Question Zen 6 Speculation Thread

[LightningZ71]

APUs and their beefy iGPUs, as opposed to the smaller "just put a picture on the screen" iGPUs installed in the rest of today's chips, have suffered from a problem called "alternative cost".

Ever since AMD started putting usable sized iGPUs on them, they have been priced higher than better performing alternatives. There was never released to general retail a laptop that has just a Rembrandt APU with the full iGPU configuration that was a better value than any of the 1650/2050/3050 equipped laptops that were typically sold for less at retail. The only thing that AMD managed to do in the market was kill off Nvidia's MX line, which likely saved NVidia money by not having to validate, design and test such a low margin part. The same has held true through Phoenix and Hawk Point. As long as NVidia has been willing to price bottom end dGPUs low enough and provide engineering assistance to make the cost of their integration minimal to the OEMs, there's never really been a point to big iGPU chips, save for very specific markets that aren't large enough to support the development costs.

As for handhelds, it's still an immature market that likely still isn't large enough to generate a big profit, if any at all.

 

[511]

No one uses igpus to game. So it can be tiny. Tiny tiny tiny.

How can you say that I have played AAA on my Intel HD 4400 including battlefield 4 in 2014

 

[marees]

Sounds like we went from "no one uses it" to "they don't get used in mainstream parts".


There's a big difference here. Especially with @Kepler_L2 mentioning an APU capable of doing 200TOPs through its shader processors.



I guess there will be both low-end APUs (Krackan successors) and CPU-focused (Granite Ridge successors) APUs with small iGPUs like there are today. But your previous suggestion that larger iGPUs are dead was, and still is, very odd. Every SoC IHV worth mentioning out there keeps putting bigger and more capable iGPUs each generation. AMD will obviously not let themselves fall behind.

Adroc is right upto rdna 3.5

However from rdna 5 onwards the npu tensor are moved into igpu. This means the tensors & matrix cores have to be powerful enough to support 50 tops are whatever it is the Microsoft tax

If the tensor / matrix cores is minimum of 50 TOPS then the shader cores must be proportional in number. So you can work backwards on the minimum shader strength based on the minimum TOPS requirement of MS (& RDNA 5 architecture)

 

[511]

Please tell me the TOPS precision . I just hope 200 TOPS is not 200 TOPS INT 2 with Sparsity.

 

[Covfefe]

One thing I would add to the SLC vs memory bandwidth debate: SRAM scaling has slowed waaay down after N7. Which means cache gets progressively more expensive on each new node.

IEDM 2022: Did We Just Witness The Death Of SRAM?

IEDM 2022: Did we just witness the death of SRAM? While foundries continue to show strong logic transistor scaling, SRAM scaling has completely collapsed.

fuse.wikichip.org

Meanwhile, new higher bandwidth memory standards are still chugging along.

 

[511]

One thing I would add to the SLC vs memory bandwidth debate: SRAM scaling has slowed waaay down after N7. Which means cache gets progressively more expensive on each new node.

CFETs will give a big Boost to this 2X Scaling if public research is to go by

 

[blackangus]

I think it just means that AMD's next igpu will have proper tensor cores, or something close enough to it.

Yeah I read into this as the trend is that the GPU's of the future will include the functionality of the NPU today.

 

[yottabit]

I would think that if they are getting rid of the NPU, it means that AI hype is already over.

I mean, AI hype is real. It’s just that “local AI” was and is not a thing people want. The gulf in capability between what you can run locally and the SOTA models on cloud is way too large.

I remember NPUs first being used for more “legacy” AI/ML tasks like neural net for face ID etc, makes sense in that context. If MSFT wanted to implement neural net style training algorithims into Windows OS that would be cool and a potentially a valid usecase but that’s not where the hype is right now.

People want generative AI and for that there is not a substitute for putting your big boy pants on and ponying up for high end GPU time. You can only run toy models on average consumer systems.

Data privacy could be another angle Microsoft could have pushed for local AI, except their track record their is awful and they completely burnt any remaining goodwill with that “Recall” feature promising free integrated 24/7 first party spyware

 

[Covfefe]

CFETs will give a big Boost to this 2X Scaling if public research is to go by

When will they be production ready? 2030? 2035? In an best case scenario where the successor to TSMC A14 is production ready in 2030 and has CFETs, and they actually improve density by 2x, but the node costs 50% more per wafer than N2, that's still only bringing cache/dollar back to N3E levels.

CFETs may mitigate the issue of SRAM scaling, but they aren't reversing the trend.

 

[511]

When will they be production ready? 2030? 2035? In an best case scenario where the successor to TSMC A14 is production ready in 2030 and has CFETs, and they actually improve density by 2x, but the node costs 50% more per wafer than N2, that's still only bringing cache/dollar back to N3E levels.

CFETs may mitigate the issue of SRAM scaling, but they aren't reversing the trend.

