AMD Ryzen 8000/9000 APU roadmap leaks out. Strix Halo delayed to 2025

[filpo]

Summary

Moore’s Law is Dead provides an update to AMD Ryzen APU deployment. More rumoured specs are seen

 

Quotes

Quote

The roadmap shows different performance segments that AMD aims to address with their upcoming Zen4/5 architectures. In 2024, the Dragon Range architecture, designed for high-performance gaming systems, will be the predominant AMD choice in the market. The roadmap doesn’t indicate any updates for the following year, but in 2025, the Fire Range series will introduce the latest Zen5 architecture.

 

More interesting are the Strix series, namely the Strix Point and Strix Halo (also known as Sarlak) APUs. The former is a proper successor to Phoenix, which was expected sometime early next year, as AMD has been releasing new series typically around that period of the year. While it is a fact that at CES AMD usually made the announcements, the actual launch typically shifted to the following months, and that’s what this roadmap is possibly showing.

 

Quote

According to this roadmap, Strix Point featuring Zen5 CPU cores and RDNA 3.5 GPU cores combined with XDNA 2 AI-cores will sometimes launch in mid-2024. The roadmap confirms that the series will feature 12 Zen5 cores and what is new is the 45 to 50 TOPS AI performance. In fact, all the upcoming premium and high-end APUs are expected to feature this level of AI performance.

 

Quote

More interestingly, the Strix Halo, which is said to offer 16 Zen5 cores and up to 40 RDNA 3.5 Compute Units, would launch in 2025, not 2024 as previously expected. This means that the new APU segment defined by AMD as ‘Elite Experiences’ will not launch next year. This chip is also said to feature XDNA2 AI chip with 45 to 50 TOPS of performance.

 

The reason for the increased AI performance is most likely a new requirement for the Windows 12 operating system, which is said to be around 50 TOPS, as mentioned by other leakers.

AMD Zen4 “Hawk Point”, Zen4 “Escher” and Zen5 “Kraken Point”

Quote

The new roadmap also confirms what will happen to Hawk Point, which is now confirmed to feature 8 Zen4 cores. This is most likely the refresh of Phoenix silicon which features RDNA3 and XDNA1 GPU/AI-cores. As the roadmaps shows it offers 16 TOPS of AI acceleration performance. Its successor, called Kraken Point is to also feature 8 cores, but of the newer Zen5 architecture and upgraded RDNA3.5 graphics. The AI-core will be upgraded to XDNA2, but the node will remain at 4nm.

 

Quote

We also finally have some data on Escher, which is an 8 Zen4 chip with RDNA3 and XDNA1 architecture. It would replace Rembrandt-R, which is surprising not going for a retirement just yet. Instead, the Rembrandt chip will fall into the low-power mainstream category within the 2025 lineup.

 

Therefore, one cannot be sure at this point if Strix Halo will remain as Ryzen 8000 or rather Ryzen 9000 series at this point. AMD’s own product naming schema suggest it should be the latter.

 

My thoughts

Not great to see that Strix Halo is coming in 2025 but it'll most likely be worth it. I like to see that we can possibly get Ryzen 8000 desktop chips late next year and hopefully X3D chips early 2025

 

Sources

AMD Ryzen 8000/9000 APU roadmap leaks out, "Strix Point" to launch mid-2024, "Fire Range" and "Strix Halo" in 2025 - VideoCardz.com

[ImWilly]

Thanks for sharing

50 minutes ago, filpo said:

Ryzen 8000 desktop chips late next year and hopefully X3D chips early 2025

At least, it seems that current Ryzen 7000 series is competitive against Intel 14th gen (still waiting for actual benchmark comparison video though, I can only speculate not huge difference compared to 13th gen because early impression is that 14th gen is just a minor adjustment from 13th gen), and current user of Ryzen 7000 series will be relevant until 2025

[PDifolco]

25 minutes ago, ImWilly said:

Thanks for sharing

At least, it seems that current Ryzen 7000 series is competitive against Intel 14th gen (still waiting for actual benchmark comparison video though, I can only speculate not huge difference compared to 13th gen because early impression is that 14th gen is just a minor adjustment from 13th gen), and current user of Ryzen 7000 series will be relevant until 2025

Yeah, I'm a bit worried about this timeline cause it means that Ryzen 9000 *may not* be AM5 as they'll be 2026+ (and AMD announced AM5 support up to 2025 only)

If AM5 is only a 2 generation thing (vs 4 on AM4) it'll end up a bit scammy and close to Intel anticonsumer strategy   

 

 

[Doobeedoo]

Interesting, the 40CU APU looks awesome, finally APU that's console like where GPU is more beefed up. Really I'd like to see them make 8c X3D model and use extra space for more CUs even. That would be amazing. 

[CephDigital]

33 minutes ago, Doobeedoo said:

Interesting, the 40CU APU looks awesome, finally APU that's console like where GPU is more beefed up. Really I'd like to see them make 8c X3D model and use extra space for more CUs even. That would be amazing. 

