- Mar 3, 2017
- 1,622
- 5,892
- 136
You're suggesting AMD will price 9950X differently depending on what the company that buys it will do with it?How it works, what ODMs pay, is exactly the same situation as with Intel, Nvidia, any other company on the market.
You have confused two different markets, on which the retail has much higher markup, compared to OEM
Never claimed Strix Halo will be sold directly to end customer. But end product (e.g. laptop) which contains that CPU will always be sold through retail. Unless the ODM will keep it for themselves. 🤣Strix Halo will never be sold via retail directly due to the 256b memory bus.
It will always have to go through ODMs. That's still treated as "sold through ODMs", not sold through retail. You're not making sense.
No. It's you that are comparing the price that ODMs pay for Strix Halo when buying directly from AMD, to 9950X end customer retail price. Which is not fair and does not make sense at all.Once again, you're correlating ODM prices with retail prices, whilst simultaneously saying the two aren't compatible. Make up your mind.
Unless you're living in some fantasy land where you think Strix Halo (a niche product) will be sold in AIOs or mini PCs (a highly niche market) at a low price. If that's the case, I can't help you.
Too late? Or when is Arrow Lake Refresh 8P+32E expected to be released, before or after Zen6?Zen 6 Q1/2 2026?
If the final retail price for a 9950X is $999, ODMs won't be paying anywhere near that much. You have no clue how large the disparity is between retail sold parts and the prices ODMs pay.Never claimed Strix Halo will be sold directly to end customer. But end product (e.g. laptop) which contains that CPU will always be sold through retail. Unless the ODM will keep it for themselves. 🤣
No. It's you that are comparing the price that ODMs pay for Strix Halo when buying directly from AMD, to 9950X end customer retail price. Which is not fair and does not make sense at all.
Why don't you instead answer this question that you tried to avoid:
But you could of course argue that ODMs will pay less for Strix Halo than for 9950X when they buy those CPUs from AMD. Is that's what you're saying, and in that case why do you think that would be reasonable when Strix Halo has lots of extras compared to 9950X?
Never claimed Strix Halo will be sold directly to end customer. But end product (e.g. laptop) which contains that CPU will always be sold through retail. Unless the ODM will keep it for themselves. 🤣
No. It's you that are comparing the price that ODMs pay for Strix Halo when buying directly from AMD, to 9950X end customer retail price. Which is not fair and does not make sense at all.
Why don't you instead answer this question that you tried to avoid:
But you could of course argue that ODMs will pay less for Strix Halo than for 9950X when they buy those CPUs from AMD. Is that's what you're saying, and in that case why do you think that would be reasonable when Strix Halo has lots of extras compared to 9950X?
…akin to the Zen 4 CCD's readiness for MI300A packaging.
Expected by whom? To me, it seems as it would be an economically difficult venture to sideline this chip off of premium laptops to a product category such as x86 SFF PCs. (Does a category of "premium SFF PCs" even exist in x86 land? That's the land in which customers who require workstation PCs can actually buy workstation PCs, rather than having to make do with SFF PCs.)
There's probably only one GPU Chiplet (Although MLID had eons ago that Strix Halo would go as low as 6C + 20 CU). So the max I expect for a cutdown version would be 8C Z5D + 32/36 CU.Why are you guys only talking about 16C Zen 5 + 40 CU RDNA3.5 for Strix Halo? That's only the top configuration. There will certainly be cut down parts.
If Strix Halo uses Zen 5 CCDs, it means there is a possibility to have only one CCD; 8C Zen5 + 40 CU RDNA3.5 SKU. Such an SKU would be perfect for a gaming laptop.
The analysis *really* depends on how AMD decides to construct the thing. Strix Point suggests the existence of an 8c Zen5c CCX (assuming its 4+8 config isn't all implemented on one big, shared CCX). It could be that for Strix Halo, there's, for example, an 8 core Zen5c CCX sharing the same die as the memory controller, in addition to a regular 8c Zen5 chiplet which then might not even need any kind of advanced packaging.
Mi300 uses unmodified Zen4 CCDs, so not necessarily. Although I'd be a bit surprised if STH went vertical.
+40% SPECINT performance may still hold up. Yes I know it was @Kepler_L2 who actually posted that specific benchmark prediction/leak, but @adroc_thurston was talking +40% in general.
Why are you guys only talking about 16C Zen 5 + 40 CU RDNA3.5 for Strix Halo? That's only the top configuration. There will certainly be cut down parts.
If Strix Halo uses Zen 5 CCDs, it means there is a possibility to have only one CCD; 8C Zen5 + 40 CU RDNA3.5 SKU. Such an SKU would be perfect for a gaming laptop.
Strix Halo will go as low as 8C/20 CUs on a 128 bit bus.There's probably only one GPU Chiplet (Although MLID had eons ago that Strix Halo would go as low as 6C + 20 CU). So the max I expect for a cutdown version would be 8C Z5D + 32/36 CU.
That's a different GPU chiplet, right? No way they're cutting down half the gpu.Strix Halo will go as low as 8C/20 CUs on a 128 bit bus.
Would be awesome if all of this is true (and we know that many of the specs are accurate), but looking at all the data in unison, it looks like someones wet dream.
Only 20-25% ipc uplift?
AFAIK even some sources that claimed 35+% in SPEC mentioned that Cinebench is ~20% uplift only.Only 20-25% ipc uplift?
What do you mean, "AMD"? They make processors, not memory. It's ODM that uses 32 bit wide memory packages (x32). Steam Deck OLED for example is using x64 chips. My main guess as to why most ODMs opt to use x32 is that lower pad count chips have lower PCB quality requirements. Apple goes around this by implementing memory directly on the soc packageWhy is AMD using 32-bit LPDDR packages?
