Question Zen 6 Speculation Thread

[MS_AT]

That thing can't reach it's full potential in that handheld thingy without burning your hands.

Aren't burnt hands small price to pay for "unlimited power"?

 

[adroc_thurston]

My point about form factor thermal limitations is proven by the initial reviews of HX 370 not maintaining peak 1T boost due to thermal throttling. Thats a single core @5.1GHz

That's an STX1 issue that is shared with nothing else.
KRK1 doesn't share it.

Bombs will be needed to keep up with Apple in ST

Well no, to win SMEbench x86 just needs to add AMX to every core.

 

[511]

Well no, to win SMEbench x86 just needs to add AMX to every core.

Alongside APX and AVX-512 xDd

 

[adroc_thurston]

Alongside APX and AVX-512 xDd

I don't think applelabs will ever fix their AVX-512 ASM.
AMX it is.

 

[511]

I don't think applelabs will ever fix their AVX-512 ASM.

I have high hopes Intel will do it for them lol

 

[Josh128]

yesterday I saw an 8inch handheld device here in this forum that has higher performance than your 9900x

hope that helps

I suggest you get glasses.

 

[Abwx]

I suggest you get glasses.

Stock 9900X does about 32000 in CB R23, AI 395 does 35000 at 100-115W, look like it require only 80W to match a 9900X.

 

[OneEng2]

Hmm okay. But the point is, could they realistically release them earlier than the rest of the Zen6 consumer product stack if they wanted to, due to N3P being available now?

I personally believe they could. I don't think they will. Traditionally, the new architecture is released in the highest margin market. This helps take the sting out of yield issues and also makes it easier to slow ramp the product (volumes are lower).

The Zen 6 N3P parts (unlike others, I believe these will be plentiful) will be squarely aimed at the volume market.

I think the way the laptop market works, is that first you've got to have good enough performance. Once you have this base line set, the key to market share is giving MARGIN to the OEMs.

I agree.

Bombs will be needed to keep up with Apple in ST

Yea, about that. Seems like lots of hoopla gets made of one or two benchmarks showing off A19 (and M3) ST prowess. This is used to show how great a laptop or desktop chip ARM variants would make.

But they wouldn't.

If A19 could be made to run a PC, it would be a terrible experience for the user.

 

[adroc_thurston]

If A19 could be made to run a PC, it would be a terrible experience for the user.

They're literally stuffing it into the entry-level Macbook and it'll be pretty good.

 

[Doug S]

If A19 could be made to run a PC, it would be a terrible experience for the user.

Why do you believe that? What does M4 have that A19 doesn't, other than a couple more P & E cores and few more GPU cores? At any rate Apple doesn't believe you because there is apparently a new lower end Macbook using A18P or A19P coming out soon.

 

[poke01]

This is used to show how great a laptop or desktop chip ARM variants would make.

But they wouldn't.

idk Qualcomms X2 Extreme looks like a great CPU to me.

 

[511]

idk Qualcomms X2 Extreme looks like a great CPU to me.

They did MoP good luck selling that to OEMs and 287mm2 N3P Silicon is not cheap.

 

[poke01]

They did MoP good luck selling that to OEMs and 287mm2 N3P Silicon is not cheap.

watch it get more laptop design wins than Strix halo

 

[511]

watch it get more laptop design wins than Strix halo

that's setting bar too low would it even surpass previous design wins that's the big question cause which OEM will take the risk especially with MoP which will take margin from OEMs.

 

[basix]

What is the latest state about the desktop IOD? RDNA 3.5 and some N4 node variant or are there some slight chances for RDNA 4/5 and N3P? Probably unlikely, right?

• I expect N4C and RDNA 3.5 for desktop
• Re-use IP from Strix Point & Halo and time to market security (GPU, NPU, CPU-to-IOD links)
• A bump to 4 CU and 3.0 GHz is still a nice step-up from 2 CU RDNA 2 (roughly 3x as fast iGPU)
• 16-Tile NPU (half of Strix) with cranked up clockrates to get the Copilot+ sticker (clocks need to be increased by at least 1.6x). Such an NPU would be roughly 7...8mm2 in size.

And what about the chipset? TSMC N6 or some Samsung node like 8LPP or SF6? AMDs Wafer-Supply-Aggreement with Globalfoundries was extended to 2025 back in 2021, but that ends now.

• I expect one of the Samsung nodes because of lower cost wafers

 

[reaperrr3]

What? Nope.

If anything, the complete opposite. IPC is higher in productivity and scientific/server workloads:

View attachment 131128
vs Client:
View attachment 131129
Hell, even in this "gaming" slide, the 3 highest and disproportionately big jumps vs the rest shown are not from games (depends how you count Dolphin emulator I guess).

At computerbase.de's review of Zen4, it showed worse IPC gains in proper ST loads than in anything remotely MT, as low as 6%, so some of the "IPC" gains seem to actually have been SMT/C2C improvements.
But alright, I'll concede this point, it was worse in my memory than it seems to have been in most cases.

Also not sure I agree with this either:
View attachment 131130
All said "28 workloads" are SPEICrate2017_int_base and SPEICrate2017_fp_base tests. Which are anything but gaming workloads and are rarely "latency sensitive".

Benchmarks in CB review showed ~11% IPC gain in non-gaming MT workloads (a bit more for ST, probably due to full L3 available to that 1 thread), vs. ~22% in games.

