a 5090 simply because they like to game at high refresh rates/resolutions in modern games, regardless of any other concern? I think you might be mistaken.

Also also, if you're the kind of person that has a one-time project you can outsource the compute work to, then sure that can make sense. If you're in a business which consumes hundreds or thousands of hours of GPU compute work on regular basis, then you have legitimate reasons to weigh up whether local hardware or cloud services are going to cost you more in the long run.

 

[yRaz]

Considering that datacentres are one of the intended/main targets for this, I'm not really sure what kind of argument you're trying to make.

Also, you don't think there are any people who bought a 5090 simply because they like to game at high refresh rates/resolutions in modern games, regardless of any other concern? I think you might be mistaken.

Also also, if you're the kind of person that has a one-time project you can outsource the compute work to, then sure that can make sense. If you're in a business which consumes hundreds or thousands of hours of GPU compute work on regular basis, then you have legitimate reasons to weigh up whether local hardware or cloud services are going to cost you more in the long run.

5090s are retarded, it's not about liking high refresh rates or anything like that. You want to play games at the propaganda settings nVidia is trying to sell us on. Noone asked for Ray tracing. Yeah, it's cool, but I notice most people's favorite part of Ray tracing is the ability to turn it off. And, sure, as a hobby people will spend whatever, but "regular people" aren't rushing out to have super high frame rates. Most people just want to play Skyrim for 2 or 3 hours A WEEK or play a few rounds of CoD after work

 

[NumberNine]

For five grand I can build a box that size or smaller that has the rest of the computer in there as well.

This is stupid pricing.
Not ignorant... stupid.

 

[Eldrach]

5090s are retarded, it's not about liking high refresh rates or anything like that. You want to play games at the propaganda settings nVidia is trying to sell us on. Noone asked for Ray tracing. Yeah, it's cool, but I notice most people's favorite part of Ray tracing is the ability to turn it off. And, sure, as a hobby people will spend whatever, but "regular people" aren't rushing out to have super high frame rates. Most people just want to play Skyrim for 2 or 3 hours A WEEK or play a few rounds of CoD after work

What people want and what you want may not align - as the sales statistics for 5090 cards that rendered them virtually unavailable for an entire year after launch would tell you.

 

[Endymio]

Since most datacenters are going completely unused right now(can't wait for the AI bros to reply to that statement)

When absurd lies are posted to a public forum, you have to expect them to be debunked.

You could build a $100,000US work station that takes weeks to process something or rent time in a data center to have something done in an couple hours for $1000.

You've never run a large mathematical or ML inference model on workstation hardware, have you? Even at today's prices, for $100K you can build an A100x8 station. A compute job that would take a couple weeks on that system would cost $10K from a major cloud vendor, and maybe $6-8K from a smaller one. That's fine if you only have one or two jobs ... but if you're running them constantly, your breakeven is just a few months time.

 

[agentnate009]

For five grand I can build a box that size or smaller that has the rest of the computer in there as well.

This is stupid pricing.
Not ignorant... stupid.

This is the problem with most all gamers, they don't have any idea what costs businesses for products like this, in fact, this is regular stuff for business. I've seen the cost of some items that businesses buy, then add in everyone wants their markup. This device may work well for some tech-oriented businesses. It was not designed for Joe Schmo gamer. This is what the business world deals with all the time, price gouging (from our perspective) costs to do business. That custom connector isn't cheap to produce, but some businesses may salivate over the bandwidth because it may allow them to cut costs elsewhere or allow them special niche case use that would be too expensive otherwise. Building a whole other computer may not work for a tech business when they can add some rack equipment to talk to the little pluggable boxes for certain use cases.

