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

[DisEnchantment]

Speculate at will

 

[Bigos]

Why are you comparing 16+8gb model with 96-128gb ones? Especially ever since dram prices skyrocketed?

 

[Thunder 57]

And the fact that Halo wasn't meant for gaming in its current form.

 

[Joe NYC]

They dont have to. They can buy a prebuilt PC at Amazon or Best Buy that will outrun a Halo minipc by a mile in gaming for less money. It makes no sense at Halos price point. If it was only $400? Then sure, yeah.

This will run circles around the cheapest Halo mini PC you can buy, and do it for a lot less money, lol. Be realistic man, not idealistic. You'd literally have to have more money than sense to by a Halo mini PC over the CyberPower for gaming, and you'd still be slower.

View attachment 134451

View attachment 134450

I was thinking a future version of Steam Machine. They already announced the current version, so no need to discuss alternate reality with current Strix Halo.

But one thing from current Strix Halo MiniPCs to consider is that Framework can sell the low end version of Strix Halo for $1,099 currently. Framework model of selling directly to customer (avoiding middlemen mark ups) is closer to Valve business model with Steam Deck and Steam Machine.

 

[Josh128]

Why are you comparing 16+8gb model with 96-128gb ones? Especially ever since dram prices skyrocketed?

lol wut? Because we are talking about gaming machines, and these are prices from today, which are still around the cheapest you could ever get a full CU count Strix Halo machine. And desktop 5060 will curbstomp Strix Halo in gaming, at any resolution, at every quality setting, thats why. Show me any Strix Halo 40 CU machine for $899 and the 5060 machine is still the better perf/dollar. Find me one around that price, by the way, since you seem to be implying that they ever existed for that price.

Bottom line, you are dumb if you pay more for less. Thats reality. Granted, there are a lot of dumb people in this world, but thats beside the point.

 

[Josh128]

ANd the fact that Halo wasn't meant for gaming in its current form.

Precisely why you'd be dumb to buy one primarily for gaming. Exactly what I said.

 

[adroc_thurston]

ANd the fact that Halo wasn't meant for gaming in its current form.

It is for gaming and they're making more gaming-specific SKUs even

 

[jpiniero]

It is for gaming and they're making more gaming-specific SKUs even

Gaming was the original point. AI hype is why it ended up coming to market.

 

[adroc_thurston]

Gaming was the original point. AI hype is why it ended up coming to market.

It was always gonna come to market, ever since 2021.

 

[jpiniero]

It was always gonna come to market, ever since 2021.

Have to have enough customers. AI hype is what got enough interested.

 

[adroc_thurston]

Have to have enough customers

It always had them.

AI hype is what got enough interested.

Had no impact on stxH whatsoever.

 

[jpiniero]

It always had them.

Had no impact on stxH whatsoever.

I have serious doubts that OEMs would have been interested otherwise. Mainly because of the price.

Once you get over $1000, it's basically 90-95% Apple. AI Hype cred is what got OEMs interested, even if they aren't going to sell much.

 

[adroc_thurston]

I have serious doubts that OEMs would have been interested otherwise. Mainly because of the price

Uhhhh

Once you get over $1000, it's basically 90-95% Apple.

Anything with dGFX including the thriving and growing gaming laptop segment is over a kilobuck.
Are you serious?

 

[jpiniero]

Anything with dGFX including the thriving and growing gaming laptop segment is over a kilobuck.
Are you serious?

Gaming laptops are probably a big chunk of the remaining share. But even then you'd have to think it's the cheaper stuff, 5050s, 5060s.

The handhelds are volume not really worth talking about other than novelty.

 

[adroc_thurston]

But even then you'd have to think it's the cheaper stuff, 5050s, 5060s.

The cheaper stuff starts at kilobuck or therein.

 

[jpiniero]

The cheaper stuff starts at kilobuck or therein.

That's right but still decently below what Halo laptops would have to be.

 

[adroc_thurston]

That's right but still decently below what Halo laptops would have to be.

Halo laptops can get much cheaper.
The more volume you ship, the lower cost per unit gets.
That's why MDSh is built using scraps from elsewhere.

 

[soresu]

That's why MDSh is built using scraps from elsewhere.

Come to think of it, designing a custom RDNA3.5 chiplet for Strix Halo rather than just waiting for N44 just doesn't make sense to me.

What's even stopping them from releasing a N44 Strix Halo refresh to tide over till MDS-H?

On the raster gfx side it's probably going to take a loss vs the older model, but on the RT and ML/FSR4 side it should be an easy win.

Or is it down to the battery power optimisations in RDNA3.5?

 

[adroc_thurston]

Come to think of it, designing a custom RDNA3.5 chiplet for Strix Halo rather than just waiting for N44 just doesn't make sense to me.

Wat.

What's even stopping them from releasing a N44 Strix Halo refresh to tide over till MDS-H?

Navi44 is not LPDDR, has woefully insufficient I/O and no d2d shoreline for CCDs.
Wtaf are you on.

 

[soresu]

Navi44 is not LPDDR, has woefully insufficient I/O and no d2d shoreline for CCDs.

That answers my question, it's easy to forget that dGPU's have very different design constraints beyond just battery power optimisations.

Is AMD going to make this STX-H die layout more common to the lower end dGPU dies going forward to foster that "using scraps from elsewhere" plan?

 

[adroc_thurston]

Is AMD going to make this STX-H die layout more common to the lower end dGPU dies going forward to foster that "using scraps from elsewhere" plan?

Bingo!
That's why AT3/4 have LPDDR shorelines.
Remove the CCD and it's a discrete GPU.
Add em back and you get an APU.

 

[luro]

Bingo!
That's why AT3/4 have LPDDR shorelines.
Remove the CCD and it's a discrete GPU.
Add em back and you get an APU.

And how they solve the bandwidth issue for dGPUs when using LPDDR?

 

[adroc_thurston]

And how they solve the bandwidth issue for dGPUs when using LPDDR?

They don't.

 

[soresu]

They don't.

Lol you missed a key opportunity for the best comeback gif ever:

 

[Tuna-Fish]

And how they solve the bandwidth issue for dGPUs when using LPDDR?

Wide bus to fast LPDDR. 384 bits to 14.4 = 614GB/s. That's halfway between 5060 and 5070.

 

[adroc_thurston]

Wide bus to fast LPDDR. 384 bits to 14.4 = 614GB/s. That's halfway between 5060 and 5070.

Getting to 14.4 will take eons.
10.6 is what you'll be getting if lucky.