Monday, September 5th 2022
Resizable-BAR a Must for Arc "Alchemist," Stick with Other Vendors if you Lack it: Intel
The PCI resizable-BAR feature is an absolute must for Intel Arc "Alchemist" graphics cards to perform as advertised, to the extent that Intel recommends sticking to "other vendors" (NVIDIA, AMD), for those on machines that lack resizable-BAR or prefer it disabled. In our testing of the Arc A380, we found a big performance gap between resizable-BAR being enabled and disabled. The company said that while it is working on driver optimizations that improve performance for machines lacking resizable-BAR, its general advice for those on older platforms is to stick with other vendors.
Resizable-BAR enables the software to see the entire video memory of a graphics card as a single large addressable block, rather than through 256 MB apertures. AMD and NVIDIA's driver architectures have optimized their memory-management to cope with these apertures through the advent of PCI-Express, but Intel Arc hasn't. Its memory-management model relies on large bursts of memory transfers, for which it needs resizable-BAR. The performance penalty for lacking it could be as high as 40 percent. In related news, Intel Graphics confirmed that the Arc A770 will be available both in 16 GB and 8 GB memory variants at launch—which is still slated for Soonuary, 2022.
Sources:
PCGH, DigitalFoundry (YouTube), VideoCardz
Resizable-BAR enables the software to see the entire video memory of a graphics card as a single large addressable block, rather than through 256 MB apertures. AMD and NVIDIA's driver architectures have optimized their memory-management to cope with these apertures through the advent of PCI-Express, but Intel Arc hasn't. Its memory-management model relies on large bursts of memory transfers, for which it needs resizable-BAR. The performance penalty for lacking it could be as high as 40 percent. In related news, Intel Graphics confirmed that the Arc A770 will be available both in 16 GB and 8 GB memory variants at launch—which is still slated for Soonuary, 2022.
53 Comments on Resizable-BAR a Must for Arc "Alchemist," Stick with Other Vendors if you Lack it: Intel
ARC A770/A750 announcement and launch will be within September.
ARC LE edition will be in stock clocks, so no overclocking implemented and the availability intent is to be limited so the partners (that decided to work with Intel) can take the majority of the sales (probably a little bit later it doesn't seem to be concurrent availability for partner models and LE cards and the number of partners is limited despite Intel's efforts to reach and work with all of them)
There will not be a A5X0 announcement at the launch event and the 5 series will be based on a different die ACM-G12 that it may come on a later date (it doesn't exclude a limited ACM-G10 batch), so essentially not just a cut-down ACM-G10 like many sites were saying and probably not A580 but A570 or A550 (my guess, they may have scraped A580).
No 8GB A770 from Intel directly only from partners and the 16GB model's price is going to be aggressive so it doesn't seem to make sense a 8GB one because possibly the difference they are giving the 16GB one from the 8GB one, is essentially their cost difference regarding memory.
They are working with MS and Nvidia to try to standardize AI based upscaling in a similar way with what DXR did to raytracing.
They are in talks with MS and MS with the other parties involved but can't agree on a single codec regarding Direct Storage GPU compression (so maybe why the delay)
Promised roughly per month major driver updates but it seems it won't be the case, probably on average per 1.5X month based on his reaction to AMD's update schedule and wording of his answer.
If I would dare a pricing prediction it would be $349 SRP for A770 16GB and $279 SRP for A750
What?
Better than Maybember, I guess. :)
Now, if only the driver optimization could just be at a level that competes with AMD (nevermind nvidia for now), that would be great.
Having said that, the biggest problem that will result in the failure of the first gen ARC GPU is the timing. We are now at the last month of Q3 2022, and the product is not even available in store. With Nvidia likely to kick off the next gen GPU announcement soon, whatever excitement with the ARC GPU will fade away very quickly, if there is any excitement at all now.
About the Rebar. Just because Intel enabled it for older graphics (Intel does not have a lot except the iGPU) it literally doesn't change a thing.
NV and AMD is so far ahead of Intel that this cannot even be any sort of dispute. I don't think timing has a major role here. Intel's GPU is competing with AMD and NV because they make graphics as well but the product does not compete in any of these: performance, features, reliability etc. People will compare Intels to other graphics vendors but the comparison is pointless. It has to be released and it is but it is a poor gpu to be fair. Kudos for trying but Intel you need to put even more effort since you are in the worse position and you don;t have an upper hand.
Rumor mill says that Intel is dropping the battlemage or whatever the next GPU is? That also says something.
And as always- don't buy 'Alchemist'. Wait for reviews on 'battelmage'.
Buy an Arc card just because it's "cheap"? Well, what about DX11 games and other games that Intel claims to lack optimisation?
And for a product like this, I don't know whether its driver is gonna be any good, and neither do I know whether this product would be soon abandoned and end support.
Well that is the gamble isnt it?
a handful of games will be tested by reviewers and Intel promises to continue support.
so you will know of a few that will work and maybe those are what you play.
We can expect future games to work well, just past games might be bad but there is a list out there already with games that work well and those that dont and I expect that list to grow.
The upside? well they claim best price performance for the stuff that does work, so up to the consumer.
I have an RX480 so a card that can match a 3060/rx6600 would be a nice upgrade for me, but we also have new cards from nvidia and AMD on the horizon so ill be waiting to see what those will be like and how that might affect Arc pricing, then ill make a choice on what to get.
i have it enabled and works just fine
and AMD rig with Nvidia GPU RE-BAR do work with the latest BIOS, tested with a friend who has the same Mobo but with a 5800X and a RTX 3070Ti
well that's it for me ... the only interest i had in ARC was a LP A380 or lower model to replace a GT730 :laugh: on a Q77 express motherboard (no RE-BAR hehe ... ) RX 6400 4gb it is then
also one ARC review here mentioned it was working well with AMD rigs :laugh:
www.techpowerup.com/review/intel-arc-a380/
I'm still convinced Intel took him hostage and he's trying his best at signalling us to save him.
The tech journalists are apparently reporting verbatim as if what Intel Tom ex Nvidia and Ryan ex PC Per scandal and Shrout research were unquestionably true.
1) When you consider what Ryan put out Q1 about DX11 in-game as Alchemist was being fabbed, you smell BS here. Ryan claimed big wins.
2) Intel Tom had claimed Arc was designed for Re-BAR but no mention was made prior to the debacle. The memory transfers have to go over a bus to arrive at dGPU .. if you look into PCIE you'll find payload sizes of 4K which would allow error correction and bus sharing.
Don't you think NVME drives wouldn't like "long burst mode transfers" to/from DRAM buffers?
256MB apertures were standard when Arc was designed, nobody used Re-BAR, yet Intel AXG were going to bet their GPU range around it?
3) Intel Tom contradicted his own statement live on PC World show suggesting the memory controller had pre-existed and they hadn't seen the significance.
That's throwing the design engineers under bus, like the driver people, anyone but the management of delayed Arc project.
Ryan Shrout is notorious for slanted benchmarks to mislead people.
Those suspecting the delays and Cannon Lake China only launch was to hide performance problems they couldn't fix in software have good reasons to doubt the. AXG marketing people.
The only real actual useful contribution I saw him do in these interviews was trying to stop Tom Petersen from overpromising regarding what OC clocks A770 can achieve, I wonder how many things from what Petersen says he actually understands and how much he pretends to.