 

[jonny888]

5090s are retarded, it's not about liking high refresh rates or anything like that. You want to play games at the propaganda settings nVidia is trying to sell us on. Noone asked for Ray tracing. Yeah, it's cool, but I notice most people's favorite part of Ray tracing is the ability to turn it off. And, sure, as a hobby people will spend whatever, but "regular people" aren't rushing out to have super high frame rates. Most people just want to play Skyrim for 2 or 3 hours A WEEK or play a few rounds of CoD after work

I'm not sure who elected you to speak on behalf of why people who buy 5090s buy them, especially as someone who (I assume from your vitriol) hasn't bought one themselves. When you're talking about a piece of hardware that at it's core is designed to output frames as fast as possible, of course people who want the best refresh rates, especially at resolutions where those rates are harder to achieve, will go for the best hardware they can reasonably afford and justify to themselves. That's true regardless of any extra features they may or may not choose to enable. Throwing out insanely sweeping statements like "most people just want to play skyrim" is... I mean you must see how silly that is. If that is all YOU want to do then that's great. Plenty of other people have more demanding needs and wants. Not to mention, that the 5090 really wasn't the crux of this article or what it was talking about. It could be literally any GPU, they just happened to use that one to showcase a card that would be bandwidth-heavy and prove the point about the connector.

 

[NumberNine]

This is the problem with most all gamers, they don't have any idea what costs businesses for products like this, in fact, this is regular stuff for business. I've seen the cost of some items that businesses buy, then add in everyone wants their markup. This device may work well for some tech-oriented businesses. It was not designed for Joe Schmo gamer. This is what the business world deals with all the time, price gouging (from our perspective) costs to do business. That custom connector isn't cheap to produce, but some businesses may salivate over the bandwidth because it may allow them to cut costs elsewhere or allow them special niche case use that would be too expensive otherwise. Building a whole other computer may not work for a tech business when they can add some rack equipment to talk to the little pluggable boxes for certain use cases.

Maybe if your business is white-boxing crypto mining rigs in Indonesia.
Otherwise, seeing this little breakout box running in a rack with real hardware would communicate the same level of professionalism as running a stack of NUCs in there on shelves.

 

[Beerfloat]

For better or for worse, PCIe is THE vendor-neutral interconnect for storage and other expansion capability that everybody in the industry supports.

The SFF-TA-1032 connector adds up to 2 meter external cabling support to CopprLink's regular internal MCIO connectivity, which is useful for bridging to neighboring racks, for example.

Questions about who would need this or whether this is for professionals are absurd. Of course it needs to exist. And as it is designed to scale to upcoming 6.0 and 7.0 versions of PCIe, it's going to be around for years to come. And of course someone was going to take this industry standard and use it for nothing more than a simple external x16 connector. Because Thunderbolt and Oculink already demonstrated that there is demand for it.

 

[Theinsanegamer]

This is the problem with most all gamers, they don't have any idea what costs businesses for products like this, in fact, this is regular stuff for business. I've seen the cost of some items that businesses buy, then add in everyone wants their markup. This device may work well for some tech-oriented businesses. It was not designed for Joe Schmo gamer. This is what the business world deals with all the time, price gouging (from our perspective) costs to do business. That custom connector isn't cheap to produce, but some businesses may salivate over the bandwidth because it may allow them to cut costs elsewhere or allow them special niche case use that would be too expensive otherwise. Building a whole other computer may not work for a tech business when they can add some rack equipment to talk to the little pluggable boxes for certain use cases.

I mean, cool, good for you.

These products still have to compete with the rest of the market, and they are way overpriced for what they offer and the compromises they make.

5090s are retarded, it's not about liking high refresh rates or anything like that. You want to play games at the propaganda settings nVidia is trying to sell us on. Noone asked for Ray tracing. Yeah, it's cool, but I notice most people's favorite part of Ray tracing is the ability to turn it off. And, sure, as a hobby people will spend whatever, but "regular people" aren't rushing out to have super high frame rates. Most people just want to play Skyrim for 2 or 3 hours A WEEK or play a few rounds of CoD after work

The 5090 has nearly as many users on Steam as the 7900xtx.....so.......

Maybe YOU just dont like the 5090? Cause nvidia has no problem selling them and consumers have no issue buying them.

 

[agentnate009]

I mean, cool, good for you.

These products still have to compete with the rest of the market, and they are way overpriced for what they offer and the compromises they make.
The 5090 has nearly as many users on Steam as the 7900xtx.....so.......

