Discussion AMD SoC Halo series GPU discussion

[marees]

After the massive market failure that was pricing Strix Halo well above any reasonable value perception, it looks like the new 388 and 392 models are being sold for cheap as Asus finally puts them in one of their value brand TUF laptops:

Asus puts Strix Halo in the TUF Gaming A14 — strong integrated graphics in a very thin chassis

Welcome, Ryzen AI Max+ 392

www.tomshardware.com

The laptop looks pretty good, actually. Too bad this didn't come last year. Looks like Panther Lake is already putting some pressure on AMD's APUs.

there is also now a rOCM box aka dgx spark competitor direct from AMD

 

[DavidC1]

Nah thats completely backwards, if you wanna buy a 128gb ryzen 395, its over 1k less than buying a similar apple product. But apple does have a perf/watt advantage at least in some tasks. And thats including all the markup from framework, there are much cheaper options for getting strix halo.

The PC vendors do not have the reputation Apple does. It sells way less than Apple for this reason.

Thats what people thought about $2000 GPUs back in 2018, look at us now.

This is one reason why Steam Deck is still king.

Unfortunately in future with the price hike of GPU and RAM, we would be only able to purchase processors like AMD Strix Halo with built-in memory and GPU. That is just sad. few weeks ago I was thinking of buying 64GB RAM. Now that is out of the question, I should have pulled the trigger earlier.

The built-in memory gives us no choice in memory price flexibility. Such future will get worse in terms of pricing, beyond demand inflated prices such as today.

The laptop looks pretty good, actually. Too bad this didn't come last year. Looks like Panther Lake is already putting some pressure on AMD's APUs.

The 10 Xe core version should achieve 90% of the top version and be available on laptop $799 and below. This is one reason why Halo iGPUs struggle, among others.

 

[Claudiovict]

yeah, mdsH is Q2'27.

given this is now the date for desktop RDNA 5, this still the date for Medusa Halo?

 

[adroc_thurston]

given this is now the date for desktop RDNA 5, this still the date for Medusa Halo?

it's AT3/4 so round there yeah.

 

[dangerman1337]

it's AT3/4 so round there yeah.

But when do we see actual products? @Kepler_L2 thinks not until 2028, maybe ITX based systems could be released or even motherboards like we're seeing with Strix Point at CES? A 388 Medusa Halo successor with 32-48GB of LPDDR5X RAM all in a single board would be a pretty cool "entry level" Performance PC.

 

[jpiniero]

But when do we see actual products? @Kepler_L2 thinks not until 2028, maybe ITX based systems could be released or even motherboards like we're seeing with Strix Point at CES? A 388 Medusa Halo successor with 32-48GB of LPDDR5X RAM all in a single board would be a pretty cool "entry level" Performance PC.

Medusa Halo is gonna be mega expensive... and 32-48 GB of RAM could easily be way more than that.

 

[Joe NYC]

But when do we see actual products? @Kepler_L2 thinks not until 2028, maybe ITX based systems could be released or even motherboards like we're seeing with Strix Point at CES? A 388 Medusa Halo successor with 32-48GB of LPDDR5X RAM all in a single board would be a pretty cool "entry level" Performance PC.

The MiniPC industry is growing, they are faster to bring new chips to market.

It went from novelty to mainstream to use notebook chips in MiniPCs. Maybe this time, the MiniPC will be even more responsive to Medusa Halo release (assuming AMD has plentiful stock on release).

 

[adroc_thurston]

But when do we see actual products? @Kepler_L2 thinks not until 2028, maybe ITX based systems could be released or even motherboards like we're seeing with Strix Point at CES? A 388 Medusa Halo successor with 32-48GB of LPDDR5X RAM all in a single board would be a pretty cool "entry level" Performance PC.

AMD has EEPs now so you should not worry about it.
They're focused on laptops so forget about the whole PC shebang.

Medusa Halo is gonna be mega expensive...

Nope. the whole point is to source it from commodity h/w. AT3 is the mainstream dGPU also.

 

[Claudiovict]

The MiniPC industry is growing, they are faster to bring new chips to market.

It went from novelty to mainstream to use notebook chips in MiniPCs. Maybe this time, the MiniPC will be even more responsive to Medusa Halo release (assuming AMD has plentiful stock on release).

i imagine stock wont be a problem since Zen 6 will be launched much earlier and RDNA 5 was supposed to be launched this year but moved to h2 27

 

[Claudiovict]

Medusa Halo can achieve GPU performance around a desktop 5070 (or even above) or my math isn't mathing?

1,25 x 1,20 x 1,05 = 1,57x 8060s performance. This would put Medusa Halo at 4070 Super perf, which falls behind only 6% against 5070

Math explanation:
Considering 8060s beats 6700 xt by 6% (here), i used 3060 ti as baseline (it beats 6700 XT by 2%).

1,25 -> A uplift of 25% from RDNA 3 to RDNA 4 (9060 XT 16gb got 35% more performance 7600 XT 16gb having same CUs count; 9070 XT got 33% more perf despite having 16 less CUs than 7900 GRE, so i low balled it to 25% to be reasonable).

1,20 -> The uplift of RDNA 5 vs RDNA 4 (100% pulled of my ahh).

1,05 -> 5% uplift from getting 20% more CUs (48 vs 40 of 8060s).

 

[reaperrr3]

Medusa Halo can achieve GPU performance around a desktop 5070 (or even above) or my math isn't mathing?

1,25 x 1,20 x 1,05 = 1,57x 8060s performance. This would put Medusa Halo at 4070 Super perf, which falls behind only 6% against 5070

Math explanation:
Considering 8060s beats 6700 xt by 6% (here), i used 3060 ti as baseline (it beats 6700 XT by 2%).

1,25 -> A uplift of 25% from RDNA 3 to RDNA 4 (9060 XT 16gb got 35% more performance 7600 XT 16gb having same CUs count; 9070 XT got 33% more perf despite having 16 less CUs than 7900 GRE, so i low balled it to 25% to be reasonable).

1,20 -> The uplift of RDNA 5 vs RDNA 4 (100% pulled of my ahh).

1,05 -> 5% uplift from getting 20% more CUs (48 vs 40 of 8060s).

Provided enough power and using LPDDR6, that sounds plausible.

An RDNA4 CU already performs at least on par (per clock) with a Blackwell SM in Raster.

MDS-H CU count matches 5070, so clockspeed in mobile TDPs and mem bw should be the only potential bottlenecks.