B-die equivalent for DDR5?

[57er6uigyuholn]

I'm looking to build and AM5 system and am looking for RAM. On my AM4 machine I have Viper Steel 4400 which is b-die, and I got lucky and all 4 sticks can really tighten down timing a lot.

What's the AM5 equivalent ≈? I read the memory controller lottery plays a big role. Plan to run a 9950x with 2x32GB sticks

What speed can 9000 series take? What kind of timings at those speeds?

Sorry I know it's a broad question, but all i can find is really outdated or contradictory info

 

[ShrimpBrime]

DDR5 Hynix A die kits scale well with voltage similar to Samsung ddr4 B-Die.

8000mhz kits, proper motherboard. Should be no problem.

 

[mkppo]

A die, or the 24GB rev.2 m-die kits. Crosshair Apex/Gene/GB Tachyon will be able to do 8200/8400 if your CPU's IMC is up to scratch. Had to go through 5 9950x's to find one that does more than 8200 on the Apex though. Single CCD CPU's can clock higher, their IMC has less strain with one GMI link.

Best kit would be the G.skill 2x16GB CL26 6000 kit or 2x24GB CL26 6000 kit. A die does tighter 8000, 24GB M die's can clock higher. Lexar's CL26 6000 does come close, but the G.skill is generally superior.

 

[ir_cow]

SK Hynix (16 Gbit) 2 GB A-Die has the lowest possible timings, excluding the oddball 8 Gbit ones that are in KLEVV FIT V. Those are dual-rank 32GB kits...very strange stuff.

SK Hynix (24 Gbit) 3 GB M-Die has the highest possible frequency currently. Not far behind A-Die for sub-timings. Both M (3GB) and A (2GB) can handle 1.5V just fine with a fan. I would say 1.6V if you have the right cooling, otherwise it will error out.

 

[Sol_Badguy]

Plan to run a 9950x with 2x32GB sticks

That means 2R sticks.
Compared to 2x16GB and 2x24GB (which are 1R) it's reasonable to expect lower max. speeds.

A good starting point if you haven't already read it:

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

It's best to decide from the start if you want lower speed and tighter timings or higher speed and looser timings. Price of the kits should help you decide.
But the priority has to be capacity, you don't want to gimp your build just to have a "high performance" 32GB RAM kit.
The advantage of the Ryzen memory sweetspot of 6000 is that it allows very generous capacity 2R kits to run at that speed, sure with usually a little bit more voltage (1.4V) and not super tight timings (C30/C32).

 

[dgianstefani]

I'm looking to build and AM5 system and am looking for RAM. On my AM4 machine I have Viper Steel 4400 which is b-die, and I got lucky and all 4 sticks can really tighten down timing a lot.

What's the AM5 equivalent ≈? I read the memory controller lottery plays a big role. Plan to run a 9950x with 2x32GB sticks

What speed can 9000 series take? What kind of timings at those speeds?

Sorry I know it's a broad question, but all i can find is really outdated or contradictory info

SK Hynix (16 Gbit) 2 GB A-Die has the lowest possible timings, excluding the oddball 8 Gbit ones that are in KLEVV FIT V. Those are dual-rank 32GB kits...very strange stuff.

SK Hynix (24 Gbit) 3 GB M-Die has the highest possible frequency currently. Not far behind A-Die for sub-timings. Both M (3GB) and A (2GB) can handle 1.5V just fine with a fan. I would say 1.6V if you have the right cooling, otherwise it will error out.

2x 2R sticks technically best you can do on AMD for the write bandwidth increase, since sweetspot is 6-6400 MT anyway, so the higher MT ratings you can get with 1R sticks are irrelevant. Going for A die if 2x 2R, or M die if 2x 1R, so 2x 32 GB, or 2x 24 GB.

Avoid any 4x setup, or even motherboards with four slots unless they're the fancy ASUS ones with optimized traces, I forget what they're called and am too lazy to look it up.
With DDR5, a 2x setup gives you more than enough capacity on dual channel systems anyway (64, 96, 128 GB). For the higher capacities, you want quad or higher channel anyway.

Note the bottleneck is still the IF, which clocks at best around 2000 MHz, maybe ~2100 or 2200 if you're extremely lucky. 6000 MT '1:1' means the memory and memory controller are both running at 3000 MHz, which still means the IF limits actual benefits of that, since it's 1.5:1 to the actual cores.

Explained here.

Zen 6 may fix the IF bottleneck as new packaging and less latency penalty, which in theory should mean X3D cache is less needed, that was a mitigation for the latency penalty of chiplet design as AMD built them.