Maybe YOU just dont like the 5090? Cause nvidia has no problem selling them and consumers have no issue buying them.

You are missing the point, it probably has more to do with the cable, the bandwidth, and as of right now, certain niche cases, but people find ways to use such a product for more than people imagined at first. Did you even read the post?? Allow me to quote the pertinent part that you seemed to have missed: "Developed by PCI-SIG for high-bandwidth >>>>>enterprise use<<<<<, CopprLink provides a full PCIe 5.0 x16 interface with throughput of 32 GT/s per lane, or 64 GB/s of total bandwidth." (emphasis mine)

Dude, just stop. I said nothing about the 5090, why are you even bringing that up or trying to paint me as something I am not? I never said ANYTHING about the 5090. I argued for the use case of such a product.

 

[yRaz]

I mean, cool, good for you.

These products still have to compete with the rest of the market, and they are way overpriced for what they offer and the compromises they make.
The 5090 has nearly as many users on Steam as the 7900xtx.....so.......

Maybe YOU just dont like the 5090? Cause nvidia has no problem selling them and consumers have no issue buying them.

I don't understand why the XTX is important. I really don't feel like participating in the circle jerk of high end graphics cards. 5090s are dumb. XTX's are dumb. I can buy both but I won't because I'm not dumb

 

[rE3ee]

Considering that datacentres are one of the intended/main targets for this, I'm not really sure what kind of argument you're trying to make.

Also, you don't think there are any people who bought a 5090 simply because they like to game at high refresh rates/resolutions in modern games, regardless of any other concern? I think you might be mistaken.

Also also, if you're the kind of person that has a one-time project you can outsource the compute work to, then sure that can make sense. If you're in a business which consumes hundreds or thousands of hours of GPU compute work on regular basis, then you have legitimate reasons to weigh up whether local hardware or cloud services are going to cost you more in the long run.

His point is that this is a consumer Product site , nobody is buying 90 série for gaming , except maybe rich kids

 

[rE3ee]

Considering that datacentres are one of the intended/main targets for this, I'm not really sure what kind of argument you're trying to make.

Also, you don't think there are any people who bought a 5090 simply because they like to game at high refresh rates/resolutions in modern games, regardless of any other concern? I think you might be mistaken.

Also also, if you're the kind of person that has a one-time project you can outsource the compute work to, then sure that can make sense. If you're in a business which consumes hundreds or thousands of hours of GPU compute work on regular basis, then you have legitimate reasons to weigh up whether local hardware or cloud services are going to cost you more in the long run.

I'm a business with few kW of GPU , I don't come here for news but entertainment, you kids are funny

 

[jonny888]

His point is that this is a consumer Product site , nobody is buying 90 série for gaming , except maybe rich kids

Please point me to your credible, non-anecdotal data source that shows that no-one is buying that type of card for those reasons. Or are you, perhaps, just spewing a narrative you like to believe?

Also, this is a tech news site. The clue is in the name. This is tech news. And as already stated, the 5090 is not even the point of the article in anyway, it's just the card the company chose to demo with to show the connector's capabilities.

 

[Megalomaniac]

As a 5090 owner, let me be first to say that I love my card, I love that I got my card under MSRP and I wouldn't want any other card but my 5090. Considering how difficult and troublesome buying new GPU will be in the future, I doubt I'll have a 6090 or 7090. Perhaps when I upgrade again, 10 years from now, the issues we have buying properly priced, unscalped equipment will return.

 

[georges01]

Outside of the childish comments, cool stuff indeed.

While Thunderbolt (& USB4) does offer fairly "high speed" connections for ether connecting portable devices to other devices, Oculink is gaining popularity for eGPU enclosures and can be found on many "mini-PCs".

Surely finding a Laptop device with TB3/4 or USB4 ports is far more likely then finding one that has a Oculink or this new Coprlink port.

Frankly, outside of bridging racks of devices I personally can't see any general office or consumer use for this standard. (after all, if one has access to the PCIe16 slot in the host, why not merely install the GPU card in the host and forgo installing the Coprlink adapter?)