News China domestic gaming GPUs receive up to 40% performance uplift — new Moore Threads driver update improves S80 and S70 gaming performance

[Admin]

Moore Threads has published a new driver update that features up to a 40% performance improvement for its S80 and S70 gaming GPUs.

China domestic gaming GPUs receive up to 40% performance uplift — new Moore Threads driver update improves S80 and S70 gaming performance : Read more

 

[TCA_ChinChin]

If they keep up their driver improvements, even these seemingly weak cards can possible reach a point where they can actually replace entry level graphics from Nvidia and AMD. Curious how much more improvement they can get out of driver + software updates versus them reaching a hardware limitation.

 

[williamcll]

If they keep up their driver improvements, even these seemingly weak cards can possible reach a point where they can actually replace entry level graphics from Nvidia and AMD. Curious how much more improvement they can get out of driver + software updates versus them reaching a hardware limitation.

Very well pass Intel considering they have decent amount of VRAM

 

[jlake3]

Very well pass Intel considering they have decent amount of VRAM

Unless you're doing GPGPU work, good luck tapping into that extra VRAM. It has FLOPS that on paper put it in-between a RTX 3060 and RTX 3060Ti with memory bandwidth matching the Ti, but that has not translated into gaming performance. It's still underperforming the GTX 1650 after nearly two years of driver development, despite having all that compute and bandwidth and a 255W(!) TDP on tap.

 

[zsydeepsky]

the DX12 support is done by emulation, so even though it can "technically run" DX12 titles but the performance is terrible. the DX11 performance boost though seems real.

yet still, it's a good sign that they are still improving through software updates. in fact, the improvements like this news happened multiple times since the release of MTT S80, but it's the first time that I saw news report on English websites.

fun fact, if you go read the news about the same card back in 2023 Feb, you will find out that it only supported DX9 back then:
https://www.tomshardware.com/news/moore-threads-mtt-s80-tested-bullslab-jay
"Moreover, the card seemed to be restricted to DX9 gaming at the time of review (it is advertised as DX11-capable)"

 

[bit_user]

If they keep up their driver improvements, even these seemingly weak cards can possible reach a point where they can actually replace entry level graphics from Nvidia and AMD.

Did you actually look at the article? The supposed big gains are just game-specific optimizations for only two titles. This is nothing special.

AMD and Nvidia don't usually have such big title-specific wins, but that's probably because a lot of game-specific optimization already happened before launch, plus the game devs tend to test and optimize their stuff on Nvidia and AMD hardware in the first place. Nonetheless, they still release "game day" drivers, because game-specific optimizations are usually necessary and usually deliver significant wins.

Curious how much more improvement they can get out of driver + software updates versus them reaching a hardware limitation.

This hardware is always going to be dog slow. There are obviously bottlenecks, bugs, etc. holding it back from its theoretical potential.

Pin your hopes on their next gen hardware. That might actually stand a half decent chance.

 

[TCA_ChinChin]

Did you actually look at the article? The supposed big gains are just game-specific optimizations for only two titles. This is nothing special.

AMD and Nvidia don't usually have such big title-specific wins, but that's probably because a lot of game-specific optimization already happened before launch, plus the game devs tend to test and optimize their stuff on Nvidia and AMD hardware in the first place. Nonetheless, they still release "game day" drivers, because game-specific optimizations are usually necessary and usually deliver significant wins.


This hardware is always going to be dog slow. There are obviously bottlenecks, bugs, etc. holding it back from its theoretical potential.

Pin your hopes on their next gen hardware. That might actually stand a half decent chance.

I mean Intel made decent improvements in optimization with their GPUs, even in terms of individual games. Yes they're individual games, but do improvments for individual games not matter? I also didn't say they were anything special in the first place? The only reason I bring up future driver and software (games included) improvements is cause Moore's Threads is a newcomer to GPUs so they might be able to improve a lot more in terms of software and game specific optimization.

In fact, its exactly how AMD & Nvidia still optimize and have "game day" drivers for specific games that leads me to think Moore's Threads have a lot more performance they can extract from their current hardware. If AMD & Nvidia can deliver software improvements even after so much prior optimization, shouldn't Moore's Threads be able to do even more on the software side, since they're so new compared to AMD & Nvidia Reminds me of how Intel improved their entry level graphics from launch till now. They a lot a better in terms of performance on their same graphics silicon.

Sure they're are plently of hardware improvements to be made, especially since it's the first couple of gens for Moore's Threads, but why dismiss large game driver improvements?