Note that the IF being 2000 MHz essentially means it didn't get faster since AM4, X570 and Zen 3 could do 2000 MHz IF/FCLK for '1:1' with 4000 MT/2000 MHz DDR4, but more common was 3800 MT.

I ran a 5800X3D with 4000/14 DRAM and 2000 MT IF, still kinda miss that rig.

Nitropath, that's what ASUS calls their improved traces, supposedly brings 4 slot boards to the level of 2 slot boards, but I am skeptical, and irritated they didn't put it on their two slot boards too.

 

[ir_cow]

Nitropath, that's what ASUS calls their improved traces, supposedly brings 4 slot boards to the level of 2 slot boards, but I am skeptical, and irritated they didn't put it on their two slot boards too.

It seems Nitropath is just shorter traces on the physical ram slots. This effectively makes those empty slots less of a RF beacon. I wonder if just putting tinfoil over the them would do the same trick.

 

[dgianstefani]

It seems Nitropath is just shorter traces on the physical ram slots.

It does work though. The question is why not put it on the two slot boards too.

 

[ir_cow]

Sorry I edited my reply before you replied.. opps.

It only affects empty slots from der8auer testing in that video. Traces to the CPU are still the same length. CAMM2 will eliminate the empty slot RF problem altogether. Also make the DRAM ICs closers to the traces as well.

 

[Kafka]

2x 2R sticks technically best you can do on AMD for the write bandwidth increase, since sweetspot is 6-6400 MT anyway, so the higher MT ratings you can get with 1R sticks are irrelevant. Going for A die if 2x 2R, or M die if 2x 1R, so 2x 32 GB, or 2x 24 GB.

Avoid any 4x setup, or even motherboards with four slots unless they're the fancy ASUS ones with optimized traces, I forget what they're called and am too lazy to look it up.
With DDR5, a 2x setup gives you more than enough capacity on dual channel systems anyway (64, 96, 128 GB). For the higher capacities, you want quad or higher channel anyway.

Note the bottleneck is still the IF, which clocks at best around 2000 MHz, maybe ~2100 or 2200 if you're extremely lucky. 6000 MT '1:1' means the memory and memory controller are both running at 3000 MHz, which still means the IF limits actual benefits of that, since it's 1.5:1 to the actual cores.

Explained here.


Zen 6 may fix the IF bottleneck as new packaging and less latency penalty, which in theory should mean X3D cache is less needed, that was a mitigation for the latency penalty of chiplet design as AMD built them.

Note that the IF being 2000 MHz essentially means it didn't get faster since AM4, X570 and Zen 3 could do 2000 MHz IF/FCLK for '1:1' with 4000 MT/2000 MHz DDR4, but more common was 3800 MT.

I ran a 5800X3D with 4000/14 DRAM and 2000 MT IF, still kinda miss that rig.

Nitropath, that's what ASUS calls their improved traces, supposedly brings 4 slot boards to the level of 2 slot boards, but I am skeptical, and irritated they didn't put it on their two slot boards too.

I remember BZ saying some optimizations are kept secret too, and there's OptiMem III to boot (not II) which appears *only* on their true high end boards but seems like they no longer use that and not much is known about it either.

As for the thread, I'd highly suggest to get Viper Xtreme 5's with 48GB (8000 MT/s) or above. 3GB M-Die is good stuff and could scale in the future for newer Ryzens even if you opt for lower speced. Think Micron B-Die towards the end of AM4 where you could easily get them to 5000 MT/s on ddr4 alder lake or just with luck on AM4, except far better than that.

The subtiming diff between A-Die and M-Die isn't that dramatic and (RAM) OCing is always up to luck anyway. I am sure you could get tRFC, for example, that does bonkers latency even on M-die with luck. I just saw a screenshot with like 150 ns, A-die does like what 130? it's not that important. You could benefit in the future from the higher MT/s with Ryzen 6.5, who knows. Just pick a decent board that's 8 layers or above imho with 2 slots if possible.

ir_cow reminded me of this absolute classic -

https://www.reddit.com/r/overclocking/comments/t82ldy/improving_the_electrical_shielding_of_ram_slots

https://www.reddit.com/r/overclocking/comments/tcfcfi/results_of_the_diy_emi_shielding_5800x_on_5100c19

 

[ir_cow]

Not gonna lie, I've put tinfoil over my slots before. Can't say it made a difference since I don't run chilled water or LN2, but I like to think it did for the effort