The only way to reverse the trend is more players which is not possible so Either Intel/Samsung has to do something.

 

[Covfefe]

The only way to reverse the trend is more players which is not possible so Either Intel/Samsung has to do something.

I don't think its a matter of pouring more money into research. SRAM scaling is dead, like clockspeed-scaling before it. There will be some incremental increases over the years, but the huge gains we had with every node prior to 5nm aren't coming back.

 

[Win2012R2]

SRAM scaling is dead

3D stacking is the valid option, packaging for it should be expanded massively to be standard and with extra layers.

 

[adroc_thurston]

3D stacking is the valid option, packaging for it should be expanded massively to be standard and with extra layers.

$$$

 

[Win2012R2]

$$$

Yeah, apparently it's so expensive and such limited capacity that AMD makes 7600X3D (that's 6 cores low end stuff)

£280 (includes 20% VAT) - that's $313 (without tax) with current exchange rates, pre-order:

https://www.overclockers.co.uk/amd-ryzen-5-7600x3d-six-core-5.30ghz-socket-am5-processor-retail-pro-amd-04829.html

 

[511]

I don't think its a matter of pouring more money into research. SRAM scaling is dead, like clockspeed-scaling before it. There will be some incremental increases over the years, but the huge gains we had with every node prior to 5nm aren't coming back.

Wafers prices ain't if we get competition

Yeah, apparently it's so expensive and such limited capacity that AMD makes 7600X3D (that's 6 cores low end stuff)

£280 (includes 20% VAT) - that's $313 (without tax) with current exchange rates, pre-order:

https://www.overclockers.co.uk/amd-ryzen-5-7600x3d-six-core-5.30ghz-socket-am5-processor-retail-pro-amd-04829.html

Those are salvaged dies why do you think they appear late it takes time to collect Salvage

 

[Win2012R2]

Those are salvaged dies why do you think they appear late it takes time to collect Salvage

AMD sells them at a loss for $313 retail price at what is known to be pretty expensive UK e-tailer?

How would salvage even work when they surely use known good 8 core dies, can error in 3D cache packaging disable couple of cores whilst keeping cache intact?

Ok, explain away this one then: 9800X3D hottest best CPU, only $450 (removed tax):

https://www.overclockers.co.uk/amd-ryzen-7-9800x3d-eight-core-5.20ghz-socket-am5-processor-retail-pro-amd-03517.html

 

[adroc_thurston]

Yeah, apparently it's so expensive and such limited capacity that AMD makes 7600X3D (that's 6 cores low end stuff)

£280 (includes 20% VAT) - that's $313 (without tax) with current exchange rates, pre-order:

That's bananas expensive ASPs by the standards of like, mobile.

 

[511]

AMD sells them at a loss for $313 retail price at what is known to be pretty expensive UK e-tailer?

when did i say that all i am saying it takes more time time to collect salvageable dies if your yield is good

 

[Win2012R2]

when did i say that all i am saying it takes more time time to collect salvageable dies if your yield is good

Ok, explain me how can 8 core known good die become 6 cores when used with 3D cache, given very high yields of 8 cores in the first place, how can they result in 6 cores?

 

[511]

Ok, explain me how can 8 core known good die become 6 cores when used with 3D cache, given very high yields of 8 cores in the first place, how can they result in 6 cores?

Maybe core can get disabled during packing can it not ?

 

[Win2012R2]

Maybe core can get disabled during packing can it not ?

Don't know, I would have thought the biggest risks would be ruining for whole thing but on those 6 cores cache is full size, so clearly that bit worked.

But ok, I'll accept this is a salvage part, meaning low volume. Fine, 8 cores however goes for $450 and this is a known highly profitable part for AMD.

 

[Joe NYC]

Yeah, apparently it's so expensive and such limited capacity that AMD makes 7600X3D (that's 6 cores low end stuff)

£280 (includes 20% VAT) - that's $313 (without tax) with current exchange rates, pre-order:

https://www.overclockers.co.uk/amd-ryzen-5-7600x3d-six-core-5.30ghz-socket-am5-processor-retail-pro-amd-04829.html

$299 at my local MicroCenter

 

[Joe NYC]

when did i say that all i am saying it takes more time time to collect salvageable dies if your yield is good

I don't think it is about salvaging dies. 7600X does that just fine. It is more about rewarding certain retailers, giving them an exclusive product.

 

[511]

I don't think it is about salvaging dies. 7600X does that just fine. It is more about rewarding certain retailers, giving them an exclusive product.

why not? packing yield is not 100% it's high 90s so there is definitely something worth salvaging. One Man's Junk is another man's gold

 

[Joe NYC]

why not? packing yield is not 100% it's high 90s so there is definitely something worth salvaging. One Man's Junk is another man's gold

Like @Win2012R2 said, I don't think it is likely that a good core gets damaged during packaging.