News Intel Arc Battlemage GPU surfaces — BMG-G31 silicon reportedly wields 32 Xe2 Cores

[Admin]

Intel has confirmed the existence of a third Battlemage die, codenamed BMG-G31. The new die is suspected to be Intel's least-powerful die in the lineup, aimed at entry-level GPU solutions.

Intel Arc Battlemage GPU surfaces — BMG-G31 silicon reportedly wields 32 Xe2 Cores : Read more

 

[oofdragon]

Dont care. Give us 8800XT news!

 

[Metal Messiah.]

The site leads to technical specifications for VRTT interposer prototypes aimed at the BMG-G31 die. However, only certain people with the proper credentials can see the technical specs of this specific component.

Full tech specs aside, at least it does tell us that the die sports the BGA 3283 format. 3283 pins, which means the chip has a larger footprint than the other Battlemage chips, and previous gen Alchemist lineup.

• ACM-G10: BGA 2660 (Alchemist)
• BMG-G21: BGA 2362 (Battlemage)
• BMG-G10: BGA 2727 (Battlemage)
• BMG-G31: BGA 3283 (Battlemage)

The new die is suspected to be Intel's least-powerful die in the lineup, aimed at entry-level GPU solutions.

If Intel follows the same route as its Arc Alchemist GPUs, this new die will most certainly be the entry-level die for Intel's Battlemage discrete GPUs (consisting of gaming, mobile, and workstation GPUs).

this is the first time we've seen mentions of Intel's G31 die.

Well, as per Intel's logic it should be the smallest die, but it could be the other way as well ? Maybe we are looking at a top-end part here, because it sports 32 cores as per rumors ?

RedGamingTech shared a video few months back, unveiling some details on Battlemage GPU SKUs and the BMG-G31 chip.

As per the leaker, while the G31 die features 32 Xe2 cores and a 256-bit interface, the Battlemage BMG-G21 die will reportedly feature 20 Xe cores across a 192-bit bus interface.

So assuming this info is accurate, what will the top high-end G30 die sport ? 64 Xe2 Cores (assuming Intel has plans to go for a bigger/stronger die) ? Because, if we go by ARC Alchemist lineup, the top die "Arc ACM-G10" came with 32 cores.

Also, Intel has apparently cancelled the G10 variant. So either this G31 is the mid-range die, with the G21 being the lowest SKU in the lineup.

**thinks, thinks, thinks**

G10 and G21 could possibly target the entry-level desktop and mid-range markets.

The G10 SKU has been dropped from Intel's pipeline, as per reliable internal gossip/rumors, which leaves only the G21 and 31 dies as of now.

 

[InvalidError]

Battlemage BMG-G21 die will reportedly feature 20 Xe cores across a 192-bit bus interface.

Makes sense if Intel is aiming for $250-300 as it did with the A750. 8GB is kind of low for that price range and would get Intel crucified for VRAM stinginess even worse than AMD and Nvidia as the complete GPU under-dog that it is. 192bits for 12GB would be the sweet spot. And they can still axe 32-64bits for $200-ish models if they want to.

 

[Metal Messiah.]

The Battlemage GPU die codename is weird to say the least ! Confusing with odd numbers, G10 , G21, and now this new G31.

Alchemist had the following, ACM-G10, G11 and G12. G1* SKUs. But with Battlemage, we now have 1, 2 and 3 as the initial starting digits, lol. 🙄


But anyway, there are already tools which were being used for verification and testing purposes for other Battlemage discrete GPUs though.

Two test tools for Battlemage GPUs were previously listed on the same Intel DesigninTools webpage. BGA2362-BMG-X2 and the BGA2727-BMG-X3-6CH (this link is currently not working).

The X2 tool features a 2362 BGA array, and the X3 tool feature a 2727 BGA array. For reference:: ACM-G10: BGA 2660 (Alchemist).

So these should correspond to the following chips:

• BMG-G21: BGA 2362 (Battlemage)
• BMG-G10: BGA 2727 (Battlemage). But this seems to be have been dropped from production/testing.

 

[Amdlova]

I want ditch my 4060ti 16gb for a Intel graphics but mid range for midrange can't do it... need moar powa baby. Something hitting a 4070 super or better

 

[Metal Messiah.]

The new die is suspected to be Intel's least-powerful die in the lineup, aimed at entry-level GPU solutions.

If Intel follows the same route as its Arc Alchemist GPUs, this new die will most certainly be the entry-level die for Intel's Battlemage discrete GPUs (consisting of gaming, mobile, and workstation GPUs).

this is the first time we've seen mentions of Intel's G31 die.

Well, maybe NO, maybe Yes. But most likely this time things might turn out to be different. To further clarify my previous first post/OP entry, and your claim as well.

