News Intel's Arc GPU Roadmap Leaked: Two Battlemage GPUs Coming in 2024

[Admin]

Intel's Battlemage family reportedly includes two GPUs, targets 150W and 225W power envelopes.

Intel's Arc GPU Roadmap Leaked: Two Battlemage GPUs Coming in 2024 : Read more

 

[thestryker]

Unless there's something easy Intel can to do resolve the hardware issues holding Alchemist's performance back I don't see how this can still be an accurate roadmap. Raja hasn't said a lot, but admitted there were some design problems that arose from assumptions the team had made based on their experience with IGPs. As it stands though Alchemist uses too much silicon for the performance (and thus price) level it's at.

 

[jkflipflop98]

I say bring it on! A 3rd player is exactly what has been needed in the GPU space for a long time. Hopefully they can knock Nvidia's stratospheric nutcase pricing structure down a peg if nothing else.

 

[JayNor]

The only hardware problem I heard discussed was that they shouldn't have required the PCIE Resizable Bar.

They could update PCIE4 to PCIE5 and get rid of Resizable Bar requirement. If their performance really is being limited by the CPU to GPU bandwidth, this seems like a no-brainer ... especially with the PCIE5 support already on Alder Lake and Raptor Lake.

 

[thestryker]

The only hardware problem I heard discussed was that they shouldn't have required the PCIE Resizable Bar.

They could update PCIE4 to PCIE5 and get rid of Resizable Bar requirement. If their performance really is being limited by the CPU to GPU bandwidth, this seems like a no-brainer ... especially with the PCIE5 support already on Alder Lake and Raptor Lake.

ReBar is definitely a part of the problem with Alchemist, but that isn't what is holding the hardware performance back. Raja spoke of other things in broad terms throughout this video, but about half way through this video is where some micro benchmarking is shown

View: https://youtube.com/watch?v=GCsy75Mtg5Y

 

[-Fran-]

Q1 '24? That is a LONG time away from now... And that's assuming they actually deliver on time and not the "we released on China only, so we meet the target!" type of shenanigans.

Let's hope they can compete with Alchemist refresh a bit better and hold on for a bit longer. We need the competition, even if it comes from such a scary place xD

Regards.

 

[bit_user]

For me, further confirmation of an Alchemist refresh is the more interesting news. I was on the fence about getting the A770, but reports of high idle power made me reconsider.

I do want an Intel GPU, since they have the best OpenCL support, and given that Battlemage won't land for > 1 year, I hope the Alchemist refresh brings enough improvements to make it a compelling option for me.

 

[bit_user]

As it stands though Alchemist uses too much silicon for the performance (and thus price) level it's at.

IMO, they went too heavy on the deep learning & ray tracing acceleration. If they'd used a similar weighting as AMD, I think it'd be more competitive at rasterization performance. I'm not saying it'd be equal, but that explains some of the gap.

I say bring it on! A 3rd player is exactly what has been needed in the GPU space for a long time. Hopefully they can knock Nvidia's stratospheric nutcase pricing structure down a peg if nothing else.

That's some good team spirit!

 

[bit_user]

The only hardware problem I heard discussed was that they shouldn't have required the PCIE Resizable Bar.

We know there's some sort of issue with idle power.

Having done some in-house firmware development, myself (not for any companies in this sector), it's almost certain there are chip bugs necessitating performance-robbing workarounds. That's what happens when you're not designing to an immutable ISA spec with a massive legacy codebase that you can't break. The hardware goes out the door once the absolute show-stopper bugs are fixed or mitigated and the burden of any remaining hardware bugs falls to the firmware team.

I once had a chat with an AMD driver developer who confirmed they do indeed face these sorts of problems.

They could update PCIE4 to PCIE5 and get rid of Resizable Bar requirement. If their performance really is being limited by the CPU to GPU bandwidth, this seems like a no-brainer ... especially with the PCIE5 support already on Alder Lake and Raptor Lake.

Why do you believe PCIe 4.0 is a bottleneck? I've combed through Tech PowerUp's PCIe scaling analysis for the RTX 4090 and even the outliers gain < 10% by going from PCIe 3.0 -> PCIe 4.0 on such a fast card.

NVIDIA GeForce RTX 4090 PCI-Express Scaling

The new NVIDIA GeForce RTX 4090 is a graphics card powerhouse, but what happens when you run it on a PCI-Express 4.0 x8 bus? In our mini-review we've also tested various PCI-Express 3.0, 2.0 and 1.1 configs to get a feel for how FPS scales with bandwidth.

www.techpowerup.com

Also, consider how the RX 6500XT beats the A380, in spite of having a PCIe 4.0 interface only half as wide (also 33% less GDDR memory, not to mention a smaller die)!

If the A770 truly is being held back by PCIe 4.0, it suggests they're doing something wrong.

 

[-Fran-]

Well, Intel GPUs are still in severe need of Linux driver optimizations, so if you want to use Linux, you'll still be subjected to problems:

https://www.phoronix.com/review/a770-windows-linux

That being said, I think all the compute stuff is working fine?

