Discussion AMD SoC Halo series GPU discussion

[gdansk]

I don't think Flex ATX PSU are known for being quieter. They have tiny, higher RPM fans.

Read again, a quieter Flex ATX PSU. Not quieter than non-flex.
For small systems that is the advantage of a lower power design.

 

[misuspita]

Depending on power needs you can get a HDPlex 250W completely passive or 500W somewhat passive PSU. I wonder why would Framework recomands a 350W+ PSU

 

[misuspita]

Sam

Same 5700G. I have been eyeing a HX 370 mini pc for a long time, but they all launched with 7500MHZ memory, and I didn't want to compromise the bandwidth. Now Halo spit in their soup. Either the Framework or that HP thing is very tempting, especially if the latter has a more affordable 12 core 64GB option, which seems to be missing from Framework. I'm not even a gamer, just interested in all that bandwidth and having 12 real cores with lots of L3 over what is in Strix Point. The 16 core Framework would be a massive upgrade for you. Medusa of course looks very enticing, but I cannot suffer this basic bitch motherboard for two more years, even if Cezanne is decent. So I must make a difficult decision within 6 months.

7500 is very good. Any non soldered RAM is much under that speed, and socketed ram is at 5600 speeds

 

[igor_kavinski]

I wonder why would Framework recomands a 350W+ PSU

Transient spikes. USB4 power delivery. Maybe special overclocking features unlocked only on the desktop mobos.

 

[gorobei]

Depending on power needs you can get a HDPlex 250W completely passive or 500W somewhat passive PSU. I wonder why would Framework recomands a 350W+ PSU

i think it is in the LTT video, the psu is actually a higher rated 500 or 600W from fsp that they then listed as 350 so that the fan could be absolutely quiet. so technically able to go passive but if you actually load it up it will still be able to cool at a noise level that the mac people cant complain about.

 

[scineram]

7500 is very good. Any non soldered RAM is much under that speed, and socketed ram is at 5600 speeds

I made peace with solderer memory precisely for higher speeds. But if Strix Point (and Halo) supports 8000, I want no compromises on that front at least.

 

[Glo.]

I wonder if we'll see an SOC with a 384bit or 512bit wide memory bus in 2-4 years.

A 96CU RDNA5 iGPU with a 512bit memory bus could be the single fastest GPU available in any laptop. AMD is finally catching up to Nvidia in perf/watt with RDNA4, Strix Halo has shown that big iGPUs have a major energy efficiency advantage over dGPUs, and RDNA4 has way more gaming performance at the same exact memory bandwidth as RDNA3. Of course, it would depend on how good the RTX 6000 series is, but it would at least beat an RTX 5090 mobile.

IF, I reapeat, IF, there will be high demand for Strix Halo and Medusa Halo from consumers, and AMD will see that there is demand for higher tier SKU, with a larger premium - yes, there will be larger design.

Don't expect it anytime soon, tho, and don't expect that it will be "mainstream" platform. Such a product could even be 1200-2000$ for SOC, with MoP, depending on performance targets, and product segment.

I do have a name for it, tho.

AMD EPYC AI.

 

[GTracing]

IF, I reapeat, IF, there will be high demand for Strix Halo and Medusa Halo from consumers, and AMD will see that there is demand for higher tier SKU, with a larger premium - yes, there will be larger design.

Don't expect it anytime soon, tho, and don't expect that it will be "mainstream" platform. Such a product could even be 1200-2000$ for SOC, with MoP, depending on performance targets, and product segment.

I do have a name for it, tho.

AMD EPYC AI.

I don't think it would have to cost that much. Since they are already doing one halo part, a bigger halo part is only one more die to design. If AMD is able to take a quarter of the dGPU laptop market share, they will have enough volume to justify two big iGPU designs at reasonable price points. As a comparison, it would be more volume than they moved with the whole RDNA3 lineup, and that had 3 unique compute dies. I agree it would be 4+ years out though.

 

[Glo.]

I don't think it would have to cost that much. Since they are already doing one halo part, a bigger halo part is only one more die to design. If AMD is able to take a quarter of the dGPU laptop market share, they will have enough volume to justify two big iGPU designs at reasonable price points. As a comparison, it would be more volume than they moved with the whole RDNA3 lineup, and that had 3 unique compute dies. I agree it would be 4+ years out though.

We are talking:
A) most likely 2 nm product.
B) a chip that has Memory on Package, which will increase BOM costs, and hence the end price with high margin/low volume product.

 

[GTracing]

We are talking:
A) most likely 2 nm product.
B) a chip that has Memory on Package, which will increase BOM costs, and hence the end price with high margin/low volume product.

My point is that if they move enough of them, it doesn't have to be high margin/low volume.

 

[Glo.]

My point is that if they move enough of them, it doesn't have to be high margin/low volume.

Of course, but it will still be priced like it is high margin/low volume one.

 

[Glo.]

AMD in APUs has extremely rare opportunity to reset their brand appeal, and make it possible to charge for their products extremely high prices, comparatively to what they have to charge now.

AMD will take this opportunity IF it happens. A lot has to fall in place(software features parity with competition, good marketing, good branding). Currently - they are VERY FAR from that place, from business perspective.

Strix Halo is the type of product that speaks for itself. It markets itself. It even has correct, very good pricing. If only it would have better branding...

