Discussion AMD SoC Halo series GPU discussion

[Joe NYC]

I would expect a MiniPC from Beelink or Minisforum for just under 2k in 6 months. I'd wish 1500 but its a pipedream.

Speaking of MinisForum, I came across a number of ads on YouTube for their latest StrixPoint MiniPC, so maybe they have some success with them. I think they will be racing to release Strix Halo MiniPC ASAP

 

[fastandfurious6]

Halo cost

9950x =~ $550
4070 = ~$600

around $1000 minimum based on market values

 

[MS_AT]

Maybe wait 6 months and pay same price for 128GB one?

I still haven't decided if I would really make use of 128GB in the expected lifetime of the machine.

I think you will be able to get one for much less than $2500 from company like MinisForum

Yes, but I prefer to have more dependable support if I am paying that amount of money. What I have read about Minisforum does not sound promising.

But Zen 5 is something of a dud in client workloads

Which workloads do you have in mind?

 

[gdansk]

Which workloads do you have in mind?

A bloated, cross-platform IDE.
And web browsers

 

[igor_kavinski]

I still haven't decided if I would really make use of 128GB in the expected lifetime of the machine.

You can, if even better coding LLMs become available and you become dependent on them for quickly pointing out any mistakes/oversights.

 

[MS_AT]

You can, if even better coding LLMs become available and you become dependent on them for quickly pointing out any mistakes/oversights.

But the bandwidth does not scale with the amount of memory and smaller models do just fine as google search replacement I want to have something that will be able to do at least 6 tk/s. With 256GB/s BW, that is aprox 40GB model.

A bloated, cross-platform IDE.
And web browsers

Granite Ridge is doing quite fine with them, do you mean that 600MHz clk deficit is making it much worse?

 

[gdansk]

Granite Ridge is doing quite fine with them, do you mean that 600MHz clk deficit is making it much worse?

No, they're cross platform. One has to look to the real competition where it's often 20%-30% slower.

 

[Glo.]

Maybe wait 6 months and pay same price for 128GB one?

I hope that 32 GB/1 TB SKU of HP Z2 Mini G1a with Ryzen 385(8C/32CU) will be around 1500$.

 

[MS_AT]

No, they're cross platform. One has to look to the real competition where it's often 20%-30% slower.

Apple Silicon is not always an option

 

[gdansk]

Apple Silicon is not always an option

Following the chain I was talking for myself. In that context it is always an option.

But Zen 5 is something of a dud in client workloads so I can't justify 2500-3000$ on any Zen 5 laptop.

 

[LightningZ71]

For at least two complete APU generations, AMD has had the ability to easily create an APU that was on par with 50 series NVIDIA GPUs of the prior generation. From Rembrandt, they had RDNA2 with the ability to use 16MB of infinity cache internal to the processor, requiring absolutely nothing extra from the OEM, save for a little extra power and decent processor cooling. It would have been near 2050 4MB performance. Yes, it would have required about 5-10% more die area than the processor they went with, but it would have commanded a comparable price premium. Phoenix and Hawk could have done the same and been quite close to 4MB 3050 performance.

Strix Point, instead of using a 4+8 config with dual CCXs could have used a regular 8 P cores config and used the saved space for a 16MB MALL cache and achieved near 4050 performance while not sacrificing a whole lot of MT performance and been a more flexible and compelling product overall.

Using OEMs as the excuse is a crutch. Yes, back in the Vega days, it would have required a more involved product, but since RDNA2, no. AMD is reaping the costs of not having their own fab and the volume limits that result.

 

[jpiniero]

Using OEMs as the excuse is a crutch.

Kinda requires two to tango.

 

[branch_suggestion]

Samsung extends LPDDR5 to 12.7 GT/s: Next-gen devices enjoy a nice speed boost

Samsung's LPDDR5-Ultra-Pro further clings into LPDDR6 territory

www.tomshardware.com

Lets see if this actually ships, regardless though Medusa Halo should have noticeably faster 5X than now.

 

[jpiniero]

On Asus' site, you can preorder the Flow. It's $2099 for the 390, $2299 for the 395 and $2799 for the 395 with 128 GB.

Guess if you are gonna buy it for the AI Hype, may as well go all in.

 

[Joe NYC]

Samsung extends LPDDR5 to 12.7 GT/s: Next-gen devices enjoy a nice speed boost

Samsung's LPDDR5-Ultra-Pro further clings into LPDDR6 territory

www.tomshardware.com

Lets see if this actually ships, regardless though Medusa Halo should have noticeably faster 5X than now.