Regards.

 

[Eximo]

Still have two common problems with my A380.

Failure to detect 'new' HDMI audio devices, requires the disabling and enabling the HD Audio driver from device manager. Happens after almost every wake up unless I absolutely get the speakers powered up before turning on the display. It is like it just can't figure out that there is now audio being transmitted unless it was already happening, absolutely stupid.

Full black screen when certain types of media are displayed. This one is more intermittent, but seems to be related mostly to advertisements on websites, even Youtube on occasion. Screen will go completely black until the offending ad is off the screen, but recovers completely. Sometimes this means alt-F4 to kill the browser if it can't be scrolled away from.

I do need to try out some DX9 titles though. Last time I had some serious stutters in even very light games.

 

[bit_user]

Failure to detect 'new' HDMI audio devices, requires the disabling and enabling the HD Audio driver from device manager.

HDMI auto-negotiation seems to be a never-ending source of issues, with all manner of devices. I wonder if the HDMI specs include good reference code & pseudo-code for doing it properly. They need to make it so simple even a trained monkey could get it right. That, and robust conformance testing.

Full black screen when certain types of media are displayed. This one is more intermittent, but seems to be related mostly to advertisements on websites, even Youtube on occasion.

Heh, probably happens when enabling HDCP, and therefore could well be another HDMI auto-negotiation issue.

I've had problems like this with an old DVI monitor and different GPU. With problems like this, it's not always obvious which party is at fault, and sometimes the issue could be specific to the version of HDCP supported by the display.

 

[Eximo]

HDMI auto-negotiation seems to be a never-ending source of issues, with all manner of devices. I wonder if the HDMI specs include good reference code & pseudo-code for doing it properly. They need to make it so simple even a trained monkey could get it right. That, and robust conformance testing.

Heh, probably happens when enabling HDCP, and therefore could well be another HDMI auto-negotiation issue.

I've had problems like this with an old DVI monitor and different GPU. With problems like this, it's not always obvious which party is at fault, and sometimes the issue could be specific to the version of HDCP supported by the display.

Never had issues before with a whole slew of AMD and Nvidia cards on the HDMI audio thing.

Forgot to add that some games also have this problem. They just won't have sound until you restart the game a time or two. Like it can't route the sound to the right place or something.


I have experienced that black screen issue on Iris Xe laptops as well, so it is either their bad drivers or something inherent in the Xe architecture, lucky for me my daily laptop also has an Nvidia MX in it, so I just default the browser to run on that, problem solved.

HDCP implementation could be it though, that seems to track with the recent behavior in Youtube. When content that is likely copyrighted plays, it will blink black for a second.

 

[bit_user]

Never had issues before with a whole slew of AMD and Nvidia cards on the HDMI audio thing.

I have an old Skylake laptop that I used for streaming movies to my TV. About 3% of the time, the laptop will fail to negotiate a connection with the TV, when I switch it to that input. Less than 1% of the time, it will fail to negotiate audio with the TV.

I should add that the laptop is usually coming out of sleep, at the time, as I never shut it down. OS is Linux.

 

[InvalidError]

They could update PCIE4 to PCIE5 and get rid of Resizable Bar requirement. If their performance really is being limited by the CPU to GPU bandwidth, this seems like a no-brainer ... especially with the PCIE5 support already on Alder Lake and Raptor Lake.

Why get rid of ReBAR? As AMD and Nvidia's drivers become more ReBAR-optimized, their GPUs are showing increasing benefits from having ReBAR enabled and a growing number of games are showing 10+% performance gains from eliminating the BAR management overhead and latency.

Also, since all PCIe5 platforms have ReBAR, it makes no sense to wish for PCIe5 without it.

 

[InvalidError]

One thing that looks rather odd is that Intel apparently plans (or at least planned) to refresh its Arc Alchemist lineup with new "Alchemist+" ACM+ G20 and ACM+ G21 graphics processors in the second half of 2023. This doesn't make a lot of sense as in H2 both AMD and Nvidia will likely introduce their next-generation entry-level and mainstream GPUs that will leave Intel's Arc Alchemist offerings behind, and investing a lot of money in a refreshed Alchemist+ lineup might be a losing proposition.

The A380 could definitely use a refresh bump to 8GB/128bits GDDR6 and ~30% more compute grunt. Toss in fixes for whatever hardware-level quirks have been identified, you'd probably get something Intel could sell for ~$170 US, especially in countries like Canada where there is nothing else available new under $300. Nvidia and AMD's current-gen pricing being what it is, they are lining up for ludicrous $400+ "low-end" GPUs.

Intel supposedly aiming for enthusiasts with BMG will likely end up in humiliation as two years probably won't be enough to polish drivers especially if they skip a year of hardware improvements to field-test their driver improvements from. ATI, Nvidia, 3dfx, S3, etc. didn't develop their drivers and hardware overnight. Back when 3D was new-ish, they were iteratively releasing new or tweaked architectures every 6-8 months to leverage the stuff they learned since the last launch and making 10-30% performance gains each time.