Assuming, the G31 is indeed the smallest entry-level die, then why is the BGA 3283 array package being used for it ?

Because if BMG-G10 is the bigger silicon, as we assume, then why does it support the BGA 2727 array instead, with only 2727 pins (see above post of mine)?

Confusing much, eh ? 🤐

 

[Metal Messiah.]

And they can still axe 32-64bits for $200-ish models if they want to.

You mean cutting on the memory Bus width to 160 or 128 bits ? Then that would also cripple its overall performance as well. For a 12GB GPU, a 192 bits bus width would be the most optimal choice.

But of course, Intel can make other variants with lower bus width with a different VRAM config, if there is a need, but this will also lead to more 'segmentation' in the same GPU market, upper mainstream, and/or mid-range, or whichever category we need to put these into.

And the price also needs to adjusted accordingly based on the then active/current market trend.

 

[thestryker]

Alchemist had the following, ACM-G10, G11 and G12. G1* SKUs. But with Battlemage, we now have 1, 2 and 3 as the initial starting digits, lol. 🙄

I have two thoughts on this:

1. Intel has decided to be rational and go bigger number = better
2. These could be die revisions and changing SKU configurations so 3 would be the newest die

Well, maybe NO, maybe Yes. But most likely this time things might turn out to be different. To further clarify my previous first post/OP entry, and your claim as well.

Assuming, the G31 is indeed the smallest entry-level die, then why is the BGA 3283 array package being used for it ?

Because if BMG-G10 is the bigger silicon, as we assume, then why does it support the BGA 2727 array instead, with only 2727 pins (see above post of mine)?

Confusing much, eh ? 🤐

Unless there's more GPU die we don't know about there's certainly no way that the biggest package is going to anything but the biggest chip. Alchemist was already a silicon/socket loser monetarily speaking so I cannot fathom them doing it again.

 

[TechyIT223]

32 cores on the lowest end GPU die this time? Strange!

 

[Alvar "Miles" Udell]

Intel doesn't have to compete with the Titan level 5090 or high end 5080, the margins are high but they're never going to lure people away from nVidia until they establish themselves. What they need is a $450 16GB card that competes with the mid range 5070 and sub-$250 8GB card that competes with the entry level 5060, much like AMD did with the HD 4870 and then HD 5870. They may not make much money but once they prove themselves viable then game studios will start taking them seriously.

 

[thestryker]

Just some random musing here as I was skimming LNL slides:

They mention 50% performance increase over MTL which have the same number of cores (albeit Xe instead of Xe2 and missing XMX). There was another slide which referred to microbenchmarking and a 1.2-12.5x performance increase over Xe (there were some nasty holes in microbenchmarking Alchemist shown by Raja in a video).

So hypothetically speaking if the 50% was comparing both at their best and the microbenchmarking means Xe2 has resolved the architectural problems it would mean 32 Xe2 cores could be close to 4070 Ti levels of performance. I don't think anyone knows for sure which manufacturing node Intel will be using for the discrete GPUs, but both N4/5 and N3 should shrink the die to about the same size as AD104 as well at worst (remembering that ACM-G10 was as big as GA104 despite using a better node).

Yes I do think there's certainly wishful thinking in there, but it's also a plausible performance area. Between optimizations, clockspeeds and memory bandwidth if Intel could deliver 4070 Ti Super levels of performance for around $400 I feel like that would be rather compelling even into the next generation of cards.

 

[Metal Messiah.]

Great ! Now this leak is gonna make things even more bleak !

Battlemage "X2" & "X3" graphics cards have been spotted. New entries spotted within customs and shipping log data from nbd.ltd and Volza.

The data says Intel is using X3 and X2 codenames for these GPUs. Not sure if these are internal codenames or some sort of branding.

The good old friend "BMG-G21" GPU also shows up here, and it would appear Intel seems to be testing a range of active-fan heatsinks with these ES/engineering samples.

https://www.nbd.ltd/customs-data?t=0&v=BMG ADD-IN-CARD


First of all, the cancelled BMG-G10 GPU has again been spotted sporting 448 EUs or 28 Xe2 cores. The entries all show an older year/date i.e. 2023. The data says the card would feature an 8-layer PCB.

BMG-G10 comes with a 256-bit bus interface and 16GB GDDR6/X memory. The SKU is listed under the yet to be finalized "Churchill Falls" platform. But I hope this card is not yet been cancelled though.

Here we can see entries for new X2 and X3 codenamed Blackmage cards.

BMG G21 entries. Intel seems to be testing active-fan heatsinks with these ES/engineering samples.

 

[usertests]

If they make another low-end die (A310/A380 successor) and it beats a GTX 970, I might get that.

If it has the same media capabilities as Lunar Lake (it doesn't have to since graphics and video have been disaggregated), then it would come with H.266 decode, which might be a curiosity for some.