Assuming you can route traces to 8 discrete chips at that speed. AMD already reduced the speed from 8533 to 8000 for Strix Halo.

It seems more and more clear that AMD needs to move to Memory On Package, and than, new SKUs can be released with faster chips inside the package, while being replaceable with previous SKUs with slower DRAM.

It would be kind of funny if at the same time Intel were to retreat from MoP, (after Intel explaining over and over it is not ideal technology) and AMD introduces Medusa Halo with MoP and kills it...

 

[jpiniero]

OEMs seemingly hate MoP. That's why Lunar Lake was a flop.

 

[branch_suggestion]

On Asus' site, you can preorder the Flow. It's $2099 for the 390, $2299 for the 395 and $2799 for the 395 with 128 GB.

Guess if you are gonna buy it for the AI Hype, may as well go all in.

$500 for the full memory config is less than I thought.

Assuming you can route traces to 8 discrete chips at that speed. AMD already reduced the speed from 8533 to 8000 for Strix Halo.

It seems more and more clear that AMD needs to move to Memory On Package, and than, new SKUs can be released with faster chips inside the package, while being replaceable with previous SKUs with slower DRAM.

It would be kind of funny if at the same time Intel were to retreat from MoP, (after Intel explaining over and over it is not ideal technology) and AMD introduces Medusa Halo with MoP and kills it...

MoP is nice until you see the cost.
You take on the burden of the memory cost when you usually only sell the chip.
You also give OEMs little flexibility, they do not like MoP at all.
Apple are the only ones shipping it en masse because they are fully vertically integrated and charge massive premiums for memory.
CAMM is a good bump in the future, Intel also has very interesting expandable memory tech which could even work for GDDR memory expansion on GPUs one day.

 

[poke01]

AMD did say MoP was avoided on purpose.

 

[Joe NYC]

OEMs seemingly hate MoP. That's why Lunar Lake was a flop.

If OEMs did not like MoP, they had an option to skip / boycott Lunar Lake. They did not.

Also, I don't think Lunar Lake is a flop.

 

[poke01]

If OEMs did not like MoP, they had an option to skip / boycott Lunar Lake. They did not.

Also, I don't think Lunar Lake is a flop.

It’s also the only x86 SoC that does video playback for a long time. Pretty nice chip. Those PMIC, media engines and MoP help a lot

 

[Joe NYC]

AMD did say MoP was avoided on purpose.

I think odds are better than 50:50 that NVidia will include MoP.

Also, maybe AMD thinks it is too risky to copy more than 1 feature from Apple in 5 years (powerful integrated GPU). We may have to wait another 5 years for AMD to copy 2nd successful feature from Apple - MoP.

Seriously, I don't know how AMD is planning on scaling bandwidth to maintain leadership if 12700 speeds are going to be available, and AMD can't even reach full speed of 8533.

Similarly, Qualcomm was unable to support full speed of 8533 without MoP.

 

[branch_suggestion]

Also, maybe AMD thinks it is too risky to copy more than 1 feature from Apple in 5 years (powerful integrated GPU).

AMD were the first to ship powerful integrated GPUs, little thing called 8th gen consoles.
Back then Apple was using discrete GPUs, AMD's very own.

We may have to wait another 5 years for AMD to copy 2nd successful feature from Apple - MoP.

PC is too cost sensitive, you need to focus on area, Apple can afford to bloatmaxx for power savings.
Now Halo parts do have high enough ASPs and low relative volumes compared to standard PC SoCs, so they can afford to do those things.
But I think MoP will be avoided to keep partners happy, MALL is a thing for good reason, as is CAMM.

Seriously, I don't know how AMD is planning on scaling bandwidth to maintain leadership if 12700 speeds are going to be available, and AMD can't even reach full speed of 8533.

There is a third dimension in play.

 

[adroc_thurston]

We may have to wait another 5 years for AMD to copy 2nd successful feature from Apple - MoP.

They explicitly avoided MoP on stxH for cost reasons.
ATi shipped MoP before iPhone was even a goddamn thing ffs.

if 12700 speeds are going to be available

Not memory on board.

 

[Joe NYC]

Not memory on board.

My assumption would be that the 12700 speed would be achievable as MoP but not as memory on board.

 

[adroc_thurston]

My assumption would be that the 12700 speed would be achievable as MoP but not as memory on board.

as always with cutting edge LPDDR, it's PoP, with the rest being 'eventually' or 'maybe'.