Support NeoGAF

[Shogmaster]

Definitely getting the 10090XT If they really release one and a Ryzen9800X3D(successor) so I most likely will have still better performance than a PS7 Pro.

PS7 Pro? Is that the one that you snort up your nose or is it the one you stick up your ass?

 

[riko]

No, I need a proper desktop gaming rig, not some gimped bullshit AI bait product with unified memory crap. 24+GB will do just fine for the time being, QWEN and other LLM can perform quite well within those constraints already.

QWEN3 14B 4-bit on a 24gb GPU is going to performance worse than any of the tiny and very cheap frontier models. I don't know why you'd bother over gpt5.4-mini, tokens are cheap. You can run 32B barely, but good luck with context.

If you're doing work of value these days you're using Sonnet/Opus anyway, nothing FOSS can get anywhere close, even on a pair of RTX 6000s.

 

[Shogmaster]

LPDDR on dgpus? Seems like a terrible idea.

Terrible idea? Or genius idea? Imagine if AMD offers RX 10060XT AI edition with 64GB of LPDDR5X for $250 more than the regular edition.

Sub $750 AI card with gaming performance better than 4080 Super? ARC B70 who?

 

[//DEVIL//]

You guys really don't learn....

Hyping yourselves just to get slapped in the face by reality in the end.
Every AMD GPU gen the same drama.

5090ish performance for a grand, lol....

No card will be touching the 5090 in performance but the 6090.

I highly doubt the 6080 will even come close considering the difference between the 5080 and 5090.

Even the 4090 is superior than the 5080 in everything outside of multi frame gen.

 

[riko]

Terrible idea? Or genius idea? Imagine if AMD offers RX 10060XT AI edition with 64GB of LPDDR5X for $250 more than the regular edition.

Sub $750 AI card with gaming performance better than 4080 Super? ARC B70 who?

I mean 250 bucks is straight up impossible today but assuming a more realistic price…
for what use case? Tinkering very slowly? The frontier models are so far ahead of open source.

Sure you can get some kinda fun inference going at very slow speeds, but a Mac Mini does that too if you are willing to give up CUDA. ROCm PyTorch is still a pita.

 

[HerjansEagleFeeder]

1. LLM values RAM amount more. 24GB will force you to use heavily quantized models. Hallucinations all day lol

2. No practical difference between AT3 in discrete card and the APU version as far as performance is concerned, unless some card ODM decides to juice the shit out of the card for OC.

What are we doing anyways? Meme image generation? Vibe coding? Making those shitty AI voice over YouTube videos? I suppose that's light years better than starting your videos with "What's up YouTube It's yer Boi blah blah blah" lol

Yeah I'm a software engineer, I don't do social media.

A 4090 smokes any unified memory system as VRAM is a LOT faster for AI related tasks than regular memory/RAM. That goes for BOTH prompt digestion AND token output. There's a reason 4090/5090 cost as much as they do.

Also your assumption regarding the amount of halluzinations is misinformed, as that varies greaty within the same model depending on the task.

 

[HerjansEagleFeeder]

QWEN3 14B 4-bit on a 24gb GPU is going to performance worse than any of the tiny and very cheap frontier models. I don't know why you'd bother over gpt5.4-mini, tokens are cheap. You can run 32B barely, but good luck with context.

If you're doing work of value these days you're using Sonnet/Opus anyway, nothing FOSS can get anywhere close, even on a pair of RTX 6000s.

Seems you know EXACTLY what my private workload looks like as you're throwing around lots of half baked recommendations taken from some stackoverflow or reddit thread. I use Anthropics products daily for work, I know what they can do. That's not what I'm after for strictly private use though.

 

[riko]

Seems you know EXACTLY what my private workload looks like as you're throwing around lots of half baked recommendations taken from some stackoverflow or reddit thread. I use Anthropics products daily for work, I know what they can do. That's not what I'm after for strictly private use though.

Well I am an ML guy and an exec at a fairly sizable AI company. I know the tech very well.

Enjoy your porn, sorry, PRIVATE use case I guess? FWIW you can sign zero retention deals with both Anthropic and OpenAI with nothing upfront and do whatever you like (No Fable for zero retention customers)

Thanks for the personal insults tho

 

[Shogmaster]

I mean 250 bucks is straight up impossible today but assuming a more realistic price…
for what use case? Tinkering very slowly? The frontier models are so far ahead of open source.

Sure you can get some kinda fun inference going at very slow speeds, but a Mac Mini does that too if you are willing to give up CUDA. ROCm PyTorch is still a pita.

Ummmm..... nVidia just AIjeculated all over Computex with N1X Spark which has far less performance than AT3 so...

 

[riko]

Ummmm..... nVidia just AIjeculated all over Computex with N1X Spark which has far less performance than AT3 so...

I know, and I have friends who are all in on Spark, but I am just not seeing it. We had a few DGX units for testing.

Best desktop setup is still a pair of RTX6000, but $$$$

 

[Shogmaster]

Yeah I'm a software engineer, I don't do social media.

A 4090 smokes any unified memory system as VRAM is a LOT faster for AI related tasks than regular memory/RAM. That goes for BOTH prompt digestion AND token output. There's a reason 4090/5090 cost as much as they do.

Also your assumption regarding the amount of halluzinations is misinformed, as that varies greaty within the same model depending on the task.

AMD also has 384bit LPDDR6 option for AT3. 614 GB/s aint exactly 1TB+/s of GDDR7 but it aint total crapola either, ya?

 

[riko]

AMD also has 384bit LPDDR6 option for AT3. 614 GB/s aint exactly 1TB+/s of GDDR7 but it aint total crapola either, ya?

I think that'll compete well with the Apple solutions, but still kinda pokey for big models.

 

[asdasdasdbb]

AMD also has 384bit LPDDR6 option for AT3. 614 GB/s aint exactly 1TB+/s of GDDR7 but it aint total crapola either, ya?

I'd hate to see how expensive that would be. AT2 might be cheaper at that rate.

 

[Shogmaster]

I think that'll compete well with the Apple solutions, but still kinda pokey for big models.

Im sure AMD will full retard an AI version of AT0 with 96GB of VRAM on 512bit bus...

 

[Shogmaster]

I'd hate to see how expensive that would be. AT2 might be cheaper at that rate.

No way LPDDR6 gonna be more costing than 36Gbps GDDR7. That'd be cray cray.

 

[asdasdasdbb]

No way LPDDR6 gonna be more costing than 36Gbps GDDR7. That'd be cray cray.

384-bit versus 192-bit. AI has fallen in love with LPDDR too so it's not like the chips would be much cheaper.

 

[Shogmaster]

384-bit versus 192-bit. AI has fallen in love with LPDDR too so it's not like the chips would be much cheaper.

Meh, that just few more mm² on the IO die. And yes, LPDDR5X got AS EXPENSIVE as GDDR7 right now bc of AI data centers, but more than? No way Ho-Zay~

 

[riko]

Meh, that just few more mm² on the IO die. And yes, LPDDR5X got AS EXPENSIVE as GDDR7 right now bc of AI data centers, but more than? No way Ho-Zay~

Part of the problem isn't the specific memory types having demand, its that the manufacturers are shifting capacity away from LPDDR and to HBM.

 

[SScorpio]

Based on what Kepler said, there won't be a Z3 Extreme. AMD is losing interest and not investing in the handheld space

That would be a pretty classic AMD move. Stop investing in handheld. The AMD powered PS5/6 Portable releases and raises interest. Intel or NVIDIA push their solution and take over the PC segment.

 

[asdasdasdbb]

That would be a pretty classic AMD move. Stop investing in handheld. The AMD powered PS5/6 Portable releases and raises interest. Intel or NVIDIA push their solution and take over the PC segment.

I thought the Z products were just the regular APUs with the NPU disabled. I don't think AMD would stop OEMs from just using the regular models instead.

 

[Shogmaster]

I thought the Z products were just the regular APUs with the NPU disabled. I don't think AMD would stop OEMs from just using the regular models instead.

Exactly. AT4 is your handheld chiplet. Madusa Premium for everybody!

 

[hardwood007]

Exactly. AT4 is your handheld chiplet. Madusa Premium for everybody!

Is Medusa premium going to use rdna5? I lost track with AMD's naming

 

[Shogmaster]

Is Medusa premium going to use rdna5? I lost track with AMD's naming

24 CU RDNA 5. 128bit LPDDR5X. Probably 24 to 32 GB.

 

[rm082e]

LPDDR on dgpus? Seems like a terrible idea.

Only on the lower end cards. The medium and high end get GDDR7.

 

[hardwood007]

24 CU RDNA 5. 128bit LPDDR5X. Probably 24 to 32 GB.

Fuck, it's still a 128bit memory bus. Same problem of memory bandwidth bottleneck

 

[kasrkinsquad]

That would be a pretty classic AMD move. Stop investing in handheld. The AMD powered PS5/6 Portable releases and raises interest. Intel or NVIDIA push their solution and take over the PC segment.

I've used Radeon my whole life and I honestly think handheld PCs need Nvidia GPUs. Nvidia is still super popular and Intel/Nvidia APUs might help make handheld PCs mainstream.

I can see AMD abandoning handhelds to Intel so they can save the silicon for more profitable product lines. OEMs love these things though. Take a laptop APU, use cheap tablet parts, and sell them for the price of a laptop.

 

[Shogmaster]

Fuck, it's still a 128bit memory bus. Same problem of memory bandwidth bottleneck

RDNA 5 is more bandwidth efficient than RDNA 3/4. 160GB/s shpuld be good enough for 24 CUs.

 

[SupremeJ]

24 CU RDNA 5. 128bit LPDDR5X. Probably 24 to 32 GB.

+ 64MB GPU cache (more "bandwidth")

It will be fine for a $599 Xbox model

 

[Shogmaster]

+ 64MB GPU cache (more "bandwidth")

It will be fine for a $599 Xbox model

HEY! Don't be stealin' my horseshit predictions!

 

[hououinkyouma00]

I would love to see AMD release straight up bangers but…

I say this every time and every time I'm let down. The only hope is that Nvidia doesn't seem to give a shit about gaming anymore so maybe AMD can capitalize on that.