That 40 CU APU is going to be a spicy chip, especially for super SFF builds. I wonder what's the max CU they could put in an APU... 

[BabaGanuche]

1 hour ago, ImWilly said:

At least, it seems that current Ryzen 7000 series is competitive against Intel 14th gen (still waiting for actual benchmark comparison video though

These have been out for a while and that answer is 13 gen = 14 gen within a 1-2% for most cases.

 

 

 

 

[porina]

43 minutes ago, Doobeedoo said:

Interesting, the 40CU APU looks awesome, finally APU that's console like where GPU is more beefed up. 

I'd caution that what we know as a socketed APU could struggle with DDR5 alone due to low bandwidth. Soldered implementations using LPDDR could help some more, but I have to wonder what else AMD might be doing that hasn't been disclosed yet. Some or all of: local VRAM, Infinity Cache, or something new?

[DANK_AS_gay]

54 minutes ago, Doobeedoo said:

Interesting, the 40CU APU looks awesome, finally APU that's console like where GPU is more beefed up. Really I'd like to see them make 8c X3D model and use extra space for more CUs even. That would be amazing. 

120 watts though? That doesn't seem like enough for that APU. The 7600 uses 32 CUs and draws 165 watts on its own. Add in a 65-85w CPU, and 120w looks anemic. 

[Doobeedoo]

38 minutes ago, CephDigital said:

That 40 CU APU is going to be a spicy chip, especially for super SFF builds. I wonder what's the max CU they could put in an APU... 

Could depend, chiplets packaging wise with both CPU and GPU of larger size. So maybe if it would be 8c + 40CU or a bit more, could be possible on AM5.

But, I'm sure on TR socket they could literally put 16c CPU + 96CU GPU basically combining consumer flagships.

24 minutes ago, porina said:

I'd caution that what we know as a socketed APU could struggle with DDR5 alone due to low bandwidth. Soldered implementations using LPDDR could help some more, but I have to wonder what else AMD might be doing that hasn't been disclosed yet. Some or all of: local VRAM, Infinity Cache, or something new?

Yes, I'd like to see onboard memory maybe HBM type, definitely something to not make such chip be bottlenecked by memory.

20 minutes ago, DANK_AS_gay said:

120 watts though? That doesn't seem like enough for that APU. The 7600 uses 32 CUs and draws 165 watts on its own. Add in a 65-85w CPU, and 120w looks anemic. 

Well 170-230W could be possible, but yeah around 300W it would be more in line like separate parts would be. And we've seen chips in this range.

[Fasterthannothing]

3 hours ago, PDifolco said:

Yeah, I'm a bit worried about this timeline cause it means that Ryzen 9000 *may not* be AM5 as they'll be 2026+ (and AMD announced AM5 support up to 2025 only)

If AM5 is only a 2 generation thing (vs 4 on AM4) it'll end up a bit scammy and close to Intel anticonsumer strategy   

 

 

Honestly I would bet that's part of the strategy. "See we gave you 2025 support" 

[PDifolco]

1 minute ago, Fasterthannothing said:

Honestly I would bet that's part of the strategy. "See we gave you 2025 support" 

Not sure really, everyone is always late and has unexpected delays in all lines of work...

Yet once in '26 they may be tempted to upgrade platform (depends on where Intel and maybe ARC will be at that future point of time) while not breaking any promises

[ImWilly]

3 hours ago, PDifolco said:

Yeah, I'm a bit worried about this timeline cause it means that Ryzen 9000 *may not* be AM5 as they'll be 2026+ (and AMD announced AM5 support up to 2025 only)

Shoot! That would be a bad curve ball from AMD if they actually create AM5 only for Ryzen 7000 series

I actually hope that AMD would support at least 3 Ryzen generation (not 3 years in parameter), where in the future I would upgrade my CPU from 7800X3D to 9800X3D (or whatever AMD marketing naming later, perhaps more "X" branding because AMD marketing seems to be a legit 90's edgy kid)

 

3 hours ago, PDifolco said:

If AM5 is only a 2 generation thing (vs 4 on AM4) it'll end up a bit scammy and close to Intel anticonsumer strategy   

I do understand that AMD need to please/support motherboard manufacturer by making shorter generational socket support (in this case, AM5) to drive more motherboard sales (just look at Intel before AMD Ryzen was cool, almost all manufacturer give the best features on Intel motherboard while AMD got whatever left because Intel is the superior market leader even though the CPU only support 2 generations, except current 12th-14th gen)

 

But still, if AMD can find a good sweet spot to please us as consumer and motherboard manufacturer, that would be great

[starsmine]

I want to point out that AMD has not committed at all to zen 6 on AM5 one way or the other. I also wouldnt say AM4 got 4 gens, zen+ wasnt a new gen. 

They have only committed to zen5. , and they said support 2025+. Zen6 can go either way, I dont know when they need to commit. 

Also people talking about changing sockets here like you do it to be scummy or anti-consumer. It is infinitely easier on the design side to not lock yourself into the constraints of a socket and co-develop a new socket with new pinouts and new supplemental features should you want them like more PCIe(though since you can daisy chain AM5 chipsets, that specifically isnt necessary).

