AMD Strix Halo APU Leak Reports the Performance of a laptop 4070, which is equivalent to the RX 6700 XT

[Arika]

press x to fucking doubt.

 

On 4/26/2023 at 2:20 AM, DrMacintosh said:

If its not ARM its not a competitor to Apple Silicon.

i've said it before, but there will never be a competitor to AS, because it's entirely segmented off from everything. it's tied to Apple's ecosystem, you can't get AS without iOS/MacOS and you cant get macOS (going forward) without it being on an AS chip. if you could buy an AS Soc in a non-apple non-macOS device, then yes, there could be competition, but they are literally in their own market segment now.

[Zodiark1593]

22 hours ago, WallacEngineering said:

I seriously doubt it. That would be like 7-8 generations worth of improvement in a single generation when we are getting close to the edge of not being able to shrink transistors any smaller without significant engineering issues or changes. Theres just no way. I said the same thing on the YT video that features the news. Its BS

Consoles are already doing this same thing. The tech is easily there, just not able to be well utilized on PC with relatively slow DDR4. 
 

With DDR5, or better, a triple or quad channel setup, a product such as this would not be a major accomplishment. 

[WallacEngineering]

4 hours ago, Zodiark1593 said:

Consoles are already doing this same thing. The tech is easily there, just not able to be well utilized on PC with relatively slow DDR4. 
 

With DDR5, or better, a triple or quad channel setup, a product such as this would not be a major accomplishment. 

 

Well see the issue is that Console Chips are custom, they wouldn't fit into a standard motherboard socket.

 

From what I hear they are as large as ThreadRipper CPUs. So putting that same amount of power into a CPU half the size while ALSO having a reasonable amount of cores AND also dealing with 5nm+ transistor heat density - its just not possible right now.

 

Thats fine though, cuz if an APU that powerful came out for desktops it would be so expensive it would cover the cost of a dedicated GPU anyways. Theres no way they wouldn't capitalize on those profits.

[Zodiark1593]

2 hours ago, WallacEngineering said:

 

Well see the issue is that Console Chips are custom, they wouldn't fit into a standard motherboard socket.

 

From what I hear they are as large as ThreadRipper CPUs. So putting that same amount of power into a CPU half the size while ALSO having a reasonable amount of cores AND also dealing with 5nm+ transistor heat density - its just not possible right now.

 

Thats fine though, cuz if an APU that powerful came out for desktops it would be so expensive it would cover the cost of a dedicated GPU anyways. Theres no way they wouldn't capitalize on those profits.

I needn’t think they need to go overly large on the GPU to achieve high performance, and can lean a bit more on clocks. RDNA 2 has shown to be able to approach 3 GHz (RX6500 XT), and if RDNA 3 can clock similarly, there would certainly be room to shave down the size. 
 

Undoubtedly, such an APU would be a fairly expensive solution, both for the SoC, and the memory subsystem to properly feed it. Given cost and power consumption, this would be best suited for high performance systems, where space is at a tight premium. High end NUC-type systems, and performance-oriented laptops. LPDDR5x on a 128-bit bus (dual-channel in pc space) would provide ample bandwidth, though you lose modularity. 

[IDProG]

On 4/26/2023 at 1:27 PM, xAcid9 said:

<95w 4070 laptop is equal to RX 5700 XT/RTX 3060 at best. 

Assuming it runs on LPDDR5X@8000MT/s, the bandwidth will be like 256GB/s?

If that the case it probably will perform like RX6600 or slightly better unless it come with significantly improve infinity cache. 

 

 

also MLID as source = weak hearsay.

Someone make a spreadsheet of MLID accuracy for AMD Zen3+4, he's just over 50% accurate. Laughable for a "reputable" "leaker" 
https://docs.google.com/spreadsheets/d/1a7OWkeyHcdFgoQ395zmi7RlWluAKCv9bG_2silBQU5Q/edit#gid=0

Nah, this comment is laughable.

 

There is one big thing that the person who made the spreadsheet missed (or probably chose to purposely ignore, depending on how jaded you are): THINGS CAN CHANGE before they're officially announced, very quickly, and very massively.

 

Remember that the pricing of the RX 7600 was changed LAST MINUTE, as confirmed by many YouTube reviewers like Hardware Unboxed and Gamers Nexus.

 

The person quoted a quote from a video that literally has phrases like "Early Whispers". I think he has stated many times that you should take his early whispers leaks with a grain of salt, because the things leaked are most likely prototypes (experimental) that can change later on.

 

There are also things that can be renamed easily. Like "NVCache" that he claimed back before Ampere came. That turned out to be true, but was called RTX IO instead of NVCache.

[leadeater]

4 hours ago, IDProG said:

THINGS CAN CHANGE before they're officially announced, very quickly, and very massively.

 

4 hours ago, IDProG said:

The person quoted a quote from a video that literally has phrases like "Early Whispers". I think he has stated many times that you should take his early whispers leaks with a grain of salt, because the things leaked are most likely prototypes (experimental) that can change later on.

When it comes to the silicon industry and silicon transistor products not really it doesn't. Any rumors about architectural things are either accurate or they are just straight up wrong and were never correct, not even remotely. If something ever gets to the stage of engineering samples and actually fabricated then it's finalized in nature and will not change drastically, not at the level of information YouTube and Twitter leakers throw out.

 

Advanced computer designs and simulations make it unnecessary to fabricate a chip to "poke around" or "test things out". It's so wildly expensive to make samples per chip cost you only ever do it for a good reason. By the time samples get out of the hands of only XYZ company and to other necessary partners these are closer to qualification samples than engineering samples too which means any leaks based on these are even less likely to deviate from the retail product so if that happens then the leaks were bogus.