 

[Metal Messiah.]

I have two thoughts on this:

1. Intel has decided to be rational and go bigger number = better
2. These could be die revisions and changing SKU configurations so 3 would be the newest die

That actually makes sense logically though. Incrementing the die numbers in a sequence. But again, the die codename might still be different than the actual GPU SKU branding name.

These new X2 and X3 entries are another mystery to add to this ongoing speculation. We currently have A770 and A750 GPUs in the alchemist series.

For Battlemage we can expect something like this: BGM B880, B870/850, B820, B810 etc. Just a guesstimate.

 

[TechyIT223]

I remember Intel's Tom Petersen saying that 30% of the Arc GPU team is working on Battlemage software while the hardware team has already fully switched to Celestial, during an interview to PCWorld back in January.

For now they hope to have Battlemage out around CES 2025. Tough competition stays ahead in front of them though.

We shall probably learn more on Battlemage during Intel's upcoming next Vision event, I think this coming September? 🤔🧐

 

[Metal Messiah.]

That's INTEL's annual Innovation event this year in San Jose, California, on the 24-25th of September, 2024. Not vision.

But I don't expect them to announce anything relate to BattleMage GPU lineup though. Expect some info on Arrow Lake, Lunar Lake, Xeon 6, & Gaudi 3 CPU lineups.

The focus is on AI PC, so the key role will definitely be on Lunar Lake. In addition, Intel Innovation 2024, which will be held in San Jose, USA on September 24 and 25, is more important, and it is confirmed that Intel will release the Intel Xeon 6 series processor code-named Granite Rapids-AP at this event, while Intel Gaudi 3 is expected to be released in Shipping starts in September.

via Benchlife (Machine Translated)

 

[thestryker]

Expect some info on Arrow Lake, Lunar Lake, Xeon 6, & Gaudi 3 CPU lineups.

Really hoping we get the architecture details on those at Hot Chips. I think there's a decent chance since retail versions of all should be coming out of the fabs by then (or already shipping) and Intel historically has been pretty open there.

 

[Metal Messiah.]

Yes. HOT CHIPS should really have some deep dive into the tech details of Intel's next-gen processors, and also on the company's Gaudi 3 AI accelerator.

Intel might have an opportunity to capitalize on the AI hype with Guadi 3 as well, given these accelerators have a price tag that makes them significantly more competitive against rivals.

But consumers won't simply pay for raw power. NVIDIA still has a stronger AI ecosystem, with specialized CUDA libraries, Grace CPUs, which makes their products suited for more widespread adoption in the markets.

https://twitter.com/x/status/1797832515945947625

View: https://x.com/TekStrategist/status/1797832515945947625/


But still, Intel is creating a full ecosystem of Gaudi 3 AI accelerators, offering several options which start at 8 accelerators, and scale up to 512, 4096 and up to 8192 super-clusters with 15 Exaflops of performance, 1 PB capacities and an insane 1.229 PB/s of bandwidth.

 

[Metal Messiah.]

But still, Intel is creating a full ecosystem of Gaudi 3 AI accelerators, offering several options which start at 8 accelerators, and scale up to 512, 4096 and up to 8192 super-clusters with 15 Exaflops of performance, 1 PB capacities and an insane 1.229 PB/s of bandwidth.

 

[InvalidError]

You mean cutting on the memory Bus width to 160 or 128 bits ? Then that would also cripple its overall performance as well. For a 12GB GPU, a 192 bits bus width would be the most optimal choice.

Talking as if the whole concept was somehow new when AMD and Nvidia have been doing it for almost 20 years already. Make one die, full die is the top trim in one bracket or bottom of the next bracket up, cut down versions populate the next 2-3 SKUs down from there, and then a "refresh" or Ti/Super variants may further shuffle those again later.

12GB on 192bits may be optimal for the upper-mid-range but there is still a plenty large market for $200ish 8-10GB GPUs. What makes the most sense for that? A bargain-basement GPU chip that is 2-3X more powerful than actually needed for people who just need extra video outputs just to make it viable at the lower-mid range or lobotomizing excess/defective G21s?

 

[Metal Messiah.]

Of course, the concept is not new. I didn't mean to say that, since they have been doing this for years. Just clarifying your previous post. Yes, there is indeed a large market for sub $200 GPUs. And it kinda also makes sense.

 

[TechyIT223]

That's INTEL's annual Innovation event this year in San Jose, California, on the 24-25th of September, 2024. Not vision.

But I don't expect them to announce anything relate to BattleMage GPU lineup though. Expect some info on Arrow Lake, Lunar Lake, Xeon 6, & Gaudi 3 CPU lineups.

Oops my fault. Yeah I meant innovation event. Got confused by the previous vision show.

🙃🙃