- Mar 3, 2017
- 1,626
- 5,910
- 136
People have to justify their purchases.
OK, so AMD is only making AM5 CPUs with more than 8C for marketing purpose. Right.Marketing.
They're niche but winning the halo is nice.
YES!
Always were, are, and will be!
The M2 Max CPU cost is tiny on MacBook? Certainly not. And they use M2 Max on both cheap and expensive MacBooks. b.L makes sense on all of them.The chip cost there is tiny relative to machine total BOM.
You're contradicting yourself. You said consumers can pay for expensive CPUs on desktop so they do not have to use b.L CPUs, yet there are a huge amount of cheap/povery AMD desktops on the market. Why is AMD not using b.L for those cheap AMD desktops? Does not compute.You can pay for that in desktop.
Whatever floats their boat.You mean they'd rather have an Intel CPU, but buy an AMD CPU anyway, and then complain about Intel having higher power consumption to justify purchasing AMD anyway
Absolutely.OK, so AMD is only making AM5 CPUs with more than 8C for marketing purpose. Right.
People like having the biggest stick even if they don't need it.Then why is the top selling AM5 CPUs on Newegg occupied by 12-16C CPUs on position 4-7?
Max is a speciality part and it's ASP is still tiny if it was a merchant laptop chip.The M2 Max CPU cost is tiny on MacBook?
Desktop ASP is $250 for AMD.You said consumers can pay for expensive CPUs on desktop so they do not have to use b.L CPUs, yet there are a huge amount of cheap/povery AMD desktops on the market
Desktop ASP is $250 for AMD.
Oh no, retail has beyond razor thin margins.Remember retailers/distributors take a decent chunk of the sales price.
With Intel undoubtly but not with AMD.Oh no, retail has beyond razor thin margins.
They both leave retailers razor thin margins. It’s not vendor specific nor is it limited to CPU sales. Retailers have almost no margin on GPUs as well.With Intel undoubtly but not with AMD.
GPUs are not sold by AMD but by OEMs, for CPUs margin wise AMD provide better numbers for retailers, intel is known for mînimal margins left, ask those who work in this area.They both leave retailers razor thin margins. It’s not vendor specific nor is it limited to CPU sales. Retailers have almost no margin on GPUs as well.
For what? I don't think you play games. Stable diffusion?I remember you warning me against buying RDNA3 because I was set on getting it.
What's stable diffusion? I wanted to get back into gaming because I hadn't played much since the early 2000's. There's so many fun looking gamines out there now a days.For what? I don't think you play games. Stable diffusion?
That sounds good, but I want to see some tests first.
Based on the supposed memory subsystem 256-bit 8500mbps LPDDR5x + 32MB cache, it is comparable to what N33 has and memory is shared with the CPU.
Then N3.
For 4060Ti level of performance, It would need ~2.5GHz.
For 7700XT level of performance, It would need ~3GHz.
I am not taking into account the memory subsystem.
For an all-rounder like Strix Halo, I am surprised they didn't save some space by using 8 Zen5 and 8 Zen5c cores, but with >100W TDP maybe It wasn't worth It.
It's this cool thing where you switch off a flag and the GPU will create NSFW art for you, just by typing whatever you wantWhat's stable diffusion?
GPU and AIE perf targets kinda define the resulting shorelone so don't worry.But if that par is also doing I/O, that would mean lot of that non-scalable I/O, SRAM, analog on N3...
Oh no, retail has beyond razor thin margins.
We're talking CPUs and both vendors have them thin on DIY CPUs.then no retailer would stock AMD GPUs
Just go on and spit out the truth. Do the retailers make $100 on selling 1000 CPUs?We're talking CPUs and both vendors have them thin on DIY CPUs.
Seems like it should be named something else if it uses bumps.SoIC is now both bump- and bumpless-3D.
In any case, MI300 is SoIC-X 9um for everything.
Stix Halo looks more like just using desktop chiplets in mobile, except with big gpu + infinity cache in IO die. It doesn't seem like it brings much in the way of system design changes except for the on package memory. I was thinking they would use infinity fabric fan out (as used for the GCD <-> MCD connection in gpus; not sure what they are calling it) rather than stacking for mainstream parts. I guess that isn't coming with Zen 5.Yea.
No.
Yes.
No, Zen5 brings quite literally 0 system design changes for parts not named Strix-Halo.
Man you're kinda-sorta-almost getting it.
Wrong focus though.
Well, 3D stacking is decidedly not mainstream, yes.
Well it's TSMC's name for trve 3D stacking so it's 3D whether it's bumped or bumpless.Seems like it should be named something else if it uses bumps.
No.Stix Halo looks more like just using desktop chiplets in mobile
MALL is the system design change!It doesn't seem like it brings much in the way of system design changes except for the on package memory.
YES.I was thinking they would use infinity fabric fan out
APUs, no.perhaps with APUs that could be connected together for high end parts
GMI is on its way out, forget about it.That would require higher speed interconnect thant GMI though. I also thought about an APU with a GMI link to add extra cpu cores.
USR over 2.5D is cheapo-cheapo and a 32MB MALL can cover for quite a lot of mem requests.The rumored device with all cores on external chiplets doesn't seem like it can get down to really low power operation.
Interesting. So for 32MB they bumped the speed or added more channels.Yea but you get more SLC bandwidth.
True. I keep forgeting Strix Halo won't be monolith, that is Strix Point.Ahhh taping out another CCD with only 8 Zen5c is a major effort since it's a totally new floorplan and not a low-effort IFOP replacement.
Remember that MALLs are per subchannel.So for 32MB they bumped the speed or added more channels.
Two CPU tiles, one GPU tile.So Strix Halo should consist of 2 CCDs, 1 IOD, maybe 1-2 GCD?,
Ah, I wish known that before buying, could have saved me some money.Whatever floats their boat.
Absolutely.
People like having the biggest stick even if they don't need it.
APUs have more rigid schedules so RDNA4 wasn't quite an option when STX-halo started.Quite odd if Strix Halo, the precursor of what is to come with Zen 6 lineup with all the modular chiplet tech, new packaging concepts etc. to not go with an RDNA4 chiplet if it launches late next year?
Good for you.Nah, I actually need more cores
Zen5 TR will be very late, unfortunately.That is exactly the reason I want more than 16 cores for Zen5.
Quite odd if Strix Halo, the precursor of what is to come with Zen 6 lineup with all the modular chiplet tech, new packaging concepts etc. to not go with an RDNA4 chiplet if it launches late next year?
A year delay like the last and current generation?Zen5 TR will be very late, unfortunately.
This is not really surprising. It already happened In the past with APUs. Cezanne used Zen3 + Vega IGP despite RDNA being already released.Quite odd if Strix Halo, the precursor of what is to come with Zen 6 lineup with all the modular chiplet tech, new packaging concepts etc. to not go with an RDNA4 chiplet if it launches late next year?