Discussion AMD SoC Halo series GPU discussion

[adroc_thurston]

RDNA4 would've been awesome with its major bandwidth efficiency gains.

and native FSR4 support, just the right fit for 1600/1800p panels.

Well, Medusa Halo with RDNA5 and hopefully an even wider LPDDR6 bus, yeah.

It's LPDDR5X and methinks they're gonna WoW plop a real chonga MALL slab to cope with the ramped b/w demand.

 

[branch_suggestion]

and native FSR4 support, just the right fit for 1600/1800p panels.

Yeah, it is a real mean uArch.

It's LPDDR5X and methinks they're gonna WoW plop a real chonga MALL slab to cope with the ramped b/w demand.

That works too, that is a pretty damn big WoW stack if true.

 

[adroc_thurston]

That works too, that is a pretty damn big WoW stack if true.

Just do a Venice and move stuff that's cores (i.e. gfx in this case) on top and everything else on the bottom.

 

[poke01]

Does Intel even have a comp part for this?

AMD will take the gaming side and Apple the creative side. why isn't Intel focusing on large APUs? They are just giving away the market.

 

[adroc_thurston]

Does Intel even have a comp part for this?

They always do and it always gets killed.

AMD will take the gaming side and Apple the creative side

No, Nvidia will take the creative side.

why isn't Intel focusing on large APUs? They are just giving away the market.

THEY WERE.
ARL-H 683 was their M3 Pro intercept part lmao.
Like they had a bunch of ADM-enabled configs for MTL and ARL and they all went poof.

 

[poke01]

No, Nvidia will take the creative side.

oh yea, is it even coming out this year?

ARL-H 683 was their M3 Pro intercept part lmao.
Like they had a bunch of ADM-enabled configs for MTL and ARL and they all went poof.

yikes

 

[adroc_thurston]

oh yea, is it even coming out this year?

Next.

yikes

Happens. Not like their GFX IP is any good anyhow.

 

[branch_suggestion]

Just do a Venice and move stuff that's cores (i.e. gfx in this case) on top and everything else on the bottom.

You can add caps to the cache layer too, also saves on fabric complexity thanks to glorious direct TSVs.
Basically everything can be in the bottom layer except the GPU, saves a lot of pricey wafer cost on the IP that matters.
Now please AMD surprise us with WoW RDNA5 flagship for the love of all that is cool.

 

[adroc_thurston]

Basically everything can be in the bottom layer except the GPU, saves a lot of pricey wafer cost on the IP that matters.

Yeah like N3p on N4c or N2 on N4c and awooga.

Now please AMD surprise us with WoW RDNA5 flagship for the love of all that is cool.

what are the odds of the next NV SM not being another practical shotgun-to-the-mouth application

 

[branch_suggestion]

Yeah like N3p on N4c or N2 on N4c and awooga.

So N2 and N4 will be SoIC compatible?

what are the odds of the next NV SM not being another practical shotgun-to-the-mouth application

Depends on how much they value CUDA/binary backcompat.
But you cannot assume NV will flop twice in a row, especially with a node bump. Gotta flex those 3D muscles before they have a real answer.

 

[adroc_thurston]

So N2 and N4 will be SoIC compatible?

Bretty sure that's what Apple gonna ship so yea.

Depends on how much they value CUDA/binary backcompat.

PTX is a thing so whatever.

But you cannot assume NV will flop twice in a row, especially with a node bump. Gotta flex those 3D muscles before they have a real answer.

SoIC stuff is just Apple and AMD for the foreseeable future.

 

[techjunkie123]

View attachment 115660
Okay bros I need your best cope as to why it actually sucks™.

No 4x MFG on?? Remeber 5070 laptop gonna have like 4X™ the performance . Matching desktop 4080.

 

[misuspita]

It kinda sucks on counter stryke, so it's a flop.



/s

 

[Abwx]

A few numbers on that matter :

AMD Claims Ryzen AI Max+ 395 "Strix Halo" APU With Radeon 8060S Up To 68% Faster Than RTX 4070 In 1080p Gaming Benchmarks

AMD has shared more gaming benchmarks of its Ryzen AI Max+ 395 "Strix Halo" APU and its integrated Radeon 8060S GPU.

wccftech.com

 

[MS_AT]

I find it curious that the only device targeted at "mainstream" audiences is a tablet and they focus their marketing efforts on that [at least I fail to see the appeal of a gaming tablet that is ill suited to gaming without tons of peripherals attached. I mean I don't see an advantage vs 14 inch laptops than can sport dedicated GPU and being of comparable weight].