 

[SyCoREAPER]

Intel really botched 'shaking up the field'. Glad we have a third player but that third player is playing on crutches while blindfolded.

Intel needs to stop and make sure these two GPUs are near perfect and a card worth getting at the right price.

Console analogy, right now Nvidia and AMD are Microsoft and Sony while Intel is Nintendo. Cutting edge vs behind but innovative.

I want to root for them but they are making it hard.

 

[Eximo]

The A380 could definitely use a refresh bump to 8GB/128bits GDDR6 and ~30% more compute grunt. Toss in fixes for whatever hardware-level quirks have been identified, you'd probably get something Intel could sell for ~$170 US, especially in countries like Canada where there is nothing else available new under $300. Nvidia and AMD's current-gen pricing being what it is, they are lining up for ludicrous $400+ "low-end" GPUs.

I think that is basically the unreleased A580 which had an expected launch date of Q4 2022, made from the same GPU as the A750 and A770. Perhaps yields were too good to justify making it? That or they didn't want to cripple sales of the A380 partner cards.

They also seem to have snuck in an A310 which is a 4GB cut down A380, doesn't seem to be anywhere. Maybe OEM only?

 

[InvalidError]

I think that is basically the unreleased A580 which had an expected launch date of Q4 2022, made from the same GPU as the A750 and A770. Perhaps yields were too good to justify making it? That or they didn't want to cripple sales of the A380 partner cards.

The A580 uses the same 400sqmm die as the A750+. That would be a pretty large chunk of silicon with ~70% of it disabled for a sub-$200 GPU which typically have dies in the 150-200sqmm range. Buffing the ACM-G10's refresh (A390?) to a more palatable level would likely bump it from 157sqmm to about 210 and its rejects could still be used as A380 substitutes.

 

[Eximo]

Not sure I take your meaning. I wasn't taking about taking G10 and making them into G11 class chips (Intel has already done that with the Arc 550m, has half the core count)

I was saying they didn't make them because the pricing wouldn't make sense. Why buy a A380 when you can get an A580 for just a little more...

A380 $140
A580 ($200?)
A750 $265
A770 8GB $305
A770 16GB $350

 

[InvalidError]

Not sure I take your meaning. I wasn't taking about taking G10 and making them into G11 class chips (Intel has already done that with the Arc 550m, has half the core count)

The G11 is a 400sqmm die. That is ~$100 of silicon and the A580's specs would scrap almost half of it. Can't put that on a $200 GPU and expect to make a profit. I doubt the defect density rate on the A770 is such that many still usable dies grade worse than the A750.

Cancelling the ACM-G11 A580 would be a perfect occasion to replace it with a new DG2-256 'refresh' one at ~250sqmm instead. As I have written before, a solid $200 contender should have been Intel's main focus for its first attempt at re-entering the dGPU space.

 

[cyrusfox]

The G11 is a 400sqmm die

You have that backwards, G10 is the bigger die, G11 the smaller for Alchemist
https://cdn.wccftech.com/wp-content/uploads/2022/03/2022-03-30_15-32-45-1030x579.jpg
The A580 was always based on G10(Cut down big die 24/32 XE).

 

[watzupken]

Q1 '24? That is a LONG time away from now... And that's assuming they actually deliver on time and not the "we released on China only, so we meet the target!" type of shenanigans.

Let's hope they can compete with Alchemist refresh a bit better and hold on for a bit longer. We need the competition, even if it comes from such a scary place xD

Regards.

I was thinking the same thing. It is too late, unless they are going to introduce a GPU that is faster than the current top end cards. 2024 should be round about the year that a new generation of Nvidia/ AMD GPUs will get announced. So Battlemage is going to end up just like the current gen ARC that came out way too late. At this rate, I think Intel's GPU division is going to be in trouble because they just cannot have any competitive advantage being almost 1 generation behind their competitors. I have doubts as to how well their current gen GPUs are selling to date.

 

[InvalidError]

So Battlemage is going to end up just like the current gen ARC that came out way too late. At this rate, I think Intel's GPU division is going to be in trouble because they just cannot have any competitive advantage being almost 1 generation behind their competitors.

Battlemage is likely DOA as a high-end product, I give Intel a practically zero chance of Intel nailing it on the second try after missing the mark by over a year on Alchemist. I thought Intel saying BM was still on schedule meant 2023, 2024 will certainly be too little too late unless BM scales way more aggressively than anything ever launched before, which would smell like a moonshot before scrapping the whole idea for missing the impossible target again.

If AMD and Nvidia's trend of pissing on the low-end continues though, that field will be wide-open for Intel to seize if it wants to. IMO, this is really what it needs to focus on until it sorts its crap out. Lower-risk, lower-rewards but also a more forgiving audience instead of a spectacular high-stakes failure.

 

[InvalidError]

You have that backwards, G10 is the bigger die, G11 the smaller for Alchemist

I've never been able to keep names straight and likely never will. In any case, the big die is far too expensive for the A580 if Intel only manages to get ~$200 out of it. It needs a more cost-effective variant.