I think availability will be fine for Medusa Halo specifically. The volumes will not be there to support mainstream Zen6 products, but a relatively low-volume halo product that's also expected to release fairly late in the cycle, I think there would be enough memory for them.
There would be a price premium.
I think that the final puzzle piece for Medusa Halo is a 3d cache chiplet. I think that the 8060s is enough for the next year, but for the applications where it matters, such as games or applications with well sized data sets, having at least a single CCD with 3d cache would make a notable difference in performance. One of the issue that Strix Halo is facing is that it's coupled with LPDDR5/X right now. That's a latency hit for the CPU cores that isn't hidden by the MALL cache. While they have a ton of bandwidth and decently sized L3, anything that spills has to wait. 3d cache could help make the difference there.
Yes cause I don't see any next gen product before 2028 for both Intel/AMD maybe Razor Lake is 2027 but idk for sure.
WE will see
Strix Halo already introduced its very own new "socket" (rather: new BGA pinout). Why shouldn't Medusa Halo come with yet another new BGA pinout also? A downside of course would be extra design and validation steps TBD by AMD and ODMs. The iGPU however would profit from the additional memory bandwidth. For work and recreation alike.
It already has VERY obvious issues with market penetration and OEM buy in. Now you want to scrap all the little work that was already done for a follow on? That's how you strangle products in their crib!
Depends on who is going to do the hard parts of the work: Still the ODMs, or AMD for a change?
If Zen 6 comes with a new IOD I would be very surprised if AMD actually ditches it within a single generation.
There are non-zero odds that there isn't a new IOD with Zen 6. And that would make sense if the timelines for DDR6 production line up for Z7.
Josh128
Senior member
- Oct 14, 2022
- 929
- 1,410
- 106
I think theres a 99% chance Zen 6 gets a new IOD. What it brings, however, remains to be seen. Real CUDIMM support and possibly LP cores?
They don't have to ditch it. Remember AM4? AM5 will continue to be supported and remain relevant for several years after the new socket launch for DDR6. Give at least two years for DDR6 widespread adoption and prices to normalize. Another two or even four years of post-DDR6 AM5 support. So the new IOD could be useful for something like 5 or 6 years.
I believe the DDR6 standard has yet to be finalized, its standardization moved slower than LPDDR6's did. We'll see LPDDR6 support appear in PCs well before DDR6.
There's a good chance we'll see DDR6 used only on the high end where maximal potential memory capacity (i.e. terabyte or greater in today's market) is required. When all the mainstream stuff leaves DDR5 behind I think it will be for LPDDR6 (CAMM or soldered, depending on market segment) not DDR6.
If you make a hybrid controller able to handle either LPDDR5X or LPDDR6 it delivers 24 bits of LPDDR6 or 16 bits of LPDDR5X, that's where that difference comes from.
A bug hybrid controller will take a lot of space it will be either DDR5 or DDR6 no in between
Because of the short sighted decision on part of AMD to NOT include Memory On Package. That one is biting AMD in the backside very badly...
Last edited:
You said no DDR6 / LPDDR6 before 2028.
Part of the problem with that statement is mixing LPDDR6 and DDR6. LPDDR6 is its own thing, for mobile phones. High volume production is supposed to start before the end of 2025. Customers are phone OEMS. Phone SoC makers will support it very soon.
Which means, LPDDR6 will already be in the phones from 2026. So I don't see how price and availability would push a higher end product (high end laptop) 2 years after the phones.
As far as no new product before 2028, again, that's mixing DDR6 into the mix, which is a separate thing, for servers and desktops, not for Medusa Halo
Zen 6 is raising core counts up by 50%. DDR5 8000 raises memory bandwidth by 43%. Seems like that covers the additional bandwidth considering that Zen 5 is RARELY bandwidth constrained and is MUCH more performance dependent on memory LATENCY.
DDR5 Memory Performance Scaling with AMD Zen 5
Our AMD Zen 5 Memory Scaling Review examines six different DDR5 memory speeds, including DDR5-8000, DDR5-6400 1:1, DDR5-6400 CL28 and others. We looked into how these speeds impact performance in gaming at four resolutions and a wide range of application workloads.
I see no case at all for AMD to go with anything other than DDR5 8000.
Even in the future for Zen 7, I think it very likely that AMD will just stick with faster DDR5 vs DDR6 (which is not even a ratified spec yet?). I think that the big bandwidth change will be with Zen 6 (going from 8 core CCD's to 12 core CCD's). Considering the trivial die shrinks expected for the next 2 TSMC process nodes, I don't see Zen 7 having any more cores than Zen 6. Likely just some strategic use of better logic in schedulers, some better buffering, lower latency trickery in the interfaces, etc to free up performance over Zen 6. No more cores.
I do think that the new IOD is going to unlock more performance than people are giving it credit for though. Lowering the main memory latency (as the designers of ARL can attest to) is a really big deal.
Funny that Intel totally threw away their RAM latency advantage. No one there had the courage to put a stop to the madness of tile architecture when it wasn't ready for prime time.
Darkmont
Member
- Jul 7, 2023
- 30
- 66
- 61
Memory latency and bandwidth are dependent variables of one another. The reason people don't see large bandwidth increases is because the fabric width and clock speed limits bandwidth per CCD. If most systems are running at an FCLK of 2000-2200 MHz you're being duped if you think that means higher memory bandwidth isn't impactful. A 32 byte interface at 2 GHz gives you 64 GB/s. 50% of the theoretical bandwidth that a 128 bit 8 Gbps memory bus supplies. You could only saturate that if you have a Dual CCD chip, and games will isolate themselves to one CCD where possible.Zen 6 is raising core counts up by 50%. DDR5 8000 raises memory bandwidth by 43%. Seems like that covers the additional bandwidth considering that Zen 5 is RARELY bandwidth constrained and is MUCH more performance dependent on memory LATENCY.
DDR5 Memory Performance Scaling with AMD Zen 5
Our AMD Zen 5 Memory Scaling Review examines six different DDR5 memory speeds, including DDR5-8000, DDR5-6400 1:1, DDR5-6400 CL28 and others. We looked into how these speeds impact performance in gaming at four resolutions and a wide range of application workloads.www.techpowerup.com
Memory bandwidth vs latency response curve
Memory bound applications are sensitive to memory latency and bandwidth that’s why it’s important to measure and monitor them.Even if this two concepts are often described independentl…
mahmoudhatem.wordpress.com
TRENDING THREADS
-
Discussion Zen 5 Speculation (EPYC Turin and Strix Point/Granite Ridge - Ryzen 9000)- Started by DisEnchantment
- Replies: 24K
-
Discussion Intel Meteor, Arrow, Lunar & Panther Lakes Discussion Threads
- Started by Tigerick
- Replies: 20K
-
-
-
Facebook
Twitter