That is a value for the entire gpu, what about the memory part itself? Also consumers don't need 300GB of it (yet).
But to answer - memory is progressing very slowly. DDR4 to DDR5 was not even a meaningful jump. Even PCIe SSDs are slowly catching up to it which is both funny and sad.
As for the usecase - I use my memory as a cache for everything. Every system in the last 15-20 years I used I maxed out memory on, I never cared much about speed of my storage, because after loading everything into RAM, the system and apps feel a lot more responsive. The difference on older systems with HDDs were especially noticeable, but even on an SSDs, things have not improved much due to latencies. Of course using any webapp connecting to the network will negate any benefits of this, but it makes a difference with desktop apps.
These days I even have enough memory to be able to run local test VMs so I don't need to use server resources.
Coincidentally, the first issue (referencing Navi 21) was the one I started these experiments with, and this turned out to be pretty informative.
Our Navi 21 would almost always go AWOL after a test run had been completed, requiring a full reboot. At some point, I noticed that this only happened when our test runner was driving the test; I never had an issue when testing interactively. I eventually realized that our test driver was simply killing the VM when the test was done, which is fine for a CPU-based test, but this messed with the GPU's state. When working interactively, I was always shutting down the host cleanly, which apparently resolved this. A patch to our test runner to cleanly shut down VMs fixed this.
And I've had no luck with iGPUs, as referenced by the second issue.
From what I understand, I don't think that consumer AMD GPUs can/will ever be fully supported, because the GPU reset mechanisms of older cards are so complex. That's why things like vendor-reset [3] exist, which apparently duplicate a lot of the in-kernel driver code but ultimately only twiddle some bits.
She also had no choice, as SBF was blaming her. The point being that they still didn't really need her help. It was obvious that he committed fraud, and there was plenty of proof of it.
I mean, the guy was constantly high on nootropics and they had no idea what actual investments FTX made. I'd imagine most of the time was just spent untangling that web, his case was more or less a slam dunk.
SemiAnalysis made this a base requirement for being appropriately ranked on their ClusterMAX report, telling me it is akin to FAA certifications, and then getting hacked themselves for not enforcing simple security controls.
reply