News Chinese GPU Dev Starts Global Sales of $245 RTX 3060 Ti Rival

[Admin]

Moore Threads MTT S80 graphics cards based on the Chunxiao architecture is now available.

Chinese GPU Dev Starts Global Sales of $245 RTX 3060 Ti Rival : Read more

 

[bit_user]

Interestingly, the MTT S80 is the world's first client graphics card with a PCIe 5.0 x16 interface

Weird. And completely unnecessary. For 3060 Ti-level performance, PCIe 3.0 x16 should be perfectly fine.

The only reason I can see for doing it is maybe as a test vehicle for debugging their PCIe controller that they plan to reuse in bigger or completely different products.

Or, if we're cynical, they did it as a way to have some selling-point where they can claim leadership.

 

[bit_user]

Moore Threads designs GPUs for many applications (except HPC) and seems intent on selling its products outside of China.

If they try to establish a real market presence in western countries, that means building a distribution channel and those distributors will be vulnerable to patent infringement claims by existing GPU players (Nvidia, AMD, Intel, Imagination, and ARM). That seems likely to limit their penetration of existing markets, where such IP claims get play.

 

[John Hogan]

Bring it on. As with the car industry and the coming rush of solid Chinese EVs, this sounds like good news. 3060ti performance should cost about $250 retail with a useful margin. I've owned nothing but Nvidia cards since my voodoo 2. But I won't miss that company if it disappears in the next few years. History is littered with companies that got to the top, took their customers for granted and then got eaten by people trying harder. Strange to say in 2022, but I'd rather give my automotive dollars to BYD and if this GPU mob is serious I'll give them a go too.

 

[artk2219]

As intel is currently finding out, and as AMD and Nvidia learned to varying degrees long ago, it comes down to the drivers. It can be the most powerful GPU ever created, and it wont matter one bit if it cant actually be used by any software because the drivers for it suck.

 

[daworstplaya]

As intel is currently finding out, and as AMD and Nvidia learned to varying degrees long ago, it comes down to the drivers. It can be the most powerful GPU ever created, and it wont matter one bit if it cant actually be used by any software because the drivers for it suck.

^This! That's basically wasted silicon without good drivers. Wonder how much IP theft happened to create a GPU with those specs. Stream processors = AMD ?

 

[Chung Leong]

Protectionism works, even when the barriers involved are erected by your bone-headed adversary.

 

[bit_user]

As intel is currently finding out, and as AMD and Nvidia learned to varying degrees long ago, it comes down to the drivers. It can be the most powerful GPU ever created, and it wont matter one bit if it cant actually be used by any software because the drivers for it suck.

Yes, but I think Intel's problems are also deeper than mere drivers.

IMO, Intel tried to do too much, in their first gen. They introduced special units for ray tracing and AI inferencing, then had to implement & optimize driver & runtime support for these + XeSS (their DLSS-equivalent). What they missed is that if the core rasterization performance and drivers for it aren't there, nobody cares much about the other stuff. Especially not ray tracing. You could make the case that XeSS can help them squeeze more performance out of their existing rasterization hardware, so I guess I'm really arguing they shouldn't have burned hardware and software resources trying to do ray tracing, also.

 

[bit_user]

I've owned nothing but Nvidia cards since my voodoo 2. But I won't miss that company if it disappears in the next few years. History is littered with companies that got to the top, took their customers for granted and then got eaten by people trying harder.

I don't see Nvidia as not trying. They tend to exploit their market leadership with very high margins, though. Also, I don't see the same creativity as AMD has shown, with their features like Infinity Cache. That complacency could be more of a recent thing, because just 2 generations earlier they introduced tensor cores and hardware ray tracing.

As a company, Nvidia is focusing very hard on datacenter AI, with products like their DPUs, and edge AI. For them, gaming and visualization is still the main money-maker, but no longer central to their mission.

Strange to say in 2022, but I'd rather give my automotive dollars to BYD and if this GPU mob is serious I'll give them a go too.

To the extent you're interested in cost-competitiveness, you'd do well to look into how heavily those products are subsidized. And consider that, if those brands ever achieve market dominance, they might cease to be so inexpensive.

 

[setx]

Weird. And completely unnecessary. For 3060 Ti-level performance, PCIe 3.0 x16 should be perfectly fine.

For useless gaming it's indeed useless. But for compute I'd absolutely love faster interface.
The cards are obviously more compute-oriented given lackluster list of supported games.

 

[blacknemesist]

These GPUs will help keep the mid-range in check, Intel could have done that if they weren't so greedy with the 1st gen card that had a lot of issues.
Maybe in the future we get more GPUs like this to keep NVidia and AMD from going overboard with pricing and rise the demand, PC are probably at the lowest point ever because of how every component is extremely expensive or extremely poor.

 

[thestryker]

Yes, but I think Intel's problems are also deeper than mere drivers.

