Why Cellular Will Not Deliver Low Latency

We are a long way from the IMT requirements

Much of the rationale for 5G was predicated on lowering latency. Getting latency numbers below 10ms, we were told, would enable haptic communications, the Internet of Senses, remote control of vehicles and robots and, yes, remote surgery too. Indeed, almost all of the new applications for 5G were based on low latency. For example, Ericsson in 2019 said[1]:

“With 5G's URLLC (Ultra Reliable Low Latency Communication), it will soon be possible to have round-trip latency in the millisecond range. This will enable the creation of real-time interactive systems – capable of delivering physically tactile experiences remotely. Round-trip sub-millisecond latency, together with 5G's carrier-grade reliability and availability, will open a new wave of 5G tactile internet applications. Such applications will increasingly enable human-machine interaction, based on haptic or tactile communication. In other words, a sense of touch can be transported to the other side of the world in real-time.”

The minimum Requirements for 5G, set out in Report ITU-R? M.2410-0 (11/2017) regarding latency were

–????????? 4 ms for eMBB

–????????? 1 ms for URLLC

Minimum, in this context, of course means that in practice the actual numbers need to be lower than this. Technologies need to meet these requirements in order to be IMT-2020 compliant with all that entails, including, in principle, access to IMT-2020 spectrum.

To put this into perspective, in 2023, four years after Ericsson told us we would “soon” have sub-ms latency, Ookla reported[2] that for the three top mobile operators in the U.S. during Q4 2023, T-Mobile had the lowest latency at 50 ms. Verizon Wireless was a close second at 57 ms. AT&T was third at 64 ms. For the 5G element alone T-Mobile registered the lowest median 5G latency with a median value of 48 ms, while their nearest competitor, Verizon Wireless was 52 ms.

So 5G latency in the US is only a little better than 4G. T-Mobile have deployed standalone (SA) 5G, but even with this, still have a latency over 10x greater than the eMBB requirement, and nearly 50x greater than the much-proclaimed 1ms ultra-reliable and low latency communications (URLLC) target.

Other countries are broadly similar, albeit slightly better. A reasonable rule of thumb is that 4G delivers around 40ms latency while 5G can reduce that to around 30ms.

By comparison, a speedtest run over Wi-Fi will, depending on the broadband connection, typically shows a latency of around 5-10ms.

Why this broadly does not matter

Ericsson, despite their earlier evangelism, in 2022 published a blog entitled “Who cares about latency in 5G?”[3]

In it, they concluded:

“Humans will hardly be able to tell – at least not appreciate much – the difference between a latency of 30 vs. 50 ms when using their everyday apps on their smartphones. This is shown in the upper chart in Figure 1. Even ambitious first-person shooter gamers would be happy with a reliable latency in the range of 30 – 50 ms. I have yet to find a sound research study that has tested humans who have been trained to react extremely fast, e.g., Olympic sprinters, race car drivers, or championship gamers. From what I have read so far, it seems that the lower limit of latency for this small fraction of the human population is not lower than 20 ms, i.e., for humans, lower latencies would not make a difference.”

Hence, for people, 4G latency levels are fine, and achieving the ITU requirements would not bring any benefits. They went on to note:

“Machines, not humans, can benefit from the ultra-low and ultra-reliable latencies that only 5G can provide. For example, think about video-controlled high-precision robots in a smart factory. Here, we are talking about latencies below 10 ms, i.e., ultra-low, and without latency spikes exceeding 10 ms, i.e., ultra-reliable. Meeting these tough requirements is one of the key drivers behind Ericsson’s launch of a new product offering targeted at Time-Critical Communication.”

The issue of machines is a point I will return to in discussing whether the targeted latency numbers are only needed in private networks.

Providing low latency is expensive

Why is it that 5G latency levels are nowhere near the IMT targets? It is the case that the SA core is needed to deliver the lowest latency, but T-Mobile and others have deployed this core, and yet latencies remain stubbornly at around 4G levels.

The simple answer is that delivering low latency is expensive. Simplistically, radio resource needs to be left idle so that there are slots free immediately that a low latency application wants to transmit. Really low latency requires smaller slots – the 5G “minislot” – so that the time waiting for the next slot to become available is minimised, but lots of small slots have a large overhead of framing structure and frame allocation information. The net effect is that system capacity is reduced when a network is engineered to allow low latency.

A recent paper[4] looked into this and concluded:

“When taking a holistic view on the performance, advanced system-level simulation results show that 1 Mbps of URLLC traffic results in an eMBB throughput reduction of up to 60 Mbps, i.e. URLLC traffic can be up to 60 times more costly than traditional best-effort. The presented results can be used by cellular service providers, such as operators, for understanding and dimensioning the tradeoffs and impact towards traditional network services when adding URLLC services to their portfolio.”

So, 1Mbits/s of low latency traffic can displace 60Mbits/s of “normal” traffic. Assuming a cell operating close to capacity, then a low-latency user would need to pay 60 times more per byte than a conventional user. Then there are other costs incurred in the core network, in the peering and more.

It could be argued that this is no different from offering FWA, where FWA subscribers generate a lot more traffic than a mobile user – something like 10x in the US but rising towards 50x in Australia. However, FWA only tends to be offered where there is spare network capacity, mostly outside dense urban areas. Low latency would likely need to be offered in the densest areas.

Hence, an MNO is unlikely to set up their network to offer URLLC just in case anyone wanted it. They would need to see a sizable application that was prepared to pay a significant premium.

