What do we think about Gorgon Point (Ryzen AI 9 HX 470)?
Posted by Everlier@reddit | LocalLLaMA | View on Reddit | 47 comments
The new APU is promised to support DDR5-6400 (102.4 GB/s) and LPDDR5X-8533 (136.5 GB/s) which should move some models that were barely usable on Strix Point to the usable territory.
However, it really seems that to utilise these capabilities, manufacturers would have to get chips that are basically inaccessible right now.
kingwhocares@reddit
It's going to be worse than Strix Halo which offers not only 256 GB/s but also has over 75 TOPS from igpu alone (same NPU). Heck, Intel's Nova Lake is gonna offer 75+ TOPS from NPU alone.
OpenMinded00@reddit
Intel Nova Lake still doesn't exist, the HX 470 you can buy and it's way beyond current intel SoC
Advanced_Paper_5061@reddit
What are we a hive mind, the question is what do you think about it?
ViRROOO@reddit
Gorgon Point is a mid-cycle refresh, its not a replacement for strix halo. That will only come around 2027 as medusa halo (as announced before). I wouldnt have my hopes up, it wil definitely be better then the current gen, but nothing crazy or even something I would consider replacing my strix halo for.
simracerman@reddit
I have a strict point with LPDDR5X 8000mt/s and this is not even an upgrade. Unless I’m missing something.
ViRROOO@reddit
Yeah, I cant understand why they even did LPDDR5X. My cynicism tells me that they did LPDDR5X just so they have a "killer" upgrade for medusa halo with LPDDR6 (maybe even with 384-bit), or could just be that the chip design didnt allow it and it required more time. Could be that they knew that the latency within the strix halo chip design would be a bottleneck regardless of the memory bandwidth and thats why medusa halo is said to be a complete chip redesign.
Meanwhile Intel will drop the most mid possible competitor with Intel Nova Lake-AX, its rumored that it will also have soldered LPDDR5X, which is laughable if true, they will be almost 2 years late to the party, with a 1 year away of medusa halo, with something that has a small edge against strix halo. Intel is so cooked.
Mochila-Mochila@reddit
Intel won't be cooked if they release something competitive with Strix Halo.
ViRROOO@reddit
Why would anyone buy an strix halo-like product, 2 years into the strix halo cycle, with untested software and support, if you can wait a year and get an objectively better product? Or, like, buy strix halo now..
redoubt515@reddit
Why would anyone buy that "objectively better product" in a year, when you could just wait 2-3 years, for an even better product or 3-5 years for an even better-er product?
The answers to this question will always be: either (1) a current need, or (2) impatience, or (3) "porque no los dos?"
ViRROOO@reddit
If you have a current need or impatience you buy strix halo, as its already available, has extremely similar specs, and with the advantage of 2 years of bug fixing and community support at the time of release of the intel offering.
QuickQuirk@reddit
This is the same product. It's a refresh year. No new technology. You've already got the benefits of driver improvements, etc, with a minor performance boost.
Donut_LordO@reddit
Doubtful, Intel will be fine. Their Panther Lake X looks good
simcop2387@reddit
I haven't heard much about it myself but my guess would be this would probably be a bit more energy efficient even if it's not any more perofrmant on memory bandwidth or compute at least. Probably not cost effective for an upgrade but maybe for anything new it'd be a better value still because of that.
waiting_for_zban@reddit
But this was never an issue with Strix Halo. The thing idles 9W fully loaded with GPT-OSS-120B (when not actively inferecing). Getting it to 6W is not a big improvements imo.
This is more of a successor of the 370, which is welcome imo, but mainly I would say targeted towards gaming.
Firm-Fix-5946@reddit
well it's true that energy efficiency was never an issue with strix halo, but who gives a shit about idle power? you buy the chip just to let it sit there and do nothing?
mindwip@reddit
It's not the 470 is refresh of 370 not the 395 strix halo.
There is a hole 400 series refreshing the 300 series.
Including new 385 that might have lower cores but same memory as 395. Meant to be cheaper for laptops. But will have same GPU core count
SpicyWangz@reddit
That release will probably be on DDR6 which would also bring fairly usable speeds for LLMs. I think that will be a game changer release
Nyghtbynger@reddit
I thought the CUDIMM DDR5 was the next standard to be adopted (they move the clock onto the chips and reconcile the signals). It allows for easy 12000MT/s. With the 200GB/s bandwidth on two stick the then 400GB/s would enable easy local use for bigger models :D
ItilityMSP@reddit
Right if you can afford the memory!