IMO, Intel tried to do too much, in their first gen. They introduced special units for ray tracing and AI inferencing, then had to implement & optimize driver & runtime support for these + XeSS (their DLSS-equivalent). What they missed is that if the core rasterization performance and drivers for it aren't there, nobody cares much about the other stuff. Especially not ray tracing. You could make the case that XeSS can help them squeeze more performance out of their existing rasterization hardware, so I guess I'm really arguing they shouldn't have burned hardware and software resources trying to do ray tracing, also.

In one of the videos they put out Raja admitted that the team's lack of experience on dGPUs led them to make some design decisions which hurt the performance. It sounded like hardware design is holding back the performance for desktop which might explain the wide variations on performance (which exists in dx12/vulcan as well) beyond just drivers.

I believe this was the video I'm thinking of:

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

 

[8086]

Its the worlds first GPU that watches you play..... instead of you watching your screen.

 

[bit_user]

for compute I'd absolutely love faster interface.
The cards are obviously more compute-oriented given lackluster list of supported games.

PCIe 5.0 x16 is nominally 64 GB/s per direction. That's close to the limit of what DDR5 can support. So, what benefit is it you see from having such a fast interface in a card presumably being used in a desktop PC?

 

[bit_user]

Intel could have done that if they weren't so greedy with the 1st gen card that had a lot of issues.

Greedy? They're getting less performance per mm^2 than AMD or Nvidia. I'm not sure they can charge much less for their GPUs.

 

[d0x360]

Greedy? They're getting less performance per mm^2 than AMD or Nvidia. I'm not sure they can charge much less for their GPUs.

On a new GPU architecture and not new as in we added tensor cores to a standard but more of it design.

The development of the interconnect fast enough to allow for mcm rendering without constant stutter is impressive and only going to keep getting better.

This time next year they will unveil the GPU that beats nVidia in price, raster, ray tracing and driver performance.

As far as this Chinese designed gpu... Not a chance I'd put homegrown Chinese hardware in my system. Sorry but I just can't bring myself to trust a country run by a psycho that's commiting genocide as I type this... They can't be trusted (the CCP not the people) but until they get overthrown and the Taiwanese take their government back but on the mainland... Nope.

Plz don't say but e everything is made in China. There is a massive difference between assembled and designed then made by a company that has to do whatever the CCP says and has no oversight.

Same goes for that new VR HMD. What happens if Xi decides one day he wants to use it to try and 0day as many PC's connected to the same network? Byte dance isn't in charge of anything.

 

[InvalidError]

Weird. And completely unnecessary. For 3060 Ti-level performance, PCIe 3.0 x16 should be perfectly fine.

The 4GB RX5500 which gains up to 70% more performance on 4.0x8 vs 3.0x8 says hi. 3.0x16 is borderline, especially at the very lower-end.

PCIe 5.0 x16 is nominally 64 GB/s per direction. That's close to the limit of what DDR5 can support. So, what benefit is it you see from having such a fast interface in a card presumably being used in a desktop PC?

128bits x 6400MT/s = 128GB/s, ~100GB/s usable after overheads, dead time, refresh, stalls, etc.

Chances are that the GPU is also chatting with the CPU at a few GB/s too. Having some spare bus bandwidth means minimal bus contention there.

 

[bjnmail]

AMD has yet to disclose the host interface of its latest RDNA 3 GPUs in general and Radeon RX 7900 XT and Radeon 7900 XTX in particular.

Is the author living under a rock or something? In the announcement presentation the RDNA3 GPUs were listed as having a PCIe gen4 interface.

 

[The Historical Fidelity]

Bring it on. As with the car industry and the coming rush of solid Chinese EVs, this sounds like good news. 3060ti performance should cost about $250 retail with a useful margin. I've owned nothing but Nvidia cards since my voodoo 2. But I won't miss that company if it disappears in the next few years. History is littered with companies that got to the top, took their customers for granted and then got eaten by people trying harder. Strange to say in 2022, but I'd rather give my automotive dollars to BYD and if this GPU mob is serious I'll give them a go too.

I’d stay away from BYD products, their quality control is poor and they have been accused multiple times of using slave labor in the manufacturing of their products.

 

[bit_user]

The development of the interconnect fast enough to allow for mcm rendering without constant stutter is impressive and only going to keep getting better.

Who? Intel? Do we have evidence of that?

This time next year they will unveil the GPU that beats nVidia in price, raster, ray tracing and driver performance.

I'll believe it when I see it (and welcome it, too). Do you have a better source for this than the Internet Rumor Mill?

 

[bit_user]

The 4GB RX5500 which gains up to 70% more performance on 4.0x8 vs 3.0x8 says hi. 3.0x16 is borderline, especially at the very lower-end.

This sort of discrepancy was also observed with the RX 6500XT, when comparing its x4 interface at 3.0 vs. 4.0 speeds. The key thing to highlight is that it's an outlier result.