 

[zsydeepsky]

This hardware is always going to be dog slow. There are obviously bottlenecks, bugs, etc. holding it back from its theoretical potential.

Pin your hopes on their next gen hardware. That might actually stand a half decent chance.

I agree, however, in @TCA_ChinChin 's defense, he was only asking for replacement of "entry-level" card.
the MTT S80 priced in China at about $170, the entry card at that level is RTX 3050.

hardware wise, MTT S80 has 55% more FLops, 100% more VRam, so it does stand a chance with poor software optimization.
(just ignore DX12 titles lol, it will be a massacre)

 

[bit_user]

hardware wise, MTT S80 has 55% more FLops, 100% more VRam, so it does stand a chance with poor software optimization.
(just ignore DX12 titles lol, it will be a massacre)

I've worked at a chip company and seen the kinds of bugs they have. They weren't GPUs, but all chips have the same sorts of bugs and if the ISA isn't something well-specified, then there's always pressure to workaround hardware bugs in software instead of doing another hugely expensive respin. Working around them can have serious performance impacts. I think that's likely what is ultimately limiting these GPUs far below their apparent potential.

Obviously, I don't know exactly what's holding them back, but I'd just caution you and @TCA_ChinChin not to get your hopes up, at least for this generation of hardware. The next generation will represent a chance both to fix their old mistakes and make some new ones!
; )

 

[TCA_ChinChin]

I've worked at a chip company and seen the kinds of bugs they have. They weren't GPUs, but all chips have the same sorts of bugs and if the ISA isn't something well-specified, then there's always pressure to workaround hardware bugs in software instead of doing another hugely expensive respin. Working around them can have serious performance impacts. I think that's likely what is ultimately limiting these GPUs far below their apparent potential.

Obviously, I don't know exactly what's holding them back, but I'd just caution you and @TCA_ChinChin not to get your hopes up, at least for this generation of hardware. The next generation will represent a chance both to fix their old mistakes and make some new ones!
; )

Of course I agree that next gen hardware is always gonna be better. I just have more faith that MT can extract more performance through software workarounds + updates similar to how AMD, Nvidia, and Intel have done.

 

[dalek1234]

Does THG get paid for posting stuff about China's tech advances, regardless if they are true, cherry-picked, or some other crypto-lies?

I need some kind of a filter add-on in my browser to hide all THG articles that contain the word 'china/chinese' in them.

Can't we just stick to news that is a bit more credible instead?

 

[bit_user]

Of course I agree that next gen hardware is always gonna be better. I just have more faith that MT can extract more performance through software workarounds + updates similar to how AMD, Nvidia, and Intel have done.

The first time building something so complex rarely goes well. For instance, look at Intel's dGPUs. Even after all the experience they had building iGPUs, they still made some mistakes that have prevented it from achieving rendering performance on par with it specifications, and two years worth of driver optimizations still haven't erased that fact. Some interviews with people involved in the project suggested that Alchemist hit some bottlenecks that they fixed in Battlemage.

In general, I think people tend to underestimate the extent to which these sorts of complex engineering project is a fundamentally iterative and evolutionary process. In the beginning, it can be hard to know where the bottlenecks will be. Once you have an actual design that you can profile and measure, there will inevitably be some surprises.

 

[bit_user]

Does THG get paid for posting stuff about China's tech advances, regardless if they are true, cherry-picked, or some other crypto-lies?

No, they write these articles because people click on them. If you don't like them, don't click on them.

Personally, I want to stay generally informed about Chinese tech. Not on a daily basis, but I would probably skim through a couple articles per week.

 

[TCA_ChinChin]

The first time building something so complex rarely goes well. For instance, look at Intel's dGPUs. Even after all the experience they had building iGPUs, they still made some mistakes that have prevented it from achieving rendering performance on par with it specifications, and two years worth of driver optimizations still haven't erased that fact. Some interviews with people involved in the project suggested that Alchemist hit some bottlenecks that they fixed in Battlemage.

That doesn't take away from how despite possible hardware defects, their dGPUs improved dramatically through driver and software updates. In fact I'd say that paints an even rosier picture for Moore's Threads.

In general, I think people tend to underestimate the extent to which these sorts of complex engineering project is a fundamentally iterative and evolutionary process. In the beginning, it can be hard to know where the bottlenecks will be. Once you have an actual design that you can profile and measure, there will inevitably be some surprises.

Yeah, it definitely is complex and an iterative engineering project. On the hardware and software side. Now that they can actually iteratively engineer their software with actual GPUs on hand, they can steadily make software improvements alongside their next generations of GPU design.