It's of course possible that things looked better in server workloads, but in desktop non-gaming applications, Zen3 wasn't anything too special in terms of IPC.

 

[511]

I expect N4C and RDNA 3.5 for desktop

Well N4C is more expensive than N4P when AMD bought it for Zen 5 🤣🤣

 

[Josh128]

What is the latest state about the desktop IOD? RDNA 3.5 and some N4 node variant or are there some slight chances for RDNA 4/5 and N3P? Probably unlikely, right?

• I expect N4C and RDNA 3.5 for desktop
• Re-use IP from Strix Point & Halo and time to market security (GPU, NPU, CPU-to-IOD links)
• A bump to 4 CU and 3.0 GHz is still a nice step-up from 2 CU RDNA 2 (roughly 3x as fast iGPU)
• 16-Tile NPU (half of Strix) with cranked up clockrates to get the Copilot+ sticker (clocks need to be increased by at least 1.6x). Such an NPU would be roughly 7...8mm2 in size.

And what about the chipset? TSMC N6 or some Samsung node like 8LPP or SF6? AMDs Wafer-Supply-Aggreement was extended to 2025 back in 2021, but that ends now.

• I expect one of the Samsung nodes because of lower cost wafers

Nobody knows at this point, but given AMD's IOD history, I'd wager its moving a node behind what the CCDs will be using. N4 or some form of N3. Whichever gives the best perf/$ for AMD.

 

[fastandfurious6]

Aren't burnt hands small price to pay for "unlimited power"?

🤣🤣🤣🤣🤣

but jokes aside, all it needs is a removable lid right above the big Halo chip so it can slot into a docking station with heatsink and fan

😉😉😉😉😉

 

[MrMPFR]

What is the latest state about the desktop IOD? RDNA 3.5 and some N4 node variant or are there some slight chances for RDNA 4/5 and N3P? Probably unlikely, right?

Most of this is pure speculation:

• N3 premium only ~20% over N4 so prob what AMD would use to unify IP across entire product lineup (Zen 6 and RDNA 5 based) family between desktop CPU, mobile and dGPU.
• No NPU given RDNA 5 based 2 WGP/CU @3ghz = ~50 dense INT8 TOPS.
• TSMC's InFO replacing SerDes
• IIRC 8000 MT/s DDR5

Boring stuff

• VCN 6.0: AV2 decode + other
• New Display controller (DCN):
• Some iterative changes to I/O (USB etc...) and other IP blocks
• PCIe prob unchanged vs Zen 4 IOD.

Do we have confirmation for one monolithic IOD? Given the massive price hike from N6 -> N3 or even N4 splitting it up would boost GM. So keep analog on N6 and everything else on N3.

 

[regen1]

They did MoP good luck selling that to OEMs and 287mm2 N3P Silicon is not cheap.

That die-size figure might be a bit off the mark. We will know in time.

that's setting bar too low would it even surpass previous design wins that's the big question cause which OEM will take the risk especially with MoP which will take margin from OEMs.

Wonder if additionally Qualcomm's mobile chips being used by some of the OEMs(eg. Samsung, Lenovo in Tablets and with Motorola, Asus in phones) play as some kind of leverage with some of them in getting more laptop design wins, both with X ELite and X2 Elite.

 

[511]

That die-size figure might be a bit off the mark. We will know in time.

Yeah but there is no doubt it's big and expensive and what's the point of this config ?
Strix Halo has it beat in MT and iGPU performanceIf you want CPU MT ARL-HX/Zen5 HX exists.
The only advantage is ST and Battery Life.

 

[marees]

Most of this is pure speculation:

• N3 premium only ~20% over N4 so prob what AMD would use to unify IP across entire product lineup (Zen 6 and RDNA 5 based) family between desktop CPU, mobile and dGPU.
• No NPU given RDNA 5 based 2 WGP/CU @3ghz = ~50 dense INT8 TOPS.
• TSMC's InFO replacing SerDes
• IIRC 8000 MT/s DDR5

Boring stuff

• VCN 6.0: AV2 decode + other
• New Display controller (DCN):
• Some iterative changes to I/O (USB etc...) and other IP blocks
• PCIe prob unchanged vs Zen 4 IOD.

Do we have confirmation for one monolithic IOD? Given the massive price hike from N6 -> N3 or even N4 splitting it up would boost GM. So keep analog on N6 and everything else on N3.

How many RDNA 5 WGPs do you need for 50 TOPS or 75 TOPS 🤔

 

[regen1]

Yeah but there is no doubt it's big and expensive and what's the point of this config ?
Strix Halo has it beat in MT and iGPU performanceIf you want CPU MT ARL-HX/Zen5 HX exists.
The only advantage is ST and Battery Life.

Yeah should be very expensive.
Last time around their flagship SKU "X Elite X1E-84-100" was kinda unobtainable, came only in one laptop model(Samsung Galaxy Book4 Edge 16) iirc. We should know next year how they fare this round.

​

 

[511]

Yeah should be very expensive.
Last time around their flagship SKU "X Elite X1E-84-100" was kinda unobtainable, came only in one laptop model(Samsung Galaxy Book4 Edge 16) iirc. We should know next year how they fare this round.

IF it didn't had MoP it would have fared better. IFS would have some packing margin even in singly % on LNL on QCOM I doubt even those.