Intel Partner Prepares Dual Arc "Battlemage" B580 GPU with 48 GB of VRAM

[igormp]

I'd presume if they want to do it better CXL, but they might still do something like incorporate TB. It would be trivial to incorporate 1 or 2 TB ports to connect them internally or externally or both. It's possible they could even daisy chain configurations between PC's with TB eventually.

I don't see TB or CXL making since given that they can just use PCIe straight away within the board itself, but we shall see.

 

[Dyatlov A]

The picture reminds me for the mighty 3dfx Voodoo5 5500 and then the company has disappeared.

 

[demirael]

I feel like you've just shown us a AI depiction of a Quad Gen 5 M.2 device with cooling woes...lol stop deep faking us with your voodoo 3DFX.

I feel like you should shut up. 3Dfx actually made prototypes of those, some with 4 GPUs on one board, and the image matches other sources and the one I personally saw once. Should've swiped it really, it was at an electronics recycler...

 

[InVasMani]

I don't see TB or CXL making since given that they can just use PCIe straight away within the board itself, but we shall see.

TB seems unlikely at least in it's current form though for external connectivity for smaller bandwidth restraints it could be moderately useful perhaps. As for CXL, you'd get direct access to pooled memory resources across multiple accelerators, which could significantly reduce memory bottlenecks compared to standard PCIe. So really depends how they flexible they intend to make a dual GPU design. Also CXL, memory pooling is far more flexible, enabling asymmetric VRAM allocation like 12GB + 4GB rather than a more rigid 8GB + 8GB split. Dynamic memory allocation in a dual GPU configuration is a big difference that opens up a lot more flexibility on how the paired GPU's work together or GPU cores.

 

[igormp]

TB seems unlikely at least in it's current form though for external connectivity for smaller bandwidth restraints it could be moderately useful perhaps

For external devices? Sure, but not really as an interconnect between GPUs in the same PCB, specially when you have PCIe right there.

As for CXL, you'd get direct access to pooled memory resources across multiple accelerators, which could significantly reduce memory bottlenecks compared to standard PCIe.

CXL is built on top of PCIe, with added coherence and other protocols.
Nonetheless, that would require the GPU to act as a CXL Host, which adds lots of complexities to a GPU.

Also CXL, memory pooling is far more flexible, enabling asymmetric VRAM allocation like 12GB + 4GB rather than a more rigid 8GB + 8GB split. Dynamic memory allocation in a dual GPU configuration is a big difference that opens up a lot more flexibility on how the paired GPU's work together or GPU cores.

WDYM by "asymmetric VRAM allocation"?
It seems like you're not talking about the interconnect between GPUs anymore, but rather how the CPU communicates with both GPUs.

 

[InVasMani]

I mean not symmetrically allocated between each. Like the two GPU's working in tandem with say 16GB VRAM total, but the secondary GPU only has access to like say 1/4 of it and the primary GPU has access to 3/4 of it for a 4GB +12GB split. It allows you to leverage the compute power of the second card, but with a more limited VRAM capacity that it holds while the primary card has a larger capacity it can store and same compute. It kind of shifts and limits how you might use use and really either could be arranged as the primary or secondary GPU for a given GPU task and rely the resulting information back to other provided the latency is reasonable enough.

I am talking about the interconnect between GPU's because in a separate dual GPU configuration you'd have to rely the data between them and sync them and it's latency and bandwidth sensitive which is a big part of the core issue. For SLI/ CF it has to mirror the information in VRAM the way it was designed to operate. It could be handled more dynamically assigned in a primary/secondary arrangement. It could be more orchestrated similarly to what we see with music sequencers and synths working in tandem and have different strengths and weaknesses and combining together cohesively in a more harmonious manner.

In essence you could use a primary or secondary gpu to perform a task prior or after to one or the other GPU's dynamically. Like running my music synth thru a music sampler and having it process it with DSP effects and/or in other ways manipulating the audio signal itself.

 

[Wirko]

CXL is built on top of PCIe, with added coherence and other protocols.
Nonetheless, that would require the GPU to act as a CXL Host, which adds lots of complexities to a GPU.

