GPUs & CPUs & Enthusiast hardware: Questions, Discussion and fanboy slap-fights - Nvidia & AMD & Intel - Separe but Equal. Intel rides in the back of the bus.


Strix Point = 4x Zen 5, 8x Zen 5c, 16 CUs RDNA 3+. 128-bit LPDDR5X-7500
Strix Halo (top) = 6-16x Zen 5 cores (1-2 chiplets shared with desktop/server), up to 40 CUs RDNA 3+, 256-bit LPDDR5X-8000, 32 MB Infinity Cache
Strix Halo ("LP") = 6-8x Zen 5 cores (1 chiplet), 20 CUs RDNA 3+, 128-bit LPDDR5X, maybe the cache

So Strix Halo has double the bus width, 6.7% faster memory speed, and the benefit of L3 cache dedicated to the graphics which increases the "effective" bandwidth. Although it isn't monolithic, it might do surprisingly well in low power scenarios from clocking a lot of CUs very low. If AMD makes a ghetto 128-bit version, it could still retain some advantages over Strix Point such as faster memory and the Infinity Cache.

Kraken Point is an 8-core (4+4) "little" version of Strix Point. Probably 8 CUs.

All of these will have an XDNA 2.0 NPU, now rumored up to 50 TOPS for Strix Point, 60 TOPS for Strix Halo. The difference is simply from being capable of clocking higher.

Sonoma Valley may be a platform compatible replacement for Mendocino, with 4x Zen 5c cores and not much else known. It may be made on Samsung 4nm instead of TSMC, which could make it a great option for cheap PCs that doesn't divert wafer supply from premium products. Moving from Zen 2 to Zen 5c would yield an absurd IPC increase, perhaps over 50%, with new features like AVX-512 support.

We don't know how much of an improvement RDNA 3+ will be, but it does introduce some changes that are intended for RDNA 4, and you can read about them here.
 
  • Informative
  • Like
Reactions: Brain Problems, Wooo, The Ghost of Kviv and 2 others
davids877 said:
I have a 7900XTX and it works ok for the big name AI like LLM and StableDiffusion. A bit slower than I'd expect though.
Click to expand...
I'm going to assume this is on Linux? I have used Rocm on Windows and it was not performant. Using latest versions on Linux (dual boot, same hardware) the difference is very, very noticeable.

The Mass Shooter Ron Soye said:

Strix Point = 4x Zen 5, 8x Zen 5c, 16 CUs RDNA 3+. 128-bit LPDDR5X-7500
Strix Halo (top) = 6-16x Zen 5 cores (1-2 chiplets shared with desktop/server), up to 40 CUs RDNA 3+, 256-bit LPDDR5X-8000, 32 MB Infinity Cache
Strix Halo ("LP") = 6-8x Zen 5 cores (1 chiplet), 20 CUs RDNA 3+, 128-bit LPDDR5X, maybe the cache

So Strix Halo has double the bus width, 6.7% faster memory speed, and the benefit of L3 cache dedicated to the graphics which increases the "effective" bandwidth. Although it isn't monolithic, it might do surprisingly well in low power scenarios from clocking a lot of CUs very low. If AMD makes a ghetto 128-bit version, it could still retain some advantages over Strix Point such as faster memory and the Infinity Cache.

Kraken Point is an 8-core (4+4) "little" version of Strix Point. Probably 8 CUs.

All of these will have an XDNA 2.0 NPU, now rumored up to 50 TOPS for Strix Point, 60 TOPS for Strix Halo. The difference is simply from being capable of clocking higher.

Sonoma Valley may be a platform compatible replacement for Mendocino, with 4x Zen 5c cores and not much else known. It may be made on Samsung 4nm instead of TSMC, which could make it a great option for cheap PCs that doesn't divert wafer supply from premium products. Moving from Zen 2 to Zen 5c would yield an absurd IPC increase, perhaps over 50%, with new features like AVX-512 support.

We don't know how much of an improvement RDNA 3+ will be, but it does introduce some changes that are intended for RDNA 4, and you can read about them here.
Click to expand...
I'm quite keen to see the next gen. I'm not expecting wild difference but I am expecting it to be a nice upgrade. Especially for laptops. I'm still on 1st gen Zen (although a nice one) and I've been holding off for sometime from upgrading. A nice little Zen 5 whole system rebuild would suit me well.
 
  • Like
Reactions: The Mass Shooter Ron Soye
Overly Serious said:
Well yes, but person I was replying to said "I have a 7900XTX" so I figured they weren't a data centre.
Click to expand...

Right. I'm just saying that the explanation for why the Windows version of ROCm sucks is AMD puts no effort into it. Hell, they put no effort into making sure the documentation is correct, the build works, or that it's compatible with the last version.

Everybody in the industry likes dealing with AMD and hates dealing with NVIDIA. But only one of those two delivers competently executed software. Culturally, it seems like AMD is also a lot more of a "ignore things that don't bring in big bucks" kind of company, while NVIDIA is more of a "win everywhere we are, regardless of how big or small it is, salt the earth with the ashes of our enemies" kind of company. I know of incidents where Jensen Huang has personally intervened to take business away from AMD when the money involved was absolute chickenshit, for no other reason than he's not okay with anyone else getting a win.
 
  • Like
  • Informative
Reactions: Anti Alien Alarm, indomitable snowman, WelperHelper99 and 2 others
I'm badly in need of a new setup.
I've heard that 'convention season' starts in June, with tons of new releases - and old stock will plunge this summer.

