Discussing memory bandwidth and DDR specifications. (Bandwidth at 192gb/s)
Discussing memory bandwidth and DDR specifications. (Bandwidth at 192gb/s)
I'm examining possible future computer designs. At 12,000MHz and 192GB bandwidth, how would different RAM sizes affect performance? Does using both sides of DDR chips double the usable space from 96GB? Is this more than just accounting for DDR speed in the bandwidth? Would it work better with a quad-channel board? Could it power an entire 192GB memory bank steadily? Are there options with 4x48GB modules? I was expecting AM5 processors to support 24 or 48 cores/threads, letting full 192GB run at 12,000MHz DDR5. I also wanted a new NVIDIA model with 48GB of VRAM. Preferably with 48GB RAM. By the way, can a GPU benefit from a high-speed RAM like this? Does it speed up communication between components? What happens if you run at twice the speed on an AMD Infinity Fabric? And did this idea ever get built?
This situation won't occur until the next generation of RAM (DDR6 or QDR1). At that point, specifications will likely shift enough that current statements won't hold. Typically, performance reached near the end of each DDR update was measured through official frequency tests within a few months. For DDR3, that was around 2600MT/s on Haswell; DDR4 reached roughly 5000MT/s, and DDR5 now sits between 10,000–11,000MT/s. This is what top memory systems, boards, and CPUs should be hitting in the last couple of years of DDR5 use. The figure you mention—12000 "MHz" (probably meant MT/s)—doesn't reflect reality. Memory speed and capacity have little impact on each other. The upper limit for memory data transfer remains constant regardless of attempts to boost it. DDR5’s 6000 in dual channel tops at about 96GB/s, irrespective of kit size. Variations come from timing, layout, and interleaving, but under ideal conditions it won’t exceed that number. Doubling channels doubles bandwidth; the math confirms this. CPU memory limits matter more than board capabilities. Achieving ~7200MT/s with four 16GB modules is possible on Intel, though AMD struggled to match stability at similar speeds. Ultimately, the focus shifts to GPU access rather than raw memory speed. The concept of Resizeable BAR highlights how GPUs interact with memory, not how much data can be moved per second. On AM5, there’s no limit to infinity fabric speed (beyond ~5200MT/s), but synchronization remains a challenge. 2:1 mode on AMD means controllers run at half the memory speed, which helps latency but adds penalty. It offers only marginal gains for high-bandwidth tasks and introduces stability issues. Single CCD CPUs like Ryzen 7000 don’t benefit from higher infinity fabric bandwidth. If you have questions, feel free to ask—I tried to cover everything thoroughly.
What are the RAM limits for this model? I’m thinking about getting a Haswell combo since there’s a used one with a 4670k for a low price. If the memory size is smaller than what I expect, I might be let down compared to Bloomfield or Gulftown. I’ve been able to run 2800C11 stable at 2.1GHz on air—especially with hot air in Indonesia—and have dual-rank Hynix CFRS chips on my i7 930 and 950. That should help the sticks last longer before they wear out further. Maybe I’ll try some 2Gb SAMSUNG dies instead, hoping they won’t degrade as quickly (around 3000 vs 2800). I don’t see this happening unless 48GB modules come out, which I doubt will be the case unless it’s for servers.
Right now, my experience shows RAM speeds are really high—around 10800 MHz on AMD, with an 8500G on a B650M HDV. Even budget boards can handle that. I’m worried about CPU memory performance too. Intel’s DRDR5 ICs are inconsistent, especially at higher speeds (around 8600-8800 for good ones). ZEN 4 tops out near 8000 with a 2:1 clock speed, which is probably not enough for quad rank or even quad sticks. AMD APUs look better in this regard, but I’m not sure yet.
The main issue here is CPU memory speeds. They’re not very reliable, and both brands seem to struggle with quad-rank and possibly quad-stick configurations. The AMD APUs might improve if voltage is adjusted, but I doubt they’ll hit 12000+ unless we get better 48GB modules. For now, I’m sticking to checking benchmarks like WPM and SPI instead of relying on just one number.
The frequencies I mentioned were just the basic ones that usually work. You can definitely improve on Haswell, but my real-world results with a 4x2GB D drive won’t push past 2666 on a Z87X-OC. If you’re aiming for single rank CFR or similar, you’ll see better speeds, though reaching 3000 remains a challenge.
I was working on a 2800 dual rank setup on a Bloomfield. At first it felt frustrating because I had to locate the correct voltage and only used two sticks since the low-end Giga X58 lacks a third channel. For some reason, trying between 2500-2600 with three sticks caused it to lock into C3 or C5 codes. If I switch to dual rank Hynix CFRS, I managed to boot it successfully, but the RAM limit was a problem and Windows booted fine. Stabilizing remains a challenge, especially since the Hynix CFRS doesn’t perform well above 2600 and the system often stalls at C1. I still can’t replicate that experience except for one time. I also tried dual rank CFRS with dual rank Hyundai, which worked easily up to 3100, but the RAM cap was an issue.
I’m close to getting a W3680 instead of a G1 Assassin after careful thought. I’m hoping it will run smoothly between 3400-3600 on air, though I’d like higher speeds if I push above 1.7V. I’ve seen these boards handle up to 2.6V without major issues, but they tend to degrade past 2V. The Z87-G45, 4670K, and a few EPLIDA sticks are in it. I doubt I’ll exceed 1600 unless I negotiate a better price.
For the X58 boards, the G1 Assassin is worth considering, but there’s also an EVGA X58 SL and FTW3 that might be better at RAM speeds. I think I’ll hold off until a more reliable model arrives.