Intel Unveils New Pro GPUs, Designed for AI Workloads

[SirShanova]

Summary

On Monday, Intel showed off three new graphics cards as part of the new "Arc Pro" lineup. The three models, Arc Pro B60 Dual, Arc Pro B60 Single, and Arc Pro B50, have relatively low TDPs while maintaining similar or greater performance (in Intel-supplied graphs) to Nvidia A1000 8GB cards at only 70% the cost. The Arc Pro B60 Dual also is the first dual-GPU die card we've seen from a manufacturer (at this pricepoint, not a Radeon W6800X Duo at $5k) in a long while, like the Nvidia GTX TITAN Z from 2014 or the AMD Radeon R9 295X2.

 

Quotes

Quote

"Internally, the Maxsun Arc Pro B60 Dual lacks a PCIe bridge chip. Since the "BMG-G21" silicon has a PCI-Express 5.0 x8 host interface, both GPUs are connected across the x16 gold finger, and rely on PCIe lane segmentation at the host level. This is precisely how M.2 NVMe riser AICs work, where they split the x16 connection among the four x4 M.2 SSDs. The primary use-case of the Arc Pro B60 Dual 48 GB is AI inferencing..."

 

My thoughts

Never in my wildest dreams did I think that Intel would make GPUs interesting again! Dual GPU cards in 2025? Please, it's been over a decade of monotony between a company who has a monopoly on performance and charges accordingly or a spottily-supported competitor whose high-midrange pricing isn't far off with an inferior product. I am so glad to see new, exciting designs and innovations again in the tech space. With 46gB of VRAM, these things are going to be a beast. If these cards are truly available and for less $1000 (which are big asks in today's market), I think we'll have to see another Intel Arc challenge. This seems like good news for Intel Arc Celestial and Druid!

 

Sources

https://www.computerbase.de/news/grafikkarten/intel-arc-pro-b60-und-b50-die-dual-gpu-grafikkarte-mit-2x24-gb-vram-ist-wirklichkeit.92707/#bilder
https://www.tomshardware.com/pc-components/gpus/intel-launches-usd299-arc-pro-b50-with-16gb-of-memory-project-battlematrix-workstations-with-24gb-arc-pro-b60-gpus

https://www.techpowerup.com/336966/maxsun-arc-pro-b60-dual-48gb-graphics-card-hands-on

[igormp]

22 minutes ago, SirShanova said:

If these cards are truly available and for less $1000 (which are big asks in today's market)

I'll be totally buying one of those for my 2nd inference desktop if that becomes reality.

 

Kinda disappointing how they did not add any extra interconnect between GPUs and in the end it's just a dumb x8/x8 design, but for that price I'm not complaining.

[Poinkachu]

28 minutes ago, SirShanova said:

-snips-

GamersNexus made a video tearing one of those down

Kinda wished he brought Buildzoid there, you know, then we'll see BZ rambling solo for an hour while looking at nothing but the PCB.

 

 

[SirShanova]

7 minutes ago, igormp said:

I'll be totally buying one of those for my 2nd inference desktop if that becomes reality.

 

Kinda disappointing how they did not add any extra interconnect between GPUs and in the end it's just a dumb x8/x8 design, but for that price I'm not complaining.

I agree that it's disappointing, but also Gen4x16 and Gen5x8 are about equal. Certainly is limiting, but 32 GT/s isn't too bad

[SirShanova]

2 minutes ago, Poinkachu said:

Kinda wished he brought Buildzoid there, you know, then we'll see BZ rambling solo for an hour while looking at nothing but the PCB.

Hopefully we'll see more teardowns in the coming days. Intel is making moves on segments other than the $250 range, I expect this will get some decent coverage

[igormp]

1 hour ago, SirShanova said:

I agree that it's disappointing, but also Gen4x16 and Gen5x8 are about equal. Certainly is limiting, but 32 GT/s isn't too bad

Yeah, specially for a setup with only one or two of those it ain't that bad. But in a setup with 4x of those (so 8x GPUs in total) you might start to see bandwidth constraints while spreading layers across all GPUs.

Nonetheless, that won't be much of an issue for inference (which is what this GPU is meant for), training would be more of an issue but then you'd be lacking any kind of horsepower to train meaningful models anyway.

[SirShanova]

Now we get an LTT vid on it!
 

 

[Fasterthannothing]

I'm just here to get excited I can get a GPU with 16gb of ram for $300 to slap in my Plex system and really push transcoding limits. 

[igormp]

30 minutes ago, Fasterthannothing said:

I'm just here to get excited I can get a GPU with 16gb of ram for $300 to slap in my Plex system and really push transcoding limits. 