C/D ?
 
  • Like
Reactions: Overly Serious
Orange Rhymer said:
I'm badly in need of a new setup.
I've heard that 'convention season' starts in June, with tons of new releases - and old stock will plunge this summer.

C/D ?
Click to expand...
It depends on what you're looking at.

Pretty much all the major producers of PC shit have responded to the post-pandemic lull in demand by slashing production and only making small quantities of stuff in short bursts (hence very swingy pricing). I don't think there's very much stock sitting out there to liquidate.

My general philosophy on PC shit is that you're never going to get a perfect deal unless you autistically devote your life to deal-hunting. At some point, it's just better to bite the bullet and buy something rather than sitting around waiting for some mythical return to sub-$500 flagship GPUs. Pretty much every PC community site right now is flooded with people having melties because they've tried to wait out the current market and have instead been rewarded with higher baseline prices.

Just buy something and get it over with. Nothing new is going to massively obsolete what you can get now and that extra few months you get to enjoy having a PC are probably going to outweigh the extra couple hundred you might hypothetically overpay if there does happen to be some kind of unprecedented downward shift in component prices.
 
  • Agree
  • Informative
  • Like
Reactions: Anti Alien Alarm, Orange Rhymer, The Ugly One and 1 other person
A friend of mine was talking about this: https://www.amazon.com/WayPonDEV-CM3588-NAS-Kit-Mali-G610/dp/B0CQZN4LSS , anyone have one?
My ghetto NAS is nearly filled and I cannot possibly cram anymore drives into it, I already have M.2 PCIE adapters in it and it's been getting really hot when I'm doing editing for an extended period of time which already makes me nervous. It would probably cost me the same amount of money to build another ghetto NAS out of unused hardware as this board, so I'm thinking about trying it out but I don't want to buy something I'll end up trashing.
 
Blue Miaplacidus said:
A friend of mine was talking about this: https://www.amazon.com/WayPonDEV-CM3588-NAS-Kit-Mali-G610/dp/B0CQZN4LSS , anyone have one?
My ghetto NAS is nearly filled and I cannot possibly cram anymore drives into it, I already have M.2 PCIE adapters in it and it's been getting really hot when I'm doing editing for an extended period of time which already makes me nervous. It would probably cost me the same amount of money to build another ghetto NAS out of unused hardware as this board, so I'm thinking about trying it out but I don't want to buy something I'll end up trashing.
Click to expand...
Seems a little odd. The RK3588 seems to have 4x PCIe3.0 lanes which means each NVMe likely only gets 1 lane, which isn't bad, but is like 1/4 the possible speed of NVMe on PCIe 3.0 Then it only has 2.5G, so there's a massive bottleneck from the possible 32Gbit/second possible even with the limited PCIe lanes on the storage.

And there's always the question of how well supported the Rockchip CPUs really are.

I assume you have heatsinks on your existing NAS NVMe, and good airflow?
 
  • Thunk-Provoking
Reactions: Blue Miaplacidus
davids877 said:
I assume you have heatsinks on your existing NAS NVMe, and good airflow?
Click to expand...
Yea, I even 3D printed custom air channels and installed dedicated 80mm fans. It's just very cramped and two drives in particular seem to get much hotter than the others. I was going to swap the chassis out eventually as a last hope.
 
Blue Miaplacidus said:
Yea, I even 3D printed custom air channels and installed dedicated 80mm fans. It's just very cramped and two drives in particular seem to get much hotter than the others. I was going to swap the chassis out eventually as a last hope.
Click to expand...
NVMe does run hot generally.

You might want to check the temps and see how they compare to others reported on-line if you haven't already.
Linux: nvme-cli is the package name on Debian/Ubuntu, similar on others probably and then
sudo nvme smart-log /dev/nvme0 (replacing with your device, obviously)
Windows: IDK, some GUI probably.

Of course now that I've said that I probably should add NVMe temps to my monitoring system.
 
  • Thunk-Provoking
  • Like
  • Informative
Reactions: Anti Alien Alarm, indomitable snowman and Betonhaus
Betonhaus said:
View attachment 5957952
Click to expand...

Intel's blaming motherboards & overclockers for running out of spec: https://videocardz.com/newz/intel-i...e-to-13-14th-gen-core-i9-cpu-stability-issues. Interesting comment on reddit:

1714768197665.png

1714768482222.png

Some interesting discussion in addition to the usual reddit-tier shrieking, sounds like MSI and ASUS are the major culprits here:
 
Last edited:
  • Informative
  • Like
  • Thunk-Provoking
Reactions: Brain Problems, George Lucas, Vecr and 2 others
They're all somewhat at fault on this one. Especially after AMD got hit by this issue with mobo vendors.

Intel plays the ignorance card because hey, it makes our performance look better in benchmarks.

*Edit* I've actually ran into simps online claiming that if you buy overclocking mobo/CPU then you should be looking at and changing default settings 1st thing anyways.

No. Factory settings should be fairly conservative and stable. Same as ram. You have to enable a profile to get the oc settings.
 
Last edited:
  • Like
  • Agree
Reactions: Brain Problems, ⠠⠠⠅⠑⠋⠋⠁⠇⠎ ⠠⠠⠊⠎ ⠠⠠⠁ ⠠⠠⠋⠁⠛, The Mass Shooter Ron Soye and 2 others
You must log in or register to reply here.
Back