Discussion AMD SoC Halo series GPU discussion

[misuspita]

I was eyeing a mini itx, Gigabyte b650i aorus ultra but I have a 5700g now and the only meaningful upgrade is a Halo, cause I need/want the extra GPU power in a small portable case, and a silent one. And no halo for desktop, so I was already content into buying a Beelink or Minisforum Halo when they would show up, but this is much better.

I have a passive case now, but it already gets hot when my current 5700g goes full tilt in work or games, so it definitely won't work with a 120W Halo. Sadly, but yeah, a Noctua will be keeping things quiet in that MQ4 case. I want.

Still debating if I should wait for Medusa Halo, cause that's where I think I'll get double the CPU performance (I need it for audio production, Studio one 7) and GPU cause when I play, I want my 40+ fps în 1440. That's all. Now I don't have that and I miss itso in the end I might just get this and be dome for the next 10 years )

 

[igor_kavinski]

Still debating if I should wait for Medusa Halo, cause that's where I think I'll get double the CPU performance (I need it for audio production, Studio one 7)

https://www.reddit.com/r/StudioOne/comments/1hx73xt/has_studio_one_7_improved_upon_cpu_efficiency_and

One guy there saying that Cubase 14 is much faster.

 

[misuspita]

Doesn't matter, I've already paid for it, and I've been using it since version 2 (10 years ago? I think?). So for me, having hundreds of projects in my current DAW that may need ressurection from some artist that needs a change, a remix, whatever, that's not happening. Also, the habits, quirks and whole work around, I am familiar with. E-learning something new? F it.

Besides, my 5700G is working great, it's 70% lust for hardware, not a real need. But oooh, boy, i was already settled in my mind that there will not be a miniITX version unless Minisforum does it. I'm pleasantly surprised. And price? Damn... 64GB of fast ram, a 8GB Samsung 9100 M2 Drive for virtual instruments and I'm set for the next decade.

 

[igor_kavinski]

After enjoying better responses from the 236B Deepseek model, 128GB suddenly feels paltry to me. RAM consumption exceeded 200GB

 

[marees]

After enjoying better responses from the 236B Deepseek model, 128GB suddenly feels paltry to me. RAM consumption exceeded 200GB

You can go for 2x epyc with 768gb ram (7 tokens/sec without GPU)

 

[igor_kavinski]

You can go for 2x epyc with 768gb ram (7 tokens/sec without GPU)

Way out of my price range but thanks

 

[misuspita]

Or use more Halos? Framework allows multiple units to be linked.

 

[fastandfurious6]

After enjoying better responses from the 236B Deepseek model, 128GB suddenly feels paltry to me. RAM consumption exceeded 200GB

how does that compare to web deepseek? does web deepseek always use the full model? seems to me they randomly switch to much weaker versions very frequently

medusa should have 256gb version, and 384GB if using bigger camm2 will be awesome

or LLMs will be so advanced at that point they can retrofit in smaller mem size

 

[igor_kavinski]

how does that compare to web deepseek? does web deepseek always use the full model? seems to me they randomly switch to much weaker versions very frequently

Haven't tried the web one but I assume they will try to return quicker and weaker responses if many users are in queue.

 

[techjunkie123]

GPU and CPU scaling with power for the top spec halo part. Seems like 60-70 W is the sweet spot for both.

 

[misuspita]

I'm curious if that's the case for this particular implementation or is the same even in a 120W capable Framework. Because the perf doesn't seem to cap at all. I guess at 120-140w in a desktop will be even better with the better cooling

 

[adroc_thurston]

I'm curious if that's the case for this particular implementation or is the same even in a 120W capable Framework

Why would "implementation" affect v/f scaling?

 

[misuspita]

Don't know, asking... The performance doesn't seem to flatten, just goes up, and R23 multicore has a jump at 90W... So it seems possible the perf might be better.

 

[igor_kavinski]

Or use more Halos? Framework allows multiple units to be linked.

If they can manage 512GB RAM for an additional $800, that would be a lot better. $3000 is the limit for me for something like that.

 

[scineram]

Sam

I was eyeing a mini itx, Gigabyte b650i aorus ultra but I have a 5700g now and the only meaningful upgrade is a Halo, cause I need/want the extra GPU power in a small portable case, and a silent one. And no halo for desktop, so I was already content into buying a Beelink or Minisforum Halo when they would show up, but this is much better.

I have a passive case now, but it already gets hot when my current 5700g goes full tilt in work or games, so it definitely won't work with a 120W Halo. Sadly, but yeah, a Noctua will be keeping things quiet in that MQ4 case. I want.

Still debating if I should wait for Medusa Halo, cause that's where I think I'll get double the CPU performance (I need it for audio production, Studio one 7) and GPU cause when I play, I want my 40+ fps în 1440. That's all. Now I don't have that and I miss itso in the end I might just get this and be dome for the next 10 years )

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.

 

[MS_AT]

If they can manage 512GB RAM for an additional $800, that would be a lot better.

Physically impossible, Halo has only 4 mem channels. And if you link them together then you can reach 512GB of capacity but with relatively slow interconnect in between and at a price that would make going with Epyc and spamming DRAM channels cheaper.

 

[Glo.]

So, the 799$ Framerwork board means that after the BOM cost of RAM, board, cooler, etc and adding the margin to the price, 8C/32 CU SOC appears to be sold by AMD to partners for around or less than 500$.

And it still might have VERY good price margin on itself.

 

[fastandfurious6]

So, the 799$ Framerwork board means that after the BOM cost of RAM, board, cooler, etc and adding the margin to the price, 8C/32 CU SOC appears to be sold by AMD to partners for around or less than 500$.

And it still might have VERY good price margin on itself.

halo 8core 32gb ($799) is around $100 cheaper than the desktop equivalent parts i.e. 9700x + 32gb + cheapest mobo + rtx4060 - just these parts without anything else

 

[MrTeal]

halo 8core 32gb ($799) is around $100 cheaper than the desktop equivalent parts i.e. 9700x + 32gb + cheapest mobo + rtx4060 - just these parts without anything else

You should be able to do that cheaper, unless you're specifically looking for a mITX board. $310 9700X + $110 B650M + $75 32GB DDR5-6000CL30 + $305 4060 puts the equivalent right at $800 as well.

 

[gdansk]

You should be able to do that cheaper, unless you're specifically looking for a mITX board. $310 9700X + $110 B650M + $75 32GB DDR5-6000CL30 + $305 4060 puts the equivalent right at $800 as well.

And a 9700X and desktop RTX 4060 should be faster at the same price.
But I think mITX board is an important part of the comparison. Halo's advantage is power consumption and size. That should allow a more compact case and quieter Flex ATX power supply. And more potential VRAM allocation when you need it.

 

[GTracing]

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.

 

[adroc_thurston]

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

no, because that's niche^3. just a difficult platform to ship.

A 96CU RDNA5 iGPU with a 512bit memory bus could be the single fastest GPU available in any laptop

Yeah but you gotta MoP to get to 512b.

 

[GTracing]

no, because that's niche^3. just a difficult platform to ship.

You're probably right, but one can dream.

 

[jpiniero]

You're probably right, but one can dream.

Alll the AI!!

 

[Heartbreaker]

And a 9700X and desktop RTX 4060 should be faster at the same price.
But I think mITX board is an important part of the comparison. Halo's advantage is power consumption and size. That should allow a more compact case and quieter Flex ATX power supply. And more potential VRAM allocation when you need it.

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