That Pro B50 should be what you're looking for, and should be cheaper than that.

[Fasterthannothing]

30 minutes ago, igormp said:

That Pro B50 should be what you're looking for, and should be cheaper than that.

I saw $299 but that would be great if it's ever cheaper. 

[igormp]

44 minutes ago, Fasterthannothing said:

I saw $299 but that would be great if it's ever cheaper. 

Seems like you're right, idk where I got the idea of a lower price from.

[Shimejii]

5 hours ago, Poinkachu said:

GamersNexus made a video tearing one of those down

Kinda wished he brought Buildzoid there, you know, then we'll see BZ rambling solo for an hour while looking at nothing but the PCB.

 

 

buildzoid is dealing with family matters atm and wouldnt be able to go sadly. Buildzoid posted a video about it. Also has GN worked with buildzoid at all recently? Its been quite a few years since they last collaborated o.o

[Poinkachu]

32 minutes ago, Shimejii said:

buildzoid is dealing with family matters atm and wouldnt be able to go sadly. Buildzoid posted a video about it. Also has GN worked with buildzoid at all recently? Its been quite a few years since they last collaborated o.o

Yeah, I watched his last vid explaining it.
I just thought it'd be nice if someone from big tech channel bring him along there, even if it's just to hear him ramble / rant unfiltered when seeing silly stuffs / interesting stuffs.

[IAmAndre]

Now it all boils down to software compatibility and driver optimization. It looks very exciting on paper but it’ll be interesting to see how they compete against Nvidia cards in 3D rendering and AI workloads. In Blender at least Nvidia has been so far ahead lately with their CUDA cores and driver optimization so I’m curious to see how the extra memory can narrow the gap.

 

Did they announce a release date ?

[igormp]

1 hour ago, IAmAndre said:

Now it all boils down to software compatibility and driver optimization. It looks very exciting on paper but it’ll be interesting to see how they compete against Nvidia cards in 3D rendering and AI workloads. In Blender at least Nvidia has been so far ahead lately with their CUDA cores and driver optimization so I’m curious to see how the extra memory can narrow the gap.

 

Did they announce a release date ?

Software wise it usually does better than AMD but worse than CUDA, almost right in between.

[IAmAndre]

39 minutes ago, igormp said:

Software wise it usually does better than AMD but worse than CUDA, almost right in between.

Indeed. Now it will be interesting to know how the dual B60 does in comparison with a similarly priced RTX 5080 in Blender.

[StDragon]

Uses GDDR6. It's "ok" for AI.

The real secret sauce is in HBM to widen that Von Neumann bottleneck. Good luck funding that cheap as there's a shortage until at least 2026.

[Mark Kaine]

On 5/19/2025 at 4:59 PM, Poinkachu said:

GamersNexus made a video tearing one of those down

Yeah, but are they faster than the regular Intel GPUs? Let's be real the only good reason to be excited is if they're cheap 4070 super etc alternatives ~

[igormp]

3 hours ago, Mark Kaine said:

Yeah, but are they faster than the regular Intel GPUs? Let's be real the only good reason to be excited is if they're cheap 4070 super etc alternatives ~

Those are not meant for games to begin with.

[Mark Kaine]

1 hour ago, igormp said:

Those are not meant for games to begin with.

Ok,is it faster than an A4000?

 

(I watch too much DDTS ¯⁠\⁠_⁠(⁠ツ⁠)⁠_⁠/⁠¯)

[igormp]

2 hours ago, Mark Kaine said:

Ok,is it faster than an A4000?

 

(I watch too much DDTS ¯⁠\⁠_⁠(⁠ツ⁠)⁠_⁠/⁠¯)

I don't think you should be comparing a dual GPU with 48GB of vram to a 16GB one, different purposes.

It can be faster or slower depending on what you meant to use it for. The dual A60 should also be cheaper.

[Dabombinable]

I wonder what the performance would be in a game with (at this point) theoretical integration of DX12's mGPU support.

[williamcll]

On 5/24/2025 at 9:27 AM, Dabombinable said:

I wonder what the performance would be in a game with (at this point) theoretical integration of DX12's mGPU support.

I don't think this has a lot of rendering cores.

[Dabombinable]

29 minutes ago, williamcll said:

I don't think this has a lot of rendering cores.

Each B60 is a B580 configured for professional use, and as there are dual versions...

[leadeater]

On 5/24/2025 at 1:27 PM, Dabombinable said:

I wonder what the performance would be in a game with (at this point) theoretical integration of DX12's mGPU support.

Sadly DX12 mGPU went nowhere, the only game I have that supported it was Civ 5?6? (forget which)