Agreed, and I've only seen CXL mentioned in association with PCIe-connected memory modules (or storage, etc.) The modules can be connected to multiple hosts through a CXL switch, which is probably more complex than a simple PCIe switch.

 

[igormp]

Like the two GPU's working in tandem with say 16GB VRAM total, but the secondary GPU only has access to like say 1/4 of it and the primary GPU has access to 3/4 of it for a 4GB +12GB split. It allows you to leverage the compute power of the second card, but with a more limited VRAM capacity that it holds while the primary card has a larger capacity it can store and same compute.

Why would you ever want to do that?

I am talking about the interconnect between GPU's because in a separate dual GPU configuration you'd have to rely the data between them and sync them and it's latency and bandwidth sensitive which is a big part of the core issue.

That's not really how most multi-GPU tasks operate. You usually keep data sharing to a minimum.
Those tasks are also often not latency sensitive, but rather bandwidth dependent.

For SLI/ CF it has to mirror the information in VRAM the way it was designed to operate.

I guess you're thinking more of a gaming scenario, where SLI/CF had tons of hardships and you need to keep tons of stuff in sync.
Such dual-GPU as in the OP would likely be used for offline rendering or machine learning tasks, where you keep different pieces of data in each GPU and only do syncs at a minimum, such as a gradient sync after each forward pass, or just doing gradient accumulation into the next layer that lives in a different GPU.

A supposedly dual-GPU like the one in the topic would appear as 2 different GPUs to the system, not as a single one. Your framework of choice would then be able to use those as appropriate.

 

[InVasMani]

I was leaning more in the direction of traditional gaming rendering scenario's like SLI/CF was originally intended for yes. Technically it doesn't have to be strictly around that. The GPU cores don't even have to be identical you could could have specialized GPU cores aim at more and different use case functions like tailored for compression/decompression by a GPU. It's more about thinking of it similarly to like Intel with SOC tile, but around GPU accelerated tasks.

I think ideally what I'd like to see is something a bit like VRAM on substrate that is accessible by other GPU cores that has like a fix allocation amount, but unused allocation at within like a certain refresh interval cycle period if it wasn't like requested to use all of it and would just otherwise be sit by idly unused could get shared for use by the other GPU cores. A little bit akin to how Intel's E cores L2 cache behaves, but with kind of some play nicely tweaks.

AMD's point to point UALink is a pretty good example of the type of coherency accessible sharing frame work I'd like to see between GPU cores and sharing of VRAM to distribute across GPU's.

 

[TechLurker]

Adding to the list of Dual GPUs (and possibly missing it being mentioned), the Radeon Pro Duo had 2 Fury X GPUs and HBM1 was also a thing.

 

[Ninehell]

Lossless scaling frame gen with 1 card if it work, instead of 2 cards (1 for render and 1 for frame gen)

 

[wolf]

I thought this thread would be more about the interesting possible dual B580, but it's apparently more about *checks notes* revisionist history and anti-AMD biased article writing for not explicitly acknowledging AMD's past dual GPU products. Amazing how we can read the same article and see such different things, I can't imagine how I'd cope day to day with such a strong persecution complex, it seems like agony.

 

[od1n]

First impression released ?!?!

https://i.postimg.cc/bY477W8D/Bild-2025-05-20-115406717.png

 

[od1n]

Gamers Nexus

 

[Scattergrunt]

View attachment 400380
First impression released ?!?!

https://i.postimg.cc/bY477W8D/Bild-2025-05-20-115406717.png

Source?

I thought this thread would be more about the interesting possible dual B580, but it's apparently more about *checks notes* revisionist history and anti-AMD biased article writing for not explicitly acknowledging AMD's past dual GPU products. Amazing how we can read the same article and see such different things, I can't imagine how I'd cope day to day with such a strong persecution complex, it seems like agony.

Its bizarre that people immediately jump to the revisionist history idea when its more likely that the author just didnt think about AMD's dual gpu products when making his comparison, or wanted to highlight a example that the average reader would know. And keep in mind, alot more, normal, average people know who NVIDIA is rather than AMD, especially as of lately.