Limp_Classroom_2645@reddit
underwhelming
IngwiePhoenix@reddit
Am I the only ones gawking at the PCIe lane count and thinking, "wait, this is mobile, right?"
If we see more MoDT boards, you could get a solid x8 slot, slap another GPU on it (or bifocate into 2x4 for a real serious bifocation hecmeck XD) and add even more compute.
I kinda like this chip. Should do quite nicely.
warnerbell@reddit
The unified memory approach is appealing for local inference. No more juggling VRAM limits.
That said, I've found bigger context window doesn't always mean better results. There's a sweet spot before quality drops off. Curious what context lengths people are actually using effectively on similar hardware?
Kubas_inko@reddit
That RAM is expensive.
power97992@reddit
Where is the 80W 2 petaflop gpu with 192gb of 4 TB/s RAm for less than 1.5 k? Maybe in 4-5 yr
AnomalyNexus@reddit
That’s the npu not gpu
Slasher1738@reddit
its fine, need to start seeing more models use the NPU
Slow_Concentrate3831@reddit
Looks better than HX 370, but still less interesting than Ryzen AI Max 395 with his 256 bits bus.
RnRau@reddit
Yup... less than half the theoretical memory bandwidth.
Look_0ver_There@reddit
Came here to say exactly this. Gorgon Point barely moves the needle. The AI Max series are the better direction to be expanding upon.
UniqueAttourney@reddit
Did they announce it already ?
RnRau@reddit
https://www.youtube.com/watch?v=UbfAhFxDomE
Everlier@reddit (OP)
In a few hours:
https://www.tickcounter.com/widget/countdown/9072533
UniqueAttourney@reddit
hh, i also found it on youtube
lacerating_aura@reddit
I doubt itll be good even if ignoring the current chip issues. I never liked the current tech scene where they have to push yearly new stuff. Most of it is pointless. I really wish for a world where we have longer cycles of say 3-4 years and the next generation of tech is actually multiple times better than the previous one, not just fractional improvements.
deadwisdom@reddit
This is the world I live in because I don’t pay attention.
SkyFeistyLlama8@reddit
It takes a long time for software to catch up to the hardware. For example, Qualcomm released the Snapdragon X chips back in May 2024 with Llama being shown to run on it, but NPU support for LLMs only dropped a year after release. CPU and GPU support on llama.cpp also took a couple of months to show up and this was with Qualcomm engineers pitching in.
AMD needs to do a heck of a lot more on the software side if it wants these APUs to be useful for LLMs and local AI.
Pale_Reputation_511@reddit
thats because NPUs are a real PIAS, every NPU its different, there is no api standard yet.
snoodoodlesrevived@reddit
I’ve seen them funding some open source, idk if they’re doing more, but that’s definitely the move
Admirable_Bag8004@reddit
Isn't that what's happening already? You can just wait & ignore news for 3-4 years and get new CPU then. I don't see how having more choice with yearly updates is a bad thing in any way.
lacerating_aura@reddit
That more choice doesn't just pop out just like that. Money, time, materials, everything is spent trying to make that, in my opinion, pointless fractional upgrade. I get it that more choice is better, but we have to think of the costs of that choice too, in every aspect. Plus many times, in case of amd, they can't even have decent stock for a lot of their offerings. And now we all know what's happening with closed ai and chips.
FullstackSensei@reddit
It's also a balancing act. With how expensive tape outs have become, companies are less willing to take big swings with major architectural changes unless they're forced to. The yearly cycle enables teams to test small changes and scale up if successful or revert if not without too much impact on the company's bottomline.
lacerating_aura@reddit
I guess i didn't think of that aspect. Thanks. But still one can wish.
_VirtualCosmos_@reddit
but but but consumerism...
PotentialFunny7143@reddit
DDR6 ram is what people are really looking for
ImportancePitiful795@reddit
470 is the successor of 370. Big HELL NO.
Still worse than 395.
teh_spazz@reddit
God forbid they give us more lanes
cibernox@reddit
I think that the real revolution will come once this kind of APUs use DDR6. But apple will be the first. And will do it with more memory lanes too