To return to the point about SA – one of the main justifications of SA was to deliver “true 5G” services which required lower latency. If lower latency is not viable, then this casts doubt on whether implementing SA is sensible. It would need to be justified on lower through life costs for a new type of core rather than on the services that it enables.

Private 5G networks do better

Ericsson’s revised position in 2022 was that only machines need latency levels below 30ms (although the original 5G hype strongly suggested people needed sub-10ms latency for the “Internet of Senses”). The examples they quoted of high-speed robots are highly likely to be within factories or a private campus. This suggests that we might see a split where public networks deliver 30-50ms latency while small-scale private networks deliver much lower levels.

Logically, this makes sense. Private networks may have ample capacity, and so could tolerate the lower throughput as a result of using low-latency configurations. The application would be clear and any extra cost might be justified by the business benefit.

Indeed, private networks do appear to have lower latency. A paper[5] looking at performance of private 5G networks found latency levels in the region 10-15ms. This is much lower than public networks but still some way off the IMT requirements. A key reason for this is that most private networks use shared spectrum in the 3.5GHz band (for example, 3.8-4.2GHz). This spectrum is set up as time division duplex (TDD) and the 5G standard specifies a 10ms frame. This means that for a round-trip message flow, the average case will be around 10ms since a user will need to wait for an uplink slot to become available then wait for the downlink slot to receive the response. Faster speeds can be achieved by having multiple transitions between the downlink and uplink within a 10ms frame but this is inefficient with guard-times needed. It will likely require FDD to reduce latencies down towards the 1ms requirement. This approximately 10ms lower limit is indeed what the measurements show. Spectrum allocations for private networks are highly unlikely to change to FDD any time soon.

While 10ms is better than public 5G, it is still worse than a decent Wi-Fi network. Those who need very low latency might want to look elsewhere than 5G.

Time to end the latency discussion?

So 5G has a latency that for public networks is not materially different from 4G. There is no obvious need for lower latency on public networks and delivering lower latency would be very expensive. There might be a need for lower latency on a small number of private networks (probably only a few thousand globally) but even here getting below around 10ms is difficult and other technologies such as Wi-Fi or bespoke solutions might do better.

Should we admit that the aim for lower latency was misguided, and focus on things that are important such as coverage? The industry does not think so, the IMT-2030 discussion document[6] “Recommendation ITU-R M.2160-0 (11/2023) Framework and overall objectives of the future development of IMT for 2030 and beyond” suggests “the research target of latency (over the air interface) could be 0.1 – 1 ms.”

Credibility, legality and lack of diligence

The conclusion that 5G will not deliver lower latency but that we do not need it anyway might not seem particularly consequential. But it does raise questions about the way that the cellular industry conducts itself.

Credibility. The cellular industry – manufacturers, academics, some operators and many others – made much of the lower latency that 5G would deliver and how this would enable revolutionary new applications. Mainstream TV adverts from large operators showed robotic surgery being conducted from the back of a church during a wedding and robotic barbers at the top of mountains. Yet lower latencies have not been delivered, will not be delivered, and neither will these applications. While memories are sometimes short, the industry is no longer one that is seen as credible. This leads to problems when significant changes are needed such as to spectrum or industry structure. The industry would do well to win back trust, and that is not achieved by suggesting 6G could deliver latency levels lower than those targeted for 5G.

Legality. The ITU requirements document for 5G is, strictly, a document that sets out the requirements that 5G has to meet to be classified as IMT-2020. Spectrum allocations are then made to IMT at WRC meetings and, in principle, only those technologies and networks that meet ITU Requirements should be allowed access to the IMT spectrum.

In practice, nationally issued licenses rarely require a particular technology or even IMT compliance, hence no operator is likely to find themselves transmitting illegally as a result of their 5G system not meeting the ITU requirements. And, in passing, 5G deployments not only fail to meet the requirements for latency, they also fail on spectrum efficiency, device connection density, energy efficiency, area traffic capacity and peak data rates. In fact, the only target that is clearly met is user-experienced data rates.

If systems that are so delinquent against the ITU Recommendations are being deployed, then this casts doubt as to whether there should be “IMT allocations” and indeed whether the ITU should have any role in mobile network technology at all.

Lack of diligence. How did the proponents of 5G get it so wrong? It was always obvious that delivering low latency would come with a cost to it and the fact that 30ms latency was sufficiently low was well known. That networks would not deliver was entirely predictable, indeed in “The 5G Myth” in 2016 I wrote “Latency is important, but 5G seems unlikely to have a materially lower latency than 4G in practical situations.” Why did those designing 5G not realise this? Broadly, the work of the academics and visionaries went unchecked by those who would need to implement, operate and most importantly pay for the networks. The operators especially should have done their due diligence before the ITU recommendations were finalised.

There are some important lessons to learn here.

[1] https://www.ericsson.com/en/blog/2019/4/5g-tactile-internet

[2] https://www.speedtest.net/global-index/united-states

[3] https://www.ericsson.com/en/blog/2022/8/who-cares-about-latency-in-5g

[4] https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=9732349

[5] 55th CIRP Conference on Manufacturing Systems, “Measurement and comparison of data rate and time delay of end-devices in licensed sub-6 GHz 5G standalone non-public networks”, Thorge Lackner et al

[6] https://www.itu.int/rec/R-REC-M.2160-0-202311-I/en