Discussion AMD SoC Halo series GPU discussion

[gdansk]

Definitely not a smaller memory bus

That is his whole point on using the halo in the handheld

There's little point to it if the GPU is smaller/lower clocked. Basically what those upcoming handhelds want is a Zen 5 + RDNA4 SoC with LLC. Strix Halo is a bad fit for handhelds due to RDNA3.5 and also the typical <20W TDP.

 

[GTracing]

There's little point to it if the GPU is smaller/lower clocked. Basically what those upcoming handhelds want is a Zen 5 + RDNA4 SoC with LLC. Strix Halo is a bad fit for handhelds due to RDNA3.5 and also the typical <20W TDP.

If you watch the video, one of his main points is that Strix Halo heavily downclocks the memory bus when running under ~30 watts. Future 256bit CPUs (either Strix Halo's successors or custom 256bit gaming chips) could heavily improve 20-25W gaming performance over Strix Halo by tuning the memory bus to run faster at lower wattages.



I don't think that a handheld with a 256bit memory bus makes sense, but I also don't think we can extrapolate 25W gaming performance for all 256bit CPUs from Strix Halo.

 

[LightningZ71]

Jeez, it doesn't have to be that complicated. Just take KRK, ditch the NPU, replace it with 16MB MALL cache and grow the whole thing to fit 16 CUs instead of the 8 is has now. That's MORE than enough for a competent handheld chip. It would have FAR greater gpu performance than PHX/HPT, able to sustain higher clocks and with more effective memory efficiency. 128bit LPDDR5X would give it plenty of memory throughput. 4+4 Zen 5 is more than enough compute for gaming with that level of GPU. If you're growing the SOC, you can ditch a lot of the shoreline I/O as it's just going to be used in a handheld with very few ports.

 

[marees]

128bit LPDDR5X would give it plenty of memory throughput.

That would consume more power

For handheld mode 256 bits is preferable

 

[jpiniero]

ditch the NPU

AI AI AI tho

 

[adroc_thurston]

He wants Microsoft/valve to order a custom apu/soc from AMD modelled on strix halo

entirely pointless.
In fact AMD already told Valve to pound sand.

 

[Bryo4321]

In fact AMD already told Valve to pound sand.

how so?

 

[Mopetar]

how so?

Well someone needs to make the silicon wafers for the chips. I guess AMD expects it to be Valve.

 

[igor_kavinski]

In fact AMD already told Valve to pound sand.

Valve's projected sales numbers must've been too abysmal for AMD to even consider making anything custom for them.

I guess Valve must now go to...

INTEL!

Six Skymont cores + enough Celestial cores to do 1080p 60fps with raytracing

 

[Kronos1996]

Valve's projected sales numbers must've been too abysmal for AMD to even consider making anything custom for them.

I guess Valve must now go to...

INTEL!

Six Skymont cores + enough Celestial cores to do 1080p 60fps with raytracing

That’s not how it works from what I understand. AMD’s semi-custom group will make anything you want for the right price. One question I have is if Sony and Microsoft are responsible for ordering wafers directly from TSMC and basically having their own WSA contracts?

 

[adroc_thurston]

That’s not how it works from what I understand

Yeah it is.

AMD’s semi-custom group will make anything you want for the right price.

Nope, you need baseline volume targets.

 

[Mopetar]

If Valve just wanted to order a million Stix Halo dies I'm sure AMD would have no problems selling them as many, wafer availability permitting, but custom silicon needs its own set of masks on top of any design work and that adds significant costs that someone needs to pay for.

 

[Kepler_L2]

entirely pointless.
In fact AMD already told Valve to pound sand.

Gaben knows he can just wait for MS to cancel their own handheld and re-use the SoC.

 

[poke01]

we finally got numbers and Halo is good in a 14" laptop. Sidenote - The 275HX is just sad of a product.


That being said HP screwed up big time in the battery life test. Battery life sucks, esp video playback.

 

[poke01]

Also Windows is a pain for AI.....
oh one last thing Thunderbolt on AMD laptop yay

 

[adroc_thurston]

That being said HP screwed up big time in the battery life test. Battery life sucks, esp video playback.

It's a byproduct of using a 1800p@120Hz OLED.

 

[poke01]

It's a byproduct of using a 1800p@120Hz OLED.

yep also the fact HP only offers 48hz or 120hz means you are stuck on 120hz. like wtf HP?

 

[adroc_thurston]

yep also the fact HP only offers 48hz or 120hz means you are stuck on 120hz. like wtf HP?

oh yeah if it does 120hz on battery at 1800p then it's nuclear holocaust for battery life under all circumstances.

 

[okoroezenwa]

yep also the fact HP only offers 48hz or 120hz means you are stuck on 120hz. like wtf HP?
View attachment 122389

No 60Hz? That is so weird.

 

[LightningZ71]

Well, at least we now know which laptop everyone is going to pull out to show that AMD "sucks". That HP one is just a very bad design choice/implementation.

 

[MS_AT]

Also Windows is a pain for AI.....

Why? I haven’t noticed why it would be worse for local AI inference than Linux.

 

[fastandfurious6]

The VAST majority of customers have ZERO idea what's inside of the SoC. A few know the relative performance without understanding why. The rest of us on the forums make up such a tiny portion that they don't really care about us.

big time

this should be permanently pinned as banner in the entire forum / all threads

 

[igor_kavinski]

AMD Ryzen AI Max+ "Strix Halo" Delivers Best Performance On Linux Over Windows 11 - Even With Gaming - Phoronix

www.phoronix.com

AMD Radeon 8060S Linux Graphics Performance With Strix Halo Review - Phoronix

www.phoronix.com

 

[MS_AT]

To be fair the Linux vs Windows comparison is a bit biased, considering it was out of the box Win11 corporate tuned vs fresh Ubuntu install. The amount of "useful" software running in the background on Win11 including HP's wolf security could well, skew the results a bit. Not to mention no data about power consumption, when Win usually is using more tuned power profiles.

https://blog.hjc.im/strix-halo-local-llm.html here Halo LLM test from David Huang using fairly vanilla llama.cpp

 

[igor_kavinski]

https://blog.hjc.im/strix-halo-local-llm.html here Halo LLM test from David Huang using fairly vanilla llama.cpp

Text generation seems sadly slow