Could AMD develop a motherboard chipset capable of handling GDDR6 or GDDR7 standards?
Could AMD develop a motherboard chipset capable of handling GDDR6 or GDDR7 standards?
If AMD has the capability to create motherboards and chipsets for Sony and MS, it's unclear why they wouldn't develop a consumer motherboard chipset with these features: an upgraded socketed CPU, support for non-upgradable GDDR6 or GDDR7 memory, and PCIe 5.0 slots. Regarding point 2, if manufacturers included soldered, non-replaceable GDDR6 or GDDR7 chips with large capacities, it might indeed make integrating such memory impractical.
GDDRwhatever is mainly related to the GPU, not the motherboard.
The xxGB might be permanently installed on the motherboard for some reason.
It would actually be quite simple, just a minor hassle to adjust a motherboard for soldered GDDR6 or GDDR7. The issue lies in the fact that the memory controller is built into the CPU, meaning it would require a CPU that can communicate with these memory types. Where would such a CPU come from? Indeed, they would need specialized CPUs as well. Unfortunately, GDDR6 and GDDR7 are optimized for high bandwidth but not ideal for low latency or power efficiency, making them less suitable compared to alternatives like LPDDR5X for typical CPU workloads.
This version is technically present, excluding the socketed CPU and available PCIe slot, yet it includes a strong GPU. Asrock introduced these devices several years ago under the name cryptominers, and they have been successfully run in Linux recently. Additionally, newer strix halo models are also available.
BC-250 essentially functions as a PS5-style APU on a board with RAM soldered in, utilizing GDDR6 memory. This choice was driven by the necessity for high memory bandwidth during mining operations (which is why various GPU vBIOSes would reduce clock speeds but increase memory frequency). Once a standard operating system was installed, the memory latency came to approximately 151ns.
That delay is quite high. I assumed FPM DRAM was designed for about 60ns. Is that the level seen by AMD and Nvidia GPUs?
151ns is actually considered extraordinarily tight for GDDR6 and GPUs typically see far worse, because there
it doesn't matter
60ns is technically the tRAC rating on that FPM and represents the minimum random access time for the first data out
at the pins
. That is definitely
not
the CPU access time since it's not SDRAM so is asynchronous, meaning the 33-to-66MHz memory controller in the chipset also has to insert wait states, plus then there is a separate FSB between the chipset and the CPU too with its own delays for data integrity
If there was L2 cache (at that point in time any L2 was connected to the chipset so ran at the bus clock rather than the CPU speed), then FPM can send a whole page of 4 columns into it one after the other, where only the first column has tRAC and the others only tPC in between, and for 60ns FPM this tended to be 35ns. So 4 sequential cells can be read at 60+35+35+35 or 165ns, rather than fully random reads of 4 cells at 4x 60ns = 240ns... again to the RAM pins rather than the CPU
In the end, that 60ns FPM was around 300ns total latency and topped out around 50-80MB/s in Everest (predecessor to AIDA64). So probably close to what GPUs see in latency, only they can get over 4000x more memory bandwidth. 300ns is of course still only 0.0003ms so even 35 year old PC RAM is way lower latency than paging to any modern SSD, despite its modest bandwidth looking very USB-like today