I mean a proper 14 inch gaming laptop would be better device to put Halo into, as it wouldn't charge price premium on top of using the Halo for being a tablet, but might be I misunderstand the market.

On the other hand they are polling github if people want to see ROCM support on Halo chips... [I mean shouldn't it be their priority so they could address the part of the market that Apple captures with their miniPCs]

 

[misuspita]

This is a cpu that needs a miniPC type of device, to have better cooling and be allowed to use more power

 

[adroc_thurston]

This is a cpu that needs a miniPC type of device, to have better cooling and be allowed to use more power

It's literally designed for laptops.
It doesn't need more cooling or more power.

 

[jpiniero]

I find it curious that the only device targeted at "mainstream" audiences is a tablet and they focus their marketing efforts on that [at least I fail to see the appeal of a gaming tablet that is ill suited to gaming without tons of peripherals attached. I mean I don't see an advantage vs 14 inch laptops than can sport dedicated GPU and being of comparable weight].

I think there might be a bit of throwing crap on the wall and see what sticks. Even AI Workstations (the main use of Strix Halo) it's not entirely clear how much demand there is for that.

 

[Mopetar]

There goes any hope of it being less expensive than an equivalent CPU plus dGPU.

As others have pointed out running it on a traditional socket would just bottleneck it so badly as to not give better price/performance anyway so the cost really doesn't matter.

AMD would need to sell the board this would be attached to and may as well solder the RAM on as well at that point. Congratulations, AMD, you've invented your own Mac without having your own OS.

Maybe some PC manufacture might decide to release something like that, but I wouldn't expect it to be a major release. It might make a nice Steambox or living room PC that replaces a console, but I don't see them investing too much into something like that. It doesn't have the volume potential to reduce the cost of making it and the higher price it will need to have to turn a profit mean that it will have low volume.

 

[Heartbreaker]

As others have pointed out running it on a traditional socket would just bottleneck it so badly as to not give better price/performance anyway so the cost really doesn't matter.

AMD would need to sell the board this would be attached to and may as well solder the RAM on as well at that point. Congratulations, AMD, you've invented your own Mac without having your own OS.

Maybe some PC manufacture might decide to release something like that, but I wouldn't expect it to be a major release. It might make a nice Steambox or living room PC that replaces a console, but I don't see them investing too much into something like that. It doesn't have the volume potential to reduce the cost of making it and the higher price it will need to have to turn a profit mean that it will have low volume.

I'm not suggesting they use a regular AM5 socket either. Just that using Threadripper MBs ruins any chance of beating conventional CPU plus dGPU pricing.

They are better off soldering it in. Yes it is much like Mac HW at that point, but some claim they want that.

Many people overestimate the benefit of the Big iGPU x86 SoC, and have been clamoring for something like this for MANY years, assuming it will be less expensive than comparable CPU and dGPU, and be a great budget gaming platform.

I've never been that optimistic. Because you have one locked in, VERY expensive chip that will likely end up in niche volumes. This is NOT a recipe for budget pricing or success.

The main thing AMD seemed to lean on in their presentation was a big memory pool for AI. If it wasn't for all the AI hype I wonder if they even would have built it.

 

[adroc_thurston]

The main thing AMD seemed to lean on in their presentation was a big memory pool for AI. If it wasn't for all the AI hype I wonder if they even would have built it

This was roadmapped for years, predates the ML bubble by a lot.

 

[jpiniero]

This was roadmapped for years, predates the ML bubble by a lot.

I think it probally would have been cancelled due to lack of OEM interest without it though.

 

[adroc_thurston]

I think it probally would have been cancelled due to lack of OEM interest without it though.

Nope, the OEM interest in Mx Pro compete is a very persistent thing.

 

[MS_AT]

Nope, the OEM interest in Mx Pro compete is a very persistent thing.

So will we see more products than 1 tablet, 1 laptop and 2 miniPCs announced so far?

 

[techjunkie123]

So will we see more products than 1 tablet, 1 laptop and 2 miniPCs announced so far?

I hope so. They need like an Asus Proart P14 type laptop. Like a professional G14.