Im wondering where the 15W options are. Like why does videocardz not show a 15W version of hawk point existing. 

Also gahd damn mendacino being carreid forward into 2025 is a long time. 

I do wonder about the future of TOPs. How much will copilot really need, and how often will I be using it. 

[NumLock21]

That Videocard Z Rumor chart is all over the place.

• Only Granite Ridge is label as Desktop CPU, and the rest isn't. So others are mobile parts or they didn't specify. APUs can also be desktop CPU, but with integrated graphics.
• Hawk Point starts with 12CU RDNA3, follow by Strix Point with an upgrade to 16CU RDNA3.5, both available in 2024.
• Fire Range comes out in 2025 and it downgraded to just 2CU RDNA2?
• Strix Halo also coming out in 2025, has 40CU RDNA3.5. From 2CU RDNA2 in Fire Range to 40CU RDNA3.5 in Strix Halo. All hail this as "Ultimate APU"!!!
• Fire Range model series Ryzen 9055, where as Strix Halo being Ultimate gets Ryzen 9050, shouldn't the 5 in 9055, represent a higher end product, like "the Ultimate".

[starsmine]

16 minutes ago, NumLock21 said:

That Videocard Z Rumor chart is all over the place.

• Only Granite Ridge is label as Desktop CPU, and the rest isn't. So others are mobile parts or they didn't specify. APUs can also be desktop CPU, but with integrated graphics.
• Hawk Point starts with 12CU RDNA3, follow by Strix Point with an upgrade to 16CU RDNA3.5, both available in 2024.
• Fire Range comes out in 2025 and it downgraded to just 2CU RDNA2?
• Strix Halo also coming out in 2025, has 40CU RDNA3.5. From 2CU RDNA2 in Fire Range to 40CU RDNA3.5 in Strix Halo. All hail this as "Ultimate APU"!!!
• Fire Range model series Ryzen 9055, where as Strix Halo being Ultimate gets Ryzen 9050, shouldn't the 5 in 9055, represent a higher end product, like "the Ultimate".

Yes, all but granite ridge are mobile chips. Whether some make it to the desktop or not as APUs for am5 we dont know. 
Im not sure what your point is for hawks and striks. Those are not all over the place Hawk Point is a revised Phoenix. Strix point will come out with zen 5 in H2. 
 

Fire range is an APU the same way dragon range is and zen 4 on desktop is. Its not. Its focus is being a desktop CPU in a laptop. It will be faster than Strix Halo purely due to power budget letting it clock higher. 

Strix Halo I agree might not have a great name. Dragon range needs a dGPU to match strix halo for gaming. But purely for CPU compute it will not be top of the line. 

 

[Mark Kaine]

So 50 TOPS? ...  is it gonna be faster than my 3070 or nah?

[Mark Kaine]

7 hours ago, ImWilly said:

That would be a bad curve ball from AMD if they actually create AM5 only for Ryzen 7000 series

Nah,  its fine, the sooner the better, its really underwhelming overall...

 

hope they go away from "ridged edges" too, looks fragile and ugly af, might force me to go intel eventually lol... 

[starsmine]

13 minutes ago, Mark Kaine said:

So 50 TOPS? ...  is it gonna be faster than my 3070 or nah?

for I(integer)OPS, not for flops. 

[Sharkyx1]

4 hours ago, starsmine said:

for I(integer)OPS, not for flops. 

TOPS is Tera Operations, not IOPS, its an ai performance operation measurement

[hishnash]

14 hours ago, Doobeedoo said:

Yes, I'd like to see onboard memory maybe HBM type, definitely something to not make such chip be bottlenecked by memory.

Have modern HBM solutions improved latency, last time I looked at this for an SOC (with cpu latency sensitive workloads) HBM was sub-optimal. 

[starsmine]

34 minutes ago, Sharkyx1 said:

TOPS is Tera Operations, not IOPS, its an ai performance operation measurement

Yes, Tera, as in 1 trillion operations. IOPS already is in use for in out.
Its not floating point. 

Gaming GPUs do not have an "efficient" path for low precisions integer ops. hell, Ampere cant even do 32bit ints faster then 32bit flops, because half the cores are FP exclusive with the other half being able to do FLOP or Integer opertation. 

Hence why a small AI chip is able to do Int8 at 40TOPS when a 3070 would only be doing 10TOPS on the cuda cores. 
 

[Arika]

christ the red and yellow """"watermarks"""" don't get any less obnoxious do they......

[williamcll]

Having previous architecture on a next generation model number is incredibly misleading. At least call it Zen4.5 or something

[starsmine]

4 hours ago, williamcll said:

Having previous architecture on a next generation model number is incredibly misleading. At least call it Zen4.5 or something

Which chip are you referring to?
 

[filpo]

Just now, starsmine said:

Which chip are you referring to?
 

probably Hawk Point. You can see it says '8x Zen4 cores'