80GB of Memory Wouldn’t Make a Mining GPU an A100
Key takeaways
- A reported capacity of 80GB does not prove that all 80GB works.
- Voltage glitching can disrupt a chip’s operation, but it cannot repair defective memory.
- A successful unlock on one card would not establish why its resources were disabled.
- More usable memory could expand AI workloads without matching an A100’s performance or reliability.
The claim that NVIDIA’s CMP 170HX mining GPU can go from 8GB to 80GB has an obvious audience: anyone who has tried squeezing an AI model into too little memory. A tenfold increase sounds like a spectacular bargain. Whether it actually delivers one depends on what “unlocked” means.
More memory changes what fits
GPU memory holds model weights and the working data generated during computation. Capacity can therefore decide whether a workload runs at all.
If a CMP 170HX genuinely gained access to 80GB of usable memory, that could substantially expand the jobs it could handle. That would be useful even without a dramatic increase in computing speed.
But fitting a model into memory is only the entrance requirement. Compute resources, clock speeds, memory bandwidth, and the connection to the rest of the system all affect how quickly it runs.
An A100 comparison also needs to account for reliability. Long computations require confidence that errors will be detected. Support matters when something goes wrong.
Matching one number on a specification sheet would leave all those questions open. The distinction is usable capacity versus overall capability.
An 80GB readout is the beginning of the test
A driver reporting 80GB would be interesting. It would not, by itself, establish that the card has 80GB of functioning memory.
The first test is straightforward in principle: write data throughout the claimed capacity, then read it back and check that it matches. Different addresses must also correspond to distinct storage locations. Otherwise, writes to apparently separate regions could overwrite each other.
Then comes sustained load. A program running briefly tells you much less than a workload repeatedly using the full memory capacity without errors.
There is a physical prerequisite, too. If the claim is that previously inaccessible onboard memory has been restored, that memory must already exist on the card. Voltage tricks do not manufacture additional memory chips.
The meaningful result would be 80GB of usable memory, backed by tests of its full capacity and stability. A larger number in a system utility is a much smaller achievement.
What voltage glitching could change
Voltage glitching briefly disturbs the voltage supplied to a chip, inducