News AMD teams up with Cisco, Nokia, and Jio Platforms for Open Telecom AI platform

[Admin]

AMD, Cisco, Nokia, and Jio Platforms will develop an open platform for telecommunication machines with advanced security based on AMD's processors and telco technologies from Cisco and Nokia.

AMD teams up with Cisco, Nokia, and Jio Platforms for Open Telecom AI platform : Read more

 

[Jame5]

I know it's not fair, as it's fundamentally different forms of AI, but I hear AI telecom solutions and my first thought is a hallucinated connection between endpoints that shouldn't be talking to each other.

Imagine calling your wife and being connection to a random person overseas, etc. Or trying to reach a server, and the AI telecom platform decides to reroute your connection due to "traffic shaping" and connects you to a completely different server. Both scenarios wildly unlikely, but also the kind of thing that comes to mind when thinking of jamming AI into telecom platforms.

I know specific, purpose driven AI is much more reliable than general purpose AI solutions, but it's hard to shake that thought.

 

[bit_user]

Under the terms of the agreement, AMD will provide high-performance computing solutions, including EPYC CPUs, Instinct GPUs, DPUs, and adaptive computing technologies.

So far, I think ROCm hasn't exactly demonstrated telco-grade reliability.

Given Nvidia's DPU portfolio, it's interesting Cisco didn't prefer to partner with them. Maybe Cisco feels too threatened by their Mellanox business?

 

[AkroZ]

Given Nvidia's DPU portfolio, it's interesting Cisco didn't prefer to partner with them. Maybe Cisco feels too threatened by their Mellanox business?

It's seems the choice has been made by JPL and Nvidia doesn't like open standards, they prefere selling a proprietary solution like Bluefield which is a concurrent to Cisco.

 

[jp7189]

I know it's not fair, as it's fundamentally different forms of AI, but I hear AI telecom solutions and my first thought is a hallucinated connection between endpoints that shouldn't be talking to each other.

Imagine calling your wife and being connection to a random person overseas, etc. Or trying to reach a server, and the AI telecom platform decides to reroute your connection due to "traffic shaping" and connects you to a completely different server. Both scenarios wildly unlikely, but also the kind of thing that comes to mind when thinking of jamming AI into telecom platforms.

I know specific, purpose driven AI is much more reliable than general purpose AI solutions, but it's hard to shake that thought.

If an AI is trained on what a 'normal' network looks like, it can be very good at recognizing 'not normal' with significant nuance. From that it can decide to drop packets, update route weights, etc. with more granularity compared to a hard coded rule.

AI is also very good at complex error recovery. Take a look at stable diffusion, though not the same at all, we can draw some parallels. SD takes a starting array of completely random pixels (pure noise) and 'fixes' the noise back to something recognizable. The concept can be applied to error prone communications like wireless. Start with a noisey, almost unrecognizable signal, and 'fix' it back to something useful.

 

[bit_user]

If an AI is trained on what a 'normal' network looks like, it can be very good at recognizing 'not normal' with significant nuance. From that it can decide to drop packets, update route weights, etc. with more granularity compared to a hard coded rule.

From my non-expert perspective, I think you have good points about anomaly detection (e.g. DDoS attack) and adaptive routing optimization.

AI is also very good at complex error recovery. Take a look at stable diffusion, though not the same at all, we can draw some parallels. SD takes a starting array of completely random pixels (pure noise) and 'fixes' the noise back to something recognizable. The concept can be applied to error prone communications like wireless. Start with a noisey, almost unrecognizable signal, and 'fix' it back to something useful.

I just don't see stream-level error-correction happening in the core. It makes much more sense to delegate this to edge devices, which already have TOPS of AI, the use of which also doesn't incur a cost to core network operators.

A more chilling use of AI, in the network core, would be for content filtering & censorship. However, the use of end-to-end encryption should limit the potential for such measures. Perhaps there's enough to be gleaned simply by looking at overall communication patterns in who's talking to whom.

 

[jp7189]

I just don't see stream-level error-correction happening in the core. It makes much more sense to delegate this to edge devices, which already have TOPS of AI, the use of which also doesn't incur a cost to core network operators.

I mentioned it because Nokia and wireless were mentioned, but you're right that something like that will be at the edge and not core.

 

[jp7189]

A more chilling use of AI, in the network core, would be for content filtering & censorship. However, the use of end-to-end encryption should limit the potential for such measures. Perhaps there's enough to be gleaned simply by looking at overall communication patterns in who's talking to whom.

I imagine your comment is focused on government/public infrastructure, but it's fairly common of corp networks to decrypt packets for inspection, and block or force downgrade of encryption that doesn't comply. E.g. Chrome connecting to Google services can be force downgraded to a standard encryption that is decryptable.

 

[bit_user]

I imagine your comment is focused on government/public infrastructure, but it's fairly common of corp networks to decrypt packets for inspection, and block or force downgrade of encryption that doesn't comply. E.g. Chrome connecting to Google services can be force downgraded to a standard encryption that is decryptable.

Yeah, I was thinking this sort of initiative involving GPUs and DPUs is aimed at core infrastructure, because I don't expect most corporations to spend so much money on their network infrastructure. Sure, you have firewalls and some of those will be enforcing content policies, but it makes much more sense for them to put anything more compute-intensive on the actual PCs, which they also control. They already paid for those, and putting it "at the edge" scales better.

 

[jp7189]

Yeah, I was thinking this sort of initiative involving GPUs and DPUs is aimed at core infrastructure, because I don't expect most corporations to spend so much money on their network infrastructure. Sure, you have firewalls and some of those will be enforcing content policies, but it makes much more sense for them to put anything more compute-intensive on the actual PCs, which they also control. They already paid for those, and putting it "at the edge" scales better.

Content filtering doesn't need packet decryption, but IPS and behavior analytics are very rudimentary without it; malware detection is impossible. Therefore, packet decryption is common at the firewall. It is by far the most compute intensive operation a firewall does and it is dominated by Intel and custom ASICs today. I'm not aware of any hardware appliance that uses AMD.