Won't be more cost & space efficient to use 64-bit or even 128-bit packages?
Apple uses 4 memory packages for M3 Max, which is a 512 bit chip. That means each memory package is 128 bit.
Well tbh this doesn't seem good to me even though cb is at the lower end. I'll still settle for anything >30% though.AFAIK even some sources that claimed 35+% in SPEC mentioned that Cinebench is ~20% uplift only.
Cinebench is really outliving its as a representative of a quick and decent MT benchmark for modern CPUs (similarily to CPU-Z long before it). It's a rather unique workflow that doesn't match as well to other CPU intensive tasks as it used to. Particularily now when actual renders in Cinema 4D are mostly done on GPUs.
Chips and Cheese has an excellent article about it:
Cinebench 2024: Reviewing the Benchmark
Maxon’s Cinebench is a perennial benchmark favorite. It’s free, easy to run, and scales across as many cores as you can give it. Its $0 cost allows the internet to provide plenty of res…chipsandcheese.com
Cinebench has been really popular for ages as it's free, runs relatively quickly and shows a pretty picture while successfully pegging your CPU to 100%. And i'ts been popular way longer than some claim. There is some odd consensus as it's rise to fame was only when AMD used it to demo Zen 1, when it was really widely used at least since the Athlon64 / Core 2 days.
A strong point of Cinebench is that it runs well on most PCs, be it Macos + ARM, Snapdragon + Windows or AMD / intel. I really wish we'd get something new in addition to it. Geekbench is fine, but it's ultra-light (as it's also needs to run on phones) and less useful for high core-count parts since version 6. I'd want something in addition to it. Something that would really stress PCs and melt phones
Strix Halo | RX 7700 XT | RX 6800 | RX 6750 XT | RX 7600 XT | |
---|---|---|---|---|---|
Arch | RDNA 3.5 | RDNA 3 | RDNA 2 | RDNA2 | RDNA 3 |
CUs (Compute Units) | 40 | 54 | 60 | 40 | 32 |
Boost Clock | ~3 Ghz | 2.54 GHz | 2.1 GHz | 2.6 GHz | 2.75 Ghz |
LLC (Last Level Cache) | 32 MB | 48 MB | 128 MB (half the BW) | 96 MB (half BW) | 32 MB |
Memory bandwidth | 273GB/s | 432.0 GB/s | 512.0 GB/s | 432 GB/s | 288.0 GB/s |
TDP | 140W (without memory but shared with CPU) | 245 W | 250 W | 250 W | 190 W |
If that s real numbers and not speculations then the perf improvement, not IPC, would be 26% in ST and 28% in MT for CB R23, if they targeted a bigger push in INT that would make sense because servers rely mainly on this kind of code.Well tbh this doesn't seem good to me even though cb is at the lower end. I'll still settle for anything >30% though.
now also on Videocardz:
AMD "Strix Halo" Zen5 & RDNA3.5 premium APU rumors take shape - VideoCardz.com
AMD first ultra-high-end APU with 16 Zen5 cores and 40 RDNA3.5 CUs is said to feature LPDDR5X memory Recent leaks and speculation were put together by Chiphell forum member. AMD is said to be launching its Strix Halo APUs next year. It’s one of the most interesting hardware releases from AMD in...videocardz.com
Overall a pretty impressive chip. Here it is compared to the RX 7700 XT , RX 6800 and RX 6850 XT
Strix Halo RX 7700 XT RX 6800 RX 6750 XT Arch RDNA 3.5 RDNA 3 RDNA 2 RDNA2 CUs (Compute Units) 40 54 60 40 Boost Clock ~3 Ghz 2.54 GHz 2.1 GHz 2.6 GHz LLC (Last Level Cache) 32 MB 48 MB 128 MB (half the BW) 96 MB (half BW) Memory bandwidth ~500 GB/s 432.0 GB/s 512.0 GB/s 432 GB/s TDP 140W (without memory but shared with CPU) 245 W 250 W 250 W
If the specs hold up, it has:
Considering the RX 6750 XT is about 13% slower than the RX 7700 XT I'd guess Strix Halo would perform overall similarily . Perhaps a few percent faster @ 1080p and slightly slower @ 440p (due to smaller cache).
- 13% less (theoretical) compute resources vs the desktop RX 7700 XT
- 15% more memory bandwidth than the desktop 7700 XT
- 33% less LLC tham the desktop RX 7700 XT
I'm particularly impressed by the memory bandwith. Thats the same that may desktop Radeon 6800 gets with a 256 bit GDDR6 bus (though at conservative clocks). And the performance won't be that far off. quite an achievement indeed.
If RDNA 3.5 has some RT improvements vs RDNA 3 (as seems to be the case for the PS5 Pro GPU at least) it could even rival it in some modern titles.
One should also see how limited is the single core clock in a mobile (premium) chip compared to the desktop version. I suspect not much, but not negligible. Again, if these numbers are real and not speculation.If that s real numbers and not speculations then the perf improvement, not IPC, would be 26% in ST and 28% in MT for CB R23, if they targeted a bigger push in INT that would make sense because servers rely mainly on this kind of code.
Makes much more sense , I'll fix that in my tableI presume that bandwidth figure is effective bandwidth like on the 7600XT rather than pure memory bandwidth.
EDIT: The maths works out to 8533 RAM on a 256 bit bus would be around 273GB/s of bandwidth so a bit more than the 7600XT.