Discussion Zen 5 Speculation (EPYC Turin and Strix Point/Granite Ridge - Ryzen 9000)

[DisEnchantment]

Speculate at will

 

[Thibsie]

Strix Halo GPU performance leak:

https://videocardz.com/newz/amd-ryzen-ai-max-pro-395-strix-halo-apu-with-raden-8060s-graphics-leaks-out

Thanks. How good/bad is this ?

 

[Thibsie]

Apple just puts their logo on their badge, next too the SKU name (that's very concise) and that's it. AMD should also just keep the Arrow logo, "R" (for Ryzen) and a single 3-4 digit product number. Ryzen + the product name is the only thing people will remember or verbally express anyway.

Imagine a PC shop talking about these over a phone. What do you think would say? 99% it would be either a "I have a shipment of Ryzen 395 incoming" or even "Strix Halo". Nobody is gonna pronounce the "AI MAX+++++ PRO" BS!

Ridiculous indeed though Apple has its own problem : e.g. M1 with how many cores ? Stupid IMO.
Less stupid than in the PC world but still stupid.

 

[okoroezenwa]

Ridiculous indeed though Apple has its own problem : e.g. M1 with how many cores ? Stupid IMO.
Less stupid than in the PC world but still stupid.

“M1 with X CPU cores and Y GPU cores”. As far as I’m aware that’s what Apple does right now: just spell it out.

 

[Gideon]

Ridiculous indeed though Apple has its own problem : e.g. M1 with how many cores ? Stupid IMO.
Less stupid than in the PC world but still stupid.

“M1 with X CPU cores and Y GPU cores”. As far as I’m aware that’s what Apple does right now: just spell it out.

I agree. And I would personally prefer 3-4 digit product numbers. Worked great on cars since time immemorial (until marketing butchered them there too) as well as old CPUs / GPUs. That's why I mentioned "Ryzen 395". That's not the part i have a problem with.

 

[FlameTail]

“M1 with X CPU cores and Y GPU cores”. As far as I’m aware that’s what Apple does right now: just spell it out.

It works fine for Apple because they don't bin by frequency, and even when they do core binning, there is usually only 2 SKUs (binned and unbinned).

 

[Gideon]

Thanks. How good/bad is this ?

Not too great. Its Vulkan scores are below a 6600M:

Radeon RX 6600M Benchmarks - Geekbench

It's either bandwidth limited or quite unoptimized considering 6600m has about the same bandwidth but considerably less RDNA2 CUs (also clocked lower AFAIK)

 

[eek2121]

It’s a mobile part, what do you guys expect?

Also, pro variant.

It would make for an amazing console SoC, however.

Oh and I wouldn’t judge it based on a single result alone.

 

[poke01]

I just want idle power usage to be low, Stand by will anyway suck cause of Windows. Especially for the lower tier Halo SKUs which will go into tablets and 14” laptops.

 

[igor_kavinski]

It’s a mobile part, what do you guys expect?

Also, it's AMD's first attempt. Assuming they don't give up on this sort of SoC in the future, every subsequent iteration of it will be better and better. AMD is moving things forward in stark contrast to Intel who only plan and then cancel (320EU iGPU anyone?).

 

[fastandfurious6]

Strix Halo GPU performance leak:

https://videocardz.com/newz/amd-ryzen-ai-max-pro-395-strix-halo-apu-with-raden-8060s-graphics-leaks-out

My guess equivalent to mobile RTX 4060. thermal management will play huge role.]

it's very impressive for iGPU actually. medusa Halo may reach even RTX 4080 levels


I wonder if they sell laptops with Halo + dgpu, where the igpu can handle secondary computations in games? but both have to be AMD cos drivers lol

 

[fastandfurious6]

I just want idle power usage to be low, Stand by will anyway suck cause of Windows. Especially for the lower tier Halo SKUs which will go into tablets and 14” laptops.

thermals/pwm probably similar as strix point

my guess on midtier Halo 385 (8 core 32 CU) vs HX 370:

ST: Halo 385 >
MT: probably similar
iGPU: Halo 385 almost x2

there's probably no point in Strix Point anymore lol

and what will be the point in Fire Range at all other than X3D??

 

[Josh128]

there's probably no point in Strix Point anymore lol

and what will be the point in Fire Range at all other than X3D??