What's going on with that, is the GPU is having to page in textures from system RAM, due to insufficient graphics memory. This is beyond the sweet spot for most games, so what most users would do is dial back the detail level and effects to a point where that's not happening. As for applying it to the card in the article, that's purported to have 16 GB, making it a non-issue.

For reference, the average performance loss from running a RTX 4090 @ 1080p at PCIe 3.0 is only 3%:

https://www.techpowerup.com/review/nvidia-geforce-rtx-4090-pci-express-scaling/28.html​

If you dig into the per-game benchmarks, there are sure to be some outliers, but that should be enough to tell us that PCIe 3.0 x16 shouldn't hold back a card with 3060 Ti-level performance. Even in those outlier cases, stepping up to 4.0 should be plenty.

128bits x 6400MT/s = 128GB/s, ~100GB/s usable after overheads, dead time, refresh, stalls, etc.

You're computing bidirectional bandwidth? That doesn't make as much sense to highlight, because the data flow isn't symmetrical. Most of the data is going from CPU and system RAM to graphics memory.

As for real-world bandwidth, if we extrapolate from this small set of results from 3DMark's PCIe Bandwidth test, it seems we should anticipate up to 53 GB/s on PCIe 5.0 x16.

https://www.reddit.com/r/overclocki..._pci_express_feature_test_share_some_results/​

​

 

[InvalidError]

This sort of discrepancy was also observed with the RX 6500XT, when comparing its x4 interface at 3.0 vs. 4.0 speeds. The key thing to highlight is that it's an outlier result.

It isn't an outlier when nearly all low-VRAM GPUs ever made benefit disproportionately from more PCIe bandwidth. The RX5500 and RX6500 are simply the most spectacular examples thanks to AMD going at least one step too far in its PCIe-cutting for a large chunk of its potential market. I omitted the RX6500 because it is mobile garbage shoehorned into a desktop thing and should just pretend it never existed now that the RX6600 almost down to $200 retail.

 

[bit_user]

It isn't an outlier when nearly all low-VRAM GPUs ever made benefit disproportionately from more PCIe bandwidth.

The test showing that delta is an outlier, in the sense that it's a corner case that pushes the graphics card outside of its sweet spot. Even with PCIe 4.0, you typically wouldn't want to use those settings.

We can imagine all sorts of tests that stress these products, but the goal should be to inform readers adequately to make the best decision for their needs. Over-emphasizing a test result from a scenario they're unlikely to encounter therefore doesn't serve their interests. Yes, it should be reported, but not blown out of proportion, as though it characterizes all games & settings.

The RX5500 and RX6500 are simply the most spectacular examples thanks to AMD going at least one step too far in its PCIe-cutting for a large chunk of its potential market.

My take on it is different. I think what it shows is that the amount of graphics memory is too small, if this kind of performance discrepancy happens in games & settings that are otherwise playable on the GPU. In general, you don't want too many texture fetches from system RAM, even at PCIe 4.0 speeds, since it can cause stuttering.

I understand they were trying to negotiate issues with the crypto market, so I give them a pass for selling 4 GB cards, in this instance.

I omitted the RX6500 because it is mobile garbage shoehorned into a desktop thing and should just pretend it never existed now that the RX6600 almost down to $200 retail.

It made sense, for the market conditions at the time. If it launched in today's market, I would judge it much more harshly.

That said, I'd like to see a passively-cooled RX 6400, and I think the RX 6500XT might still be interesting at < a $100 price point. These are enough for business users, who still need a dGPU for ThreadRipper workstations and any non-APU AM4 machines still being built.

Aside from the lack of AV1 decode, these products are quite competitive against Intel's A380. In terms of price/perf, they should continue to be a more compelling value, given the Intel GPU's larger die and memory size. All are TSMC N6, making die-size comparisons very relevant.

 

[InvalidError]

The test showing that delta is an outlier, in the sense that it's a corner case that pushes the graphics card outside of its sweet spot. Even with PCIe 4.0, you typically wouldn't want to use those settings.

While enthusiasts may balk at playing games at anything less than 60fps, people on a budget or who cannot be bothered to spend $300+ on a GPU may be more than willing to make that sacrifice if it gets them meaningfully better visual quality while still maintaining generally playable frame rates. On my GTX1050, I take whatever details I can get as long as I hit at least 40fps most of the time. The only reason that GPU is still remotely usable today is because it has 3.0x16 to partially cover for its 2GB of VRAM.

 

[setx]

PCIe 5.0 x16 is nominally 64 GB/s per direction. That's close to the limit of what DDR5 can support. So, what benefit is it you see from having such a fast interface in a card presumably being used in a desktop PC?

One of the biggest issues with GPGPU is it's often not worth it to move some computation to discreet GPU even if that part of algorithm is well suited for GPU parallelism because data transfer time would take more than speedup of computation.

So, faster interfaces would allow more GPU-acceleration in wide variety of programs.