If Zen6 remains AM5-compatible AND becomes dependent on CUDIMMs, that means getting new boards since existing AM5 boards don't support CUDIMMs. Which would sort-of break backwards compatibility with older AM5 boards, unless you just want terrible performance due to bandwidth starvation.9600MT/s CUDIMM.
Literally the same page as my post saying som people think there is a market for 4 channel boards, and someone is asking for 4 channel mainstream boards. Remarkably quick. Chapeau.Moving to quad channel would mitigate any memory bandwidth concerns.
Honestly, I think desktop would be great with 32 cores. It would allow AMD to resolve the odd HEDT problem that Threadripper has.
Example:
Athlon - Dual channel memory - Up to 8 cores -20 PCIE 4.0 lanes
Ryzen - Quad Channel Memory - 8 - 32 cores - 40 PCIE 5.0 lanes (example prices: $1299 32c, $999 24c, $699 16c, X3D +$100)
Threadripper - 8 channel memory - 32+ cores - 80 PCIE 5.0 lanes
EPYC - 12+ Channel memory 32+ cores 80+ PCIE lanes.
Of course it probably won’t ever happen, but one can dream. 🤣
My old Zen 1 Threadripper had quad channel boards for $300-$500. While that was DDR4 and before inflation, I imagine they could get it done for $600-$800. The X chipset could be quad channel while the B chipset could be dual channel. The A chipset could be similar to B except designed for Athlon.
Indeed. It's a small fraction of users. How many people need massive multi thread grunt but don't get enough/any economic benefit from it to be able to afford actual massive multi thread grunt. It's tiny. Since users don't generate economic benefit the price has to be low to sell it, and then you have the risk of some customers that actually pay a decent amount now choosing this cheap option. So this tiny market would have to be substantial to pay for the downside of the customers downgrading to this new option rather than buying thread ripper or epyc.Well, that is kina my point.
Most (the vast majority) of desktop/laptop users would benefit much more from one "super core" instead of 24 cores. For those that need LOTS more cores, you are really better off with a workstation chip that is specifically designed to feed those cores with lots of memory .... but I am guessing that this is a small fraction of users.
Eh? I thought CUDIMM's just had a timer on the memory pcb's? No active support needed by the board and cpu?If Zen6 remains AM5-compatible AND becomes dependent on CUDIMMs, that means getting new boards since existing AM5 boards don't support CUDIMMs.
To prevent compatibility issues, the chipset and motherboard manufacturers will roll out BIOS updates to legacy motherboards that disable the clock driver if a CUDIMM/CSODIMM is installed. When a CUDIMM or CSODIMM is installed on a legacy system, it will automatically operate at the supported speed of that chipset and processor.
If Zen6 remains AM5-compatible AND becomes dependent on CUDIMMs, that means getting new boards since existing AM5 boards don't support CUDIMMs. Which would sort-of break backwards compatibility with older AM5 boards, unless you just want terrible performance due to bandwidth starvation.
That actually surprises me. I had thought they were completely socket-incompatible and had a different pin interface.Eh? I thought CUDIMM's just had a timer on the memory pcb's? No active support needed by the board and cpu?
Edit... a quick google;
Source
If Zen6 remains AM5-compatible AND becomes dependent on CUDIMMs, that means getting new boards since existing AM5 boards don't support CUDIMMs. Which would sort-of break backwards compatibility with older AM5 boards, unless you just want terrible performance due to bandwidth starvation.
That being said, we don't know that Zen6 will remain on AM5 or what will happen with it.
It will find its niche I guess. How big remains to be seen, but they might take-over the SFF community. Also it's not like the socket has to go, they could still do a daughter board (SoC + Mem) plugged into motherboard (IO expansion, power delivery, CXL additional RAM etc). But I would expect for general DIY PC market, the socketed CPUs are here to stay for couple more years for x64. What WoA crowd will do will also be interesting in this context.I wonder how popular a "no socket" option becomes, when CPUs like Strix Halo and Medusa Helo are sold permanently attached to mini motherboards.
Wrong.64MB zen6c CCX
It is my opinion and I know many disagree with this. There is no more low-hanging fruit when it comes to single thread performance. There is only so much instruction parallelism that can be extracted from a sequence of instructions that have dependencies on one another. It's not a problem of figuring it out or adding more transistors. The problem is the code itself.
I would expect the next generation of cpu cores from AMD and Intel to show an average IPC increase between 5 and 15%. With my estimate being 10%. I know this will have a lot of people up in arms, what's the use? But this is just a reality of decreasing your distance to a wall by half with every step. We've been walking towards the wall for a long time and the steps are indeed very small at this point.
