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

[DisEnchantment]

Speculate at will

 

[igor_kavinski]

Too high of a gen on gen bump.

Hey, well, the current 9800X3D situation kinda confirms your $999 halo price hypothesis had vanilla Zen 5 delivered a solid 40% IPC uplift.

 

[adroc_thurston]

Hey, well, the current 9800X3D situation kinda confirms your $999 halo price hypothesis had vanilla Zen 5 delivered a solid 40% IPC uplift.

Just perf uplift, really.
Turin does and boy the ASPs are kinda stinky.

 

[Joe NYC]

Halo won't be cheap.

Laptops with NVidia 4070 dGPU are not cheap either

 

[LightningZ71]

If AMD really wants the laptop gaming performance crown, then they can release a 9745HX3D any day of the week. They didn't release that for Dragon range, and I suspect that won't do it this time either.

If you look at Strix Point and Kraken point, you realize that neither one of them really make a lot of sense without the MALL cache that they were originally (supposedly) designed with.

 

[igor_kavinski]

If AMD really wants the laptop gaming performance crown, then they can release a 9745HX3D any day of the week. They didn't release that for Dragon range, and I suspect that won't do it this time either.

The OEMs are also to blame. Nothing is stopping them from releasing 18 inch laptops with 9800X3D power limited to 65W TDP. It's the age old problem. The sellers don't want to sell us something WE want. They want to sell us something that benefits them more than it does us.

We live in a world where hordes of people are sometimes paying more for SBCs than mini PCs that would smoke them in a second in terms of performance. Utterly incomprehensible to me why manufacturers are so myopic when it comes to gaming options.

 

[igor_kavinski]

https://videocardz.com/press-release/g-skill-releases-low-latency-ddr5-6000-cl26-cl28-kits-for-ryzen-9000-series

NICE

 

[Win2012R2]

Now imagine Zen 6 x3d with this low latency RAM... yum!

 

[igor_kavinski]

Now imagine Zen 6 x3d with this low latency RAM... yum!

I like that AMD partners are taking the opposite approach of Intel's high speed RAM support by reducing latency. Competition is awesome!

Also, DDR5 seems to be a lot more successful than DDR4. We are at least two years away from general availability of DDR6 and DDR5 is already getting close to CL18 of many DDR4-3600 kits.

 

[Win2012R2]

I reckon we won't see DDR6 for desktops for longer, this is going to be server first RAM, DDR5 got a lot of juice left in it anyway

 

[poke01]

Laptops with NVidia 4070 dGPU are not cheap either

Sure but the gaming market wants GeForce

 

[igor_kavinski]

Sure but the gaming market wants GeForce

They will change their tune once they see Strix Halo GPU accessing more than 32GB RAM on the 64GB laptops.

Nvidia can only make that possible on their ARM laptops.

 

[LightningZ71]

They're already talking about high memory mobile GPUs. They can leverage the 3GB GDDR7 modules to get 24GB or more with higher end GPUs...

 

[Win2012R2]

once they see Strix Halo GPU accessing more than 32GB RAM on the 64GB laptops.

Very very slow snail like RAM, where as Nvidia got cheetah GDDR7

 

[igor_kavinski]

Very very slow snail like RAM, where as Nvidia got cheetah GDDR7

But there's the benefit of unified RAM. Depending on their driver implementation, it could mean instant access for shared data between CPU and GPU and avoid bus transfers.

 

[Win2012R2]

But there's the benefit of unified RAM

It's still very slow RAM, now if it was unified GDDR7...

Plus when talking about unification you probably mean console scenario where games don't have to copy few times same data - however I believe (correct me if I am wrong here) Windoze games using DirectX are just not used to unified scenario, so laptop GPU gets it's "own" memory selected from that unified one...

 

[igor_kavinski]

laptop GPU gets it's "own" memory selected from that unified one...

I think that's true of current implementations of iGPU by AMD (not sure about Intel) but maybe with Strix Halo they could conceivably tap their console experience and build special drivers to take advantage of a large pool of unified RAM.

 

[Win2012R2]

maybe with Strix Halo they could conceivably tap their console experience and build special drivers to take advantage of a large pool of unified RAM.

How can it be done if the software does not support it? Hardware was there a long time ago, but Windoze model isn't there, so (and that's my understanding) you can't just have unified memory and have say CyberPunk 2077 run in it - on Windowz of course, they can do that fine on consoles where API designed for it. It will probably take DirectX 20 to get it all working...

Now for "AI" having large pool of unified memory should be good to load big models, that's for sure.

 

[igor_kavinski]

How can it be done if the software does not support it?

Tricks. If the CPU requests some data that the driver "knows" is present in the GPU's portion of the RAM (maybe through indexing or data tags or whatever), then it could save precious latency cycles on expensive memory transfers. That's my hope at least.

 

[igor_kavinski]

Now for "AI" having large pool of unified memory should be good to load big models, that's for sure.

Yeah. I guess that's what they are going to tout in their marketing materials.

 

[StinkyPinky]

Anyone see that some of these Chinese mini pc companies are putting a Ai HX370 in the form factor and selling them for a thousand dollars. Kinda expensive but that would be an interesting PC.

 

[igor_kavinski]

Kinda expensive but that would be an interesting PC.

Very.

Doesn't throttle at full load: https://www.servethehome.com/beelink-ser9-amd-ryzen-ai-9-hx-370-mini-pc-review/

 

[Win2012R2]

If the CPU requests some data that the driver "knows" is present in the GPU's portion of the RAM

Console Peasant: pathetic weak CPU (or maybe even GPU) requests compressed data from disk and it gets decompressed by additional hardware and put where it needs to be - done.

PC Master Race: strong CPU reads data from disk into main memory, then decompresses it into main memory, then copies it over slow PCIE interface to GPU memory - even if memory is unified and no slow PCIE needed, then still memory is logically partitioned to RAM and "VRAM" - copying will still happen, plus likely higher usage as some stuff potentially will be duplicated in VRAM and RAM.

It's something like that.

Microsoft needs to pull a finger and introduce DirectX 13 which will be whatever they've got for Xbox, only in Windows 12...

 

[igor_kavinski]

https://videocardz.com/newz/amd-set-to-launch-ryzen-9-9955hx3d-flagship-mobile-cpu-with-3d-v-cache

Can they manage to divert enough supply to laptops to turn the tide?

 

[LightningZ71]

They can supply to wherever they can make a profit. The issue is more that the laptop motherboards for these products are rather unique and, given the price range that they sell into, low volume. Since this generation is still on AM5, we'll likely see the few (only?) vendor(s) that made Zen4 series laptops reuse the same board and cases for these products.

 

[igor_kavinski]

Since this generation is still on AM5, we'll likely see the few (only?) vendor(s) that made Zen4 series laptops reuse the same board and cases for these products.

They might save a few bucks on the cooling solutions since Zen 5 generally runs cooler. Laptops could be lighter too.