 

[GTracing]

Of course, but it will still be priced like it is high margin/low volume one.

Why? If AMD can increase market share and overall profit by taking a lower margin on halo SKUs, that seems like a win. 40% margin on 5 million units sold is better than 60% margin on 2 million units sold.

 

[Glo.]

Why? If AMD can increase market share and overall profit by taking a lower margin on halo SKUs, that seems like a win. 40% margin on 5 million units sold is better than 60% margin on 2 million units sold.

That product WIL NOT BE to increase market share. It will be designed to increase profits.

 

[GTracing]

That product WIL NOT BE to increase market share. It will be designed to increase profits.

Increasing market share will increase overall profit.

You keep saying that AMD has high margins on CPUs so they aren't going to lower them, but how did they get to the point of high margins in the first place? AMD sold early Zen CPUs with a comparitively small margin.

If they sell big iGPUs at a reasonable margin they have a real chance to take some of the dGPU laptop market. That's something that they've been trying and failing to do for years. They don't even need to sell them at a loss, just at a lower profit margin.

This wouldn't affect the profit on mainstream laptop CPUs either. They can sell mainstream chips with a 50% margin, 256bit halo with a 45% margin, and 512bit halo with a 40% margin.

 

[Glo.]

Increasing market share will increase overall profit.

You keep saying that AMD has high margins on CPUs so they aren't going to lower them, but how did they get to the point of high margins in the first place? AMD sold early Zen CPUs with a comparitively small margin.

If they sell big iGPUs at a reasonable margin they have a real chance to take some of the dGPU laptop market. That's something that they've been trying and failing to do for years. They don't even need to sell them at a loss, just at a lower profit margin.

This wouldn't affect the profit on mainstream laptop CPUs either. They can sell mainstream chips with a 50% margin, 256bit halo with a 45% margin, and 512bit halo with a 40% margin.

Can you get M4 Max and M2 Max in sub 2000$ desktop PC and Sub 2500$ laptop?

No. Expect that will be exactly the same for hypothetical large APU from AMD.

Who is there to compete with AMD in the front of large APUs?

Apart from Apple, and maybe at some point - Qcomm and Nvidia, nobody.

Since they are ARM based products, who is there to compete with AMD on x86 front?

Nobody.

 

[adroc_thurston]

Who is there to compete with AMD in the front of large APUs?

NVL-AX.
If it survives, that is.

 

[GTracing]

Can you get M4 Max and M2 Max in sub 2000$ desktop PC and Sub 2500$ laptop?

No. Expect that will be exactly the same for hypothetical large APU from AMD.

Who is there to compete with AMD in the front of large APUs?

Apart from Apple, and maybe at some point - Qcomm and Nvidia, nobody.

Since they are ARM based products, who is there to compete with AMD on x86 front?

Nobody.

The M4 Macbook starts at $1100. It has a ~170mm^2 SOC (from what I could find, could be a little off). I don't think what apple charges is particularly relevant.

And my whole argument is that big iGPUs can compete with dGPUs if AMD chooses to do so.

 

[Glo.]

The M4 Macbook starts at $1100. It has a ~170mm^2 SOC (from what I could find, could be a little off). I don't think what apple charges is particularly relevant.

And my whole argument is that big iGPUs can compete with dGPUs if AMD chooses to do so.

Why do you bring M4 to this discussion?

Comparison:
Apple M4 - Strix Point.
Apple M4 Pro - Strix Halo.
Apple M4 Max - XXXX Max.

Thats how those products would stack up, and compete with each other.

AMD will NOT make large iGPU to go into products that are supposed to be mainstream.

And no, even on 2 nm process it is not possible to cram 96 CUs into even 200 mm2 die size, if thats what you wanted.

 

[adroc_thurston]

AMD will NOT make large iGPU to go into products that are supposed to be mainstream.

An 8+32 or 8+36 Halo config could be plenty mainstream.

 

[Glo.]

An 8+32 or 8+36 Halo config could be plenty mainstream.

Mainstream in this context means something different .

Yes, 8+32 especially on 2 nm will be hella mainstream. We are talking however about iGPU that is 80-96 CUs thats compared to current gen.

And with current gen 80 CU APU would be very much NOT a mainstream APU.

 

[adroc_thurston]

We are talking however about iGPU that is 80-96 CUs thats compared to current gen

80 you're never getting, since gfx12/13 focus is per-CU perf bumps.

 

[Glo.]

80 you're never getting, since gfx12/13 focus is per-CU perf bumps.

Yeah, just like we would never get Strix Halo soldered into desktop motherboards.

 

[adroc_thurston]

Yeah, just like we would never get Strix Halo soldered into desktop motherboards.

I mean it.

 

[GTracing]

Why do you bring M4 to this discussion?

Comparison:
Apple M4 - Strix Point.
Apple M4 Pro - Strix Halo.
Apple M4 Max - XXXX Max.

Thats how those products would stack up, and compete with each other.

AMD will NOT make large iGPU to go into products that are supposed to be mainstream.

And no, even on 2 nm process it is not possible to cram 96 CUs into even 200 mm2 die size, if thats what you wanted.

My point is Apple charges a lot relative to the amount of silicon they put in their laptops. Anyway, this is getting off topic, so this is probably my last comment. Cheers!