News Intel Arc Battlemage GPU slips past Arc A750 in new benchmark — G21 GPU surfaces with 20 Xe2 Cores up to 2.85 GHz and 12GB VRAM

[Admin]

A prototype of an upcoming Battlemage GPU was tested and sadly only performed as well as an Arc A750 in Geekbench’s OpenCL test

Intel Arc Battlemage GPU slips past Arc A750 in new benchmark — G21 GPU surfaces with 20 Xe2 Cores up to 2.85 GHz and 12GB VRAM : Read more

 

[Eximo]

A750 has 28 Xe cores. So if 20 Xe cores of Battlemage is roughly equivalent, that is a decent 40% jump in performance. Seems to be mostly from the clock speed.

This looks to be the G11 successor. So potentially still a 96bit bus, just larger memory modules to get to 12GB. (Though it could be 192 bit, doesn't seem likely)

A770 has 32 Xe cores on a 256 bit bus and was the G10 chip. Trades blows with the 3060Ti and 3070 in gaming.

If they stick with a doubling that seems likely, G20 should have up to 64 Xe cores on the same 256 bit bus.

With a 40% uplift that would place it around 4070 Ti or 4070 Ti Super range, which is a little above the expected 4070 range of performance from way back in the day. But it will vary from title to title.

 

[thestryker]

This looks to be the G11 successor. So potentially still a 96bit bus, just larger memory modules to get to 12GB. (Though it could be 192 bit, doesn't seem likely)

All 3 memory vendors still list a maximum of 16Gb density for GDDR6 so if it was 96-bit that would mean clamshell.

 

[jlake3]

Battlemage is obviously still in pre-production form, and the software for it is probably not finalized.

Intel reportedly plans to launch Battlemage before the holiday season, so it shouldn’t be long until we see what Intel brings to the GPU market.

The writers here on Tom's keep signing off every piece about Battlemage rumors saying we're going to see Battlemage before the holiday season, but is that realistic if it's currently in "prototype" or "pre-production" form?

November is five and a half weeks away. At this point I feel AIBs would have to be done with design and tooled up, and actively running cards and putting them on boats if they wanna get them on shelves in November. Boat shipping from China is what, 35 days? Plus getting them through the distribution network of your retailers.

They could do some or all of; launching with reference designs only, shipping them air-freight at a higher cost, having a low initial supply, or squeaking them out at the last minute in mid-December and claiming they were still "before the holidays", but it feels like 2024 is increasingly unlikely and could feel kinda like a paper launch if there's a launch at all.

 

[watzupken]

A750 has 28 Xe cores. So if 20 Xe cores of Battlemage is roughly equivalent, that is a decent 40% jump in performance. Seems to be mostly from the clock speed.

This looks to be the G11 successor. So potentially still a 96bit bus, just larger memory modules to get to 12GB. (Though it could be 192 bit, doesn't seem likely)

A770 has 32 Xe cores on a 256 bit bus and was the G10 chip. Trades blows with the 3060Ti and 3070 in gaming.

If they stick with a doubling that seems likely, G20 should have up to 64 Xe cores on the same 256 bit bus.

With a 40% uplift that would place it around 4070 Ti or 4070 Ti Super range, which is a little above the expected 4070 range of performance from way back in the day. But it will vary from title to title.

I do think it should come with a 192 bit bus in order to be the likely replacement of the A750/ A580. The A770 in most cases is behind a RTX 3060 Ti, so I won't think they are trading blows. In synthetic benchmarks, the A770 is almost as fast as a RTX 3070, but in actual performance, its generally between a RTX 3060 and 3060 Ti. An example is Horizon Zero Dawn where my A770 is not faster than my RTX 3060.

To me, the biggest problem with Intel's GPU now is the driver optimization. The performance in different games whipsaws a lot. I feel Intel has done a lot, but being the new kid on the block, they will need more time to address this.

 

[dalek1234]

This looks like discrete Battlemage. Aren't those cancelled because hardly anybody bought the previous gen and Intel is just losing money on discrete GPUs?

 

[Eximo]

This looks like discrete Battlemage. Aren't those cancelled because hardly anybody bought the previous gen and Intel is just losing money on discrete GPUs?

The tech also goes into their AI accelerators, so the division was not shut down.

Yes, it is likely Intel sold all of the Alchemist cards at a loss.

 

[Eximo]

I do think it should come with a 192 bit bus in order to be the likely replacement of the A750/ A580. The A770 in most cases is behind a RTX 3060 Ti, so I won't think they are trading blows. In synthetic benchmarks, the A770 is almost as fast as a RTX 3070, but in actual performance, its generally between a RTX 3060 and 3060 Ti. An example is Horizon Zero Dawn where my A770 is not faster than my RTX 3060.

To me, the biggest problem with Intel's GPU now is the driver optimization. The performance in different games whipsaws a lot. I feel Intel has done a lot, but being the new kid on the block, they will need more time to address this.

If you look at the G21 with 20 Xe cores as a doubling of the previous generation, then it is still a relatively 'small' part. G11 is the size of an RTX 4060 die. This is the same class of chip that was often seen in mobile solutions as well (though some where cut in half G10 chips, interestingly). A580 and up all used the bigger chip. I think it would be really nice of them to use a 192bit bus on their lower end products, but I don't expect it. Probably a 128bit bus on the top end (though all the 128 bit options were cut down G10 chips)

It really depends on the title. Some games it gets up to 3070 territory, some games it falls well below a 3060. Usually the DX11 titles suffer, but Vulcan and DX12 titles do quite well. More popular games got direct driver attention and are often better, but clearly some game engines just aren't suited to Alchemist. Hopefully the architectural changes Intel made after the lessoned they learned help with more general game performance.

 

[systemBuilder_49]

To me, the biggest problem with Intel's GPU now is the driver optimization. The performance in different games whipsaws a lot. I feel Intel has done a lot, but being the new kid on the block, they will need more time to address this.

False. Intel blew the design - no amount of optimization in ARC drivers will fix a bad design. Maybe they fixed the #1 bottleneck in Battlemage. All i know is that A770 is a gigantic chip, it was supposed to beat 3070, it was 2 years late, and it achieved ~3060 performance, hence, Intel was 4Y behind when A770 was releaseed. The optimism over Battlemage is misplaced.