Making my post even more grim is the fact that clock speeds using conventional cooling also seem to have topped out.
Of course I will be happy to be proven wrong when Zen 6 arrives with a 20% IPC gain. I just don't think it will happen.
Where is left to go? Well, it is obviously better multi-thread performance which is exactly where AMD and Intel are headed.
Increased performance, assuming CPU manufacturers provide us with massive multi-threaded performance is going to be in the hands of software developers that will need to make use of this multi-threaded compute.
Of course. I am hopeful too. Sandy and Zen 3 were back when the fruit was pretty low and easily picked.People have been saying this, but then we get things like Sandy Bridge and Zen 3. You may be right, but I like to remain hopeful.
It's hard to believe this when there exist competitors with easily 50% higher IPC and clock rates that are closing in on Zen every year.Of course. I am hopeful too. Sandy and Zen 3 were back when the fruit was pretty low and easily picked.
In your dreams.....It's hard to believe this when there exist competitors with easily 50% higher IPC and clock rates that are closing in on Zen every year.
And yet Apple M cores are able to do so. Since their design clocks lower but is wider and it can match or exceed Intel and AMD in not SIMD workloads. So if anyone could be standing next to a wall it is Apple what gives AMD and Intel still few steps to make to catch up. After all somebody has shown it is possible.There is only so much instruction parallelism that can be extracted from a sequence of instructions that have dependencies on one another. It's not a problem of figuring it out or adding more transistors. The problem is the code itself.
It's hard to believe this when there exist competitors with easily 50% higher IPC and clock rates that are closing in on Zen every year.
I believe he means Apple. This year Apple may reach 4.95GHz on its M5 P core, closing in on mobile Zen5 which is at 5.1GHz. Apple has a node advantage this year, so I would wait till late 2026 to see Zen6 mobile on N2 and M6 core on N2 and then compare.In your dreams.....
Please name this unheard of competitor.
Every single ARM-based performance core from 2024 is able to do so tbh. Oryon (L), Blackhawk, M4-p.And yet Apple M cores are able to do so. Since their design clocks lower but is wider and it can match or exceed Intel and AMD in not SIMD workloads. So if anyone could be standing next to a wall it is Apple what gives AMD and Intel still few steps to make to catch up. After all somebody has shown it is possible.
I don't think Apple M runs Windows with very good IPC, right?. I don't give two sh*ts about Apple and their confounding and restrictive OS. It's great for some people, but not me. I like to be able to "get in the box" and run literally millions of apps from yesterday and 50 years ago. Windows runs everything.. from apps designed by hundreds of highly qualified engineers to smart kids in their basements. Apple is East Berlin before the wall came down. Everything is controlled and set up for Apple. x86 has to do it all. Yes, freedom is harder to lock down for optimizing. Apple? No thanks.And yet Apple M cores are able to do so. Since their design clocks lower but is wider and it can match or exceed Intel and AMD in not SIMD workloads. So if anyone could be standing next to a wall it is Apple what gives AMD and Intel still few steps to make to catch up. After all somebody has shown it is possible.
Also please don't read that as me disagreeing as a whole with what you wrote but I think others are showing there is still room to grow.
I agree that apple has good IPC, but not 50% higher, and its not a competitor of Zen or Intel. And 50% higher clock rates ? 7.5 ghz .. No way.I believe he means Apple. This year Apple may reach 4.95GHz on its M5 P core, closing in on mobile Zen5 which is at 5.1GHz. Apple has a node advantage this year, so I would wait till late 2026 to see Zen6 mobile on N2 and M6 core on N2 and then compare.
The point is if a ARM core thats wider than x86 is able clock near to mobile Zen 6 next year, then AMD and Intel still have low fruit to pick.
32C CCD is just for serversI've been hearing about a 32 core Zen 6c desktop variant: https://wccftech.com/amd-next-gen-ryzen-zen-6-medusa-ridge-cpus-12-24-32-core-up-to-128-mb-l3-cache/
I am assuming that this would be accomplished by using a single server 32c CCD and an IOD. What is really interesting about this setup is the potential for a 64 core Zen 6c with two 32c CCD's and an IOD.
The worry for such cores for me would be .... can you feed them with only dual channel memory?
I can see feeding a 32c Zen 6c with DDR5 8000 (maybe with a 1:2 divider?), but 64? Starts looking pretty iffy. Perhaps that will end up being the reason AMD only has its non-thread ripper desktop processors reach 32c. Beyond that you might well need quad channel memory to feed it.... and then it would be Threadripper.
So its likely R5 8C/16T, R7 12/24T, R9 24/48T32C CCD is just for servers