News A 48GB dual-GPU Intel Arc B580 is reportedly in the works — Computex reveal rumored

[Admin]

A new rumor claims that the Intel Arc B580 could get a dual-GPU variant with 48GB of memory.

A 48GB dual-GPU Intel Arc B580 is reportedly in the works — Computex reveal rumored : Read more

 

[I touch grass]

SLI is back maybe
(I do know that SLI is propriety to Nvidia so a SLI equivalent from intel)

 

[Pierce2623]

Seems kind of overboard since it’s unlikely to hit 5070ti performance considering the 5070ti is 136% faster than a 4060.

 

[usertests]

Goofy product, but I would like to see more AIB customization. There's no reason that there can't be a 32 GB 9070 XT already.

 

[AkroZ]

The B580 chip doesn't support Xe Link and neither Deep Link. It will most likely be considered as two separated cards on the same PCB sharing the same PCIe port, by example you can find PCIe card where you can put multiples NVME drives.
Useful for AI, but in gaming you have not many games that can use multiples GPU (6 published games, like Ashe of Singularity).

 

[Thunder64]

Seems kind of overboard since it’s unlikely to hit 5070ti performance considering the 5070ti is 136% faster than a 4060.

It wouldn't be for gaming. It probably wouldn't be great for AI either. Sounds more like a gimmick.

 

[Jabberwocky79]

I owned a GTX 690 for years... awesome, beast of a card. But the dual-GPU, a.k.a SLI design created a lot of problems in games. Most of the time I had to run it in single mode just to keep weird bugs and anomalies from happening.

This sounds cool, but I'm having trouble seeing how it would be a huge advantage.

 

[Zerk2012]

Not a gaming card for sure.
Workstation.

 

[greenreaper]

Realistically you could perhaps do as well for cheaper using multi-GPU capable software with two B580s. Most PSUs support dual 8-pins and it only uses PCIe x8 anyway.

 

[cyrusfox]

This is clearly for LLM, and would sell regardless, create your own local LLM, I imagine price would be $800-1000 and would be capable of 60 tokens per second, compare that to Chatgpt, which charges ~1$ per million tokens, then a B580x2 48gb would be capable of 60*3600*24=5.18 million tokens a day, so equivalent of ~$5. Roughly pay for itself in 200 days, will be a little slower with electricity cost. This is the new Crypto...

 

[Mr Majestyk]

I'd rather a 24GB B770/780 for desktop. I hope AMD can release a 24GB 9080XT with ~ 6000 cores on 384 bit bus, or even better a GDDR7 version for 256bit bus.

 

[Pierce2623]

It wouldn't be for gaming. It probably wouldn't be great for AI either. Sounds more like a gimmick.

My point is that it’s not even great for AI. Who’s talking about gaming? Sure it has a lot of VRAM but it’s still not competive performance. How many people need 48GB for LLMs BUT they also don’t care if it’s much slower than other 48GB solutions? I just don’t see the draw when the card will cost more than a Strix Halo machine and OpenAPI is way behind even AMD for AI compatibility.

 

[Eximo]

I'd rather a 24GB B770/780 for desktop. I hope AMD can release a 24GB 9080XT with ~ 6000 cores on 384 bit bus, or even better a GDDR7 version for 256bit bus.

If they make a B770 it would be 16GB like the A770, but it isn't cost effective to do so. 3Gb GDDR7 would require a new chip entirely. I think they are going to hold out for Celestial.

Even with the comparable improvements that we can translate from the B580 and the small node shrink, a potential size matching chip might have 30 Xe cores, if you stretch to 32 it would be a bit bigger than a RTX 5080 die. (RTX 5070 die is smaller than a B580 FYI) GDDR6 as well, so they would be in the same boat as AMD, 9070 XT die is only slightly smaller than a 5080 die.

So unless Intel paid for TSMC 4N like Nvidia, or moved straight to their own 18A node, they could launch an expensive card that performs half as well as the competition. So more selling at cost or loss.

 

[Pierce2623]

If they make a B770 it would be 16GB like the A770, but it isn't cost effective to do so. 3Gb GDDR7 would require a new chip entirely. I think they are going to hold out for Celestial.

Even with the comparable improvements that we can translate from the B580 and the small node shrink, a potential size matching chip might have 30 Xe cores, if you stretch to 32 it would be a bit bigger than a RTX 5080 die. (RTX 5070 die is smaller than a B580 FYI) GDDR6 as well, so they would be in the same boat as AMD, 9070 XT die is only slightly smaller than a 5080 die.

So unless Intel paid for TSMC 4N like Nvidia, or moved straight to their own 18A node, they could launch an expensive card that performs half as well as the competition. So more selling at cost or loss.

They literally just said “those waiting on a b770 stay tuned” today….

 

[Eximo]

I've been tuned for years. We'll see.

 

[Mindstab Thrull]

I've been tuned for years. We'll see.

I'm completely untuned. Overheating, require too much power, and have to shut down regularly to keep functional...

... oh, my PC is fine, thanks for asking!