Price, price, price. Production cost of Halo should be way more than Strix Point due to advanced packaging, 4nm compute + 3nm large GPIO vs 4nm monolithic.

 

[SteinFG]

Also, it's AMD's first attempt. Assuming they don't give up on this sort of SoC in the future, every subsequent iteration of it will be better and better. AMD is moving things forward in stark contrast to Intel who only plan and then cancel (320EU iGPU anyone?).

That would be 20 Xe cores, right? bonkers, considering that B580 has 20 Xe cores.

 

[Josh128]

Kind of sane. Though I wonder why they didn't go higher. Probably to reduce power consumption.

Thermals 100%. Even Strix Point is limited to 5.1GHz and often doesnt even reach that due to thermal throttling. This thing is larger and not monolithic, thermals will be even harder to constrain.

 

[SteinFG]

there's probably no point in Strix Point anymore lol

and what will be the point in Fire Range at all other than X3D??

Fire range + RTX 5090/5080/5070 would still have a niche.

Strix point is a lot cheaper than Strix halo. Strix halo laptops are probably 1.5-2x the price of Strix point

 

[FlameTail]

Big SoCs like Strix Halo are the future.

https://www.reddit.com/r/hardware/comments/1h8ymj4/the_rise_of_big_socs_with_large_igpus_and_how_it

 

[GTracing]

Big SoCs like Strix Halo are the future.

https://www.reddit.com/r/hardware/comments/1h8ymj4/the_rise_of_big_socs_with_large_igpus_and_how_it

I agree with the post that we'll see more and more big SOCs with larger memory buses and larger iGPUs, but I don't think that they're for gaming (at least not strix halo). They make the most sense for premium mobile workstations.

The RX 7600m and RTX 4060 mobile each have 256GB/s of memory bandwidth. A 256bit LPDDR5X bus (like Strix Halo or the M4 pro) has about 250GB/s. But in gaming the iGPU will have less bandwidth to work with since it's sharing with the CPU.

I'm not sure if LPDDR will ever have enough bandwidth for big iGPUs to be cost competitive with mid range dedicated GPUs in gaming.

 

[igor_kavinski]

But in gaming the iGPU will have less bandwidth to work with since it's sharing with the CPU.

Despite less bandwidth it has one huge advantage, elimination of PCIe bus transfers from system RAM to VRAM. It's just a single large pool of high speed system RAM so results in a huge reduction of latency over discrete solutions like how great the consoles perform despite having older GPU tech. Based on performance per square mm, consoles are king due to unified memory pool.

 

[fastandfurious6]

Strix halo laptops are probably 1.5-2x the price of Strix point

dang 3k for a halo laptop with 4060 perf? nobody will buy that

STX will be price reduced and Halo will sit around $2k laptops, has to...

Yes Halo is way more expensive to produce but in the market there will be a lot of overlap

 

[Josh128]

dang 3k for a halo laptop with 4060 perf? nobody will buy that

STX will be price reduced and Halo will sit around $2k laptops, has to...

Yes Halo is way more expensive to produce but in the market there will be a lot of overlap

Its purpose is more for enterprise/engineering AI and GPU workloads then gaming. Those same people will pay 3k+ for it.

 

[fastandfurious6]

so Halo can exploit 128gb ram for LLMs while even 4090 is limited to Vram? that's the whole selling point?

 

[Josh128]

so Halo can exploit 128gb ram for LLMs while even 4090 is limited to Vram? that's the whole selling point?

Appears to be one of the main ones, at least.

 

[Glo.]

Expect that in the long run, BGA based stuff like this will also be developed for desktop.

Mobo's with soldered in SOCs, but with possibility of adding your own RAM, SSD, and holes for your own cooling systems.

 

[adroc_thurston]

BGA based stuff like this will also be developed for desktop.

No.

Mobo's with soldered in SOCs, but with possibility of adding your own RAM, SSD, and holes for your own cooling systems.

No.
Just not gonna float in commercial et al.

 

[marees]

so Halo can exploit 128gb ram for LLMs while even 4090 is limited to Vram? that's the whole selling point?

Not just LLMs. Microsoft games such as Indiana Jones or MSFS can hog full vram while scaling up, I think