Intel Arc B580

[lexluthermiester]

Hardware Canucks are not in the same league as HUB for what you describe.

I was not referring to HC, I know very little about them. HU however is a different story. Their data/info is dubious at times and as a result can not be trusted.

 

[Lew Zealand]

I haven't found HUB's data to be significantly different than any other place that does a lot of CPU/GPU testing. All places' data are skewed based on where in-game they test and their test system configurations. I disagree with all reviewers' conclusions fairly frequently because my preferred criteria often don't match a reviewer's (I own and enjoy a GPU that HUB quite dislikes) but I haven't found any of the major reviewers' data to be incorrect or misleading when I've bought a GPU or CPU.

 

[lexluthermiester]

I haven't found HUB's data to be significantly different than any other place that does a lot of CPU/GPU testing. All places' data are skewed based on where in-game they test and their test system configurations. I disagree with all reviewers' conclusions fairly frequently because my preferred criteria often don't match a reviewer's (I own and enjoy a GPU that HUB quite dislikes) but I haven't found any of the major reviewers' data to be incorrect or misleading when I've bought a GPU or CPU.

The data is sometimes less than consistent with the results of other testers. However you nailed the conclusions issue. Their conclusions and opinions are often so daft and nitwited as to be without any reason whatsoever. They're certainly without merit.

 

[10tothemin9volts]

I want a Arc 24GB VRAM clamshell (just like the 4060 Ti 16GB one is) version for AI inference and later, if there's going to be a Arc B700 series, a B750/B770 32GB VRAM clamshell version as well, because the 24GB VRAM are kinda not enough for fast inference to fit bigger LLMs.

Tbh, I think my (and maybe others') demands have changed since the release of the GeForce 5090 which has 32GB VRAM: To fit a decent 30B-32B LLM quant (e.g. Qwen2.5-32B-Instruct-Q6_K.gguf) in 1 GPU, because my other PCI(e) slots are used (maybe I could reshuffle stuff but then it wouldn't be optimal) or because I don't want to bother with 2 GPUs, one would need a Arc B770 / B780 32GB VRAM GPU clamshell version. So it would be nice if Intel priorized the 256bit clamshell version over the 192bit one.

Another way to make the Arc Battlemage lots-of-vram-for-ai clamshell maybe cheaper: Use GDDR7 (NV already show that despite GDDR7, 5070' price is still 549) and reduce the memory bus width from 256 to 192 bit to make the chip smaller and cheaper and use 3GB GDDR7 chips:
Before: 8 chips * 2GB per chip * 2 [clamshell] = 32GB VRAM (@ ~608 GB/s)
After: 6 chips * 3GB per chip * 2 [clamshell] = 36GB VRAM (@ ~672 GB/s)

 

[10tothemin9volts]

Can't take seriously anything from HU. Got a better source of info?

Why, any examples?

That's personal preference. You'd be surprised just how good a mix of medium/high settings looks. Finding the right balance takes some tweaking but it's worth it to get the most from the experience.

True. What I mean is ideally I'd just set everything on Ultra/High and forget about it, but modern games can still look great even at e.g. medium and the tweaking can be very worth it.

 

[lexluthermiester]

Why, any examples?

There are too many to list. That's not a topic we should traverse here.

but modern games can still looks great even at e.g. medium and the tweaking can be very worth it.

Exactly! Most modern games even look great on a combination of low & medium which can make for excellent framerates and very smooth gameplay.

 

[10tothemin9volts]

There are too many to list. That's not a topic we should traverse here.

Exactly! Most modern games even look great on a combination of low & medium which can make for excellent framerates and very smooth gameplay.

I understand that u don't want to go off topic (which I do by post this very reply, but I have to reply), but if there are really too many to list, you may crease a new topic, list them and maybe ppl want to discuss it too. The only recent thing I can remember that I don't like is HU (still) not testing the 9800X3D in 1440p native/no upscaling or 4K with DLSS Quality (internal resolution 2560*1440) and it's kinda ingenuine to do so, but I guess this is why we also have TechPowerUp. People who buy the 9800X3D are not going to play at 1080p and even their own monitor article says that 1440p is the new 1080p. (I understand their theoretical reason for why only 1080, but it's simply not too practical, they might as well just test in 720p because that's even less GPU bound, but then even more people would complain..)