Bondrewd
Veteran
Not quite, but yeah, it works.They claim ROCm has reached software parity with CUDA for LLMs
Just not on client parts (APUs in particular) which is the ~issue~.
Not quite, but yeah, it works.They claim ROCm has reached software parity with CUDA for LLMs
10-15% would be healthy. Something like 19% for Zen 3 was unusual and was done by chasing a pretty low hanging fruit with the unified L3. I think without doing V-cache as standard(which AMD have made clear they aren't doing, at least not with Zen 5), there's not really any such low hanging fruit again to boost IPC like that in a single gen. Any talk of like 25%+ should just be laughed out of the room(as I did when MLID tried to claim this for Zen 4 early on).https://videocardz.com/newz/amd-x86...ana-zen6-morpheus-microarchitectures-detailed
IPC figure is rather conservative for ZEN5, far form this 25%-30% IPC "leaks" from RGT/AdoredTV
finally 6 ALU´s
new 16 core complex
Latest drivers are crashing on me whenever I start fiddling with the performance controls. It seems specific to the voltage control. Every time I try and undervolt it the screen goes black and I need to reboot.
8GB RX 580, and lots of annoyance.
Oh there is.there's not really any such low hanging fruit again to boost IPC like that in a single gen
They are making an actual LITTLE but it's not due for Zen5.it suggests a potential 3rd design apart from the main and 'c' variants.
I’m not a believer in moats
You have been consistently wrong about most everything regarding AMD for the past few years, but you talk so confidently about things that you still manage to get people buying into your rhetoric anyways. There's genuinely no reason to believe you actually know anythingOh there is.
They are making an actual LITTLE but it's not due for Zen5.
"and neither in AI."
Oooh someone's upset.You have been consistently wrong about most everything regarding AMD for the past few years
They do, just that the real LITTLE shouldn't be an infra throughput core like Cortex-A or Atom.So they don't "believe" in the big and little core solution done by others
This is not constructive conversation. You need to add details. What are the low hanging fruit to improve IPC and why hasn't it been used so far? What's the source on the new 'LITTLE' cores for Zen5? Give people something to actually talk about.Oh there is.
They are making an actual LITTLE but it's not due for Zen5.
I just wish you'd qualify your comments with more appropriate language, unless you actually KNOW something is true, in which case you should be telling us more about it, which you almost never do. You just continually talk as if you're some insider with secret knowledge, yet rarely ever substantiate anything you say, not to mention that you've been wrong many times in things you've talked so confidently about. It's even worse when you're replying to others trying to say they're wrong, without ever really detailing why. You just make it sound like you know we're wrong and we just have to trust you for some reason.Oooh someone's upset.
Yes, RDNA3 missed.
Nothing else did.
They do, just that the real LITTLE shouldn't be an infra throughput core like Cortex-A or Atom.
One of the event’s highlights was the preview of Samsung’s next-generation flagship mobile processor, the Exynos 2400 with Xclipse 940 GPU based on the latest AMD RDNA™ 3 architecture. A live demo showcased the processor’s substantially enhanced ray tracing capability, promising improved realism and immersion in gaming through a range of optical effects including global illumination, reflection and shadow rendering.
Achieving significant advancements in computing performance, the Exynos 2400 processor features a 1.7x increase in CPU performance and a remarkable 14.7x boost in AI performance compared to the previous Exynos 2200 product. Additionally, Samsung introduced a new AI tool designed for upcoming smartphones, demonstrating text-to-image AI generation using its Exynos 2400 reference board.