 

Sure big changes can happen, like pricing, but not the real actual important stuff that the technology minded people want to know.

[doubleflower]

On 4/25/2023 at 11:51 PM, My_Computer_Is_Trash said:

Summary

AMD's Strix Halo Mobile APU has been leaked and it appears to have the graphics performance of a laptop 4070. (The 4070 Laptop has the performance benchmarks of a Radeon RX 6700 XT) Estimations of its arrival are in the second half of 2024. This looks like AMD attempt to hit three birds with one stone. The birds being Intel, Nvidia, and Apple.

 

Quotes

 

My thoughts

I hope AMD delivers at their normal budget price point. I would like to see this in desktops eventually.

 

Sources

https://overclock3d.net/news/cpu_mainboard/amd_s_next-gen_zen_5_strix_halo_apu_should_worry_both_apple_and_nvidia/1

https://www.laptopmag.com/news/amd-ryzen-8000-series-leak-macbook-tier-efficiency-4070-graphics-incredible-performance

https://www.kitguru.net/components/cpu/joao-silva/amd-zen-5-based-ryzen-8000-mobile-cpu-specs-reportedly-leak/

https://www.notebookcheck.net/Massive-AMD-Ryzen-8000-Zen-5-leak-Strix-Halo-chiplet-APUs-targeting-Apple-silicon-efficiency-with-RTX-4070-class-graphics-along-with-new-Strix-Point-and-Fire-Range-APUs.709446.0.html

 

More

The video that leaked these APUs has much more information and is here:

 

As with any hardware 'leakers', always take those rumors with a large grain of NaCl

 

While the chip looks very impressive and I don't doubt AMD can design and build a chip like this, the main issue for AMD mobile chips is that their chips are so supply starved, to the point that it might be a paper launch, or worse still a vaporware product (see laptops with Zen 4 mobile chips, which was first announced in CES this year and they still haven't reached any form of general availability, while you can buy laptops with 13th gen Intel chips practically anywhere now)

 

Quote

 

I hope AMD delivers at their normal budget price point.

 

 

 

Judging from the specs it probably won't be cheap and they might be even pricier than a normal performance oriented mobile CPU + GPU combo (which will push prices closer to what Apple offers in their hardware, and at that point many will question why would you want to get this over an equivalent Apple hardware)

[IDProG]

On 6/4/2023 at 5:12 PM, leadeater said:

 

When it comes to the silicon industry and silicon transistor products not really it doesn't. Any rumors about architectural things are either accurate or they are just straight up wrong and were never correct, not even remotely. If something ever gets to the stage of engineering samples and actually fabricated then it's finalized in nature and will not change drastically, not at the level of information YouTube and Twitter leakers throw out.

If the leak was spread 1 quarter max, then yes, you are correct.

 

However, anything above 2 quarters is fair game to change. Even node used is fair game, as Nvidia changed Ampere's node from TSMC N7 to Samsung 8LPP in less than a year. Though, to be fair, they probably thought of this possibility more than a year ago and designed Ampere on two different nodes.

 

The person who made the spreadsheet was clearly not being scientific with the analysis (self-proclaimed and also, like, 50% of the leaks are nowhere to be found in the spreadsheet), which is fine for personal analysis, but unacceptable in this case, since he/she spreads his/her analysis to a lot of people, and people aren't going to fact-check his/her analysis and will just shout "That leaker is unreliable. Only 51% accuracy". In fact, it actually happened here.

[leadeater]

3 hours ago, IDProG said:

If the leak was spread 1 quarter max, then yes, you are correct.

 

However, anything above 2 quarters is fair game to change. Even node used is fair game, as Nvidia changed Ampere's node from TSMC N7 to Samsung 8LPP in less than a year. Though, to be fair, they probably thought of this possibility more than a year ago and designed Ampere on two different nodes.

Which company is going to fabricate a product(s) isn't really the same thing as the GPU/CPU architecture design. What I mean by this is the actual architectural changes to the execution engines and front end happen before and without reliance on who or where it's going to be made.

 

So at lot of rumors about that area apply as I stated originally. I also doubt the TSMC or Samsung fabrication decision played out anything like in public reports, time frames included.

 

I think a lot of people have a rose tinted view of how silicon development happens, spontaneous changes aren't really a thing and the more critical it is whether technology or supply chain are much less likely to happen.

[porina]

31 minutes ago, leadeater said:

I think a lot of people have a rose tinted view of how silicon development happens, spontaneous changes aren't really a thing and the more critical it is whether technology or supply chain are much less likely to happen.

I'm sure I'm grossly over-simplifying but I see three main stages:

1, microarchitectural (logical) design - what do you expect the product to do?

2, silicon design - actually implementing the design for a particular silicon process

3, volume manufacture

 

If a company wants to hedge their bets, #2 above is where they can design for multiple processes in parallel, if they have the people available, otherwise sequentially. We have an idea of roughly how long that process can take. Rocket Lake was essentially an implementation of Ice Lake, but on 14nm. Intel have said they started work on it early 2019 when Ice Lake was already done, so it took about 2 years to be on the shelves. Maybe it could be faster or slower depending on the resources thrown at it. This isn't something you do on impulse. It takes long term planning.

 

At the cost of increased engineering expenditure, you can spread manufacturing risk by designing something for 2 or more processes at the same time with parallel teams.

[Franck]

On 4/25/2023 at 1:24 PM, My_Computer_Is_Trash said:

(The 95 Watt 4070 Max-Q Has the performance benchmarks of a 1660ti) 

Where do they get that from ? I am at my second video on youtube and it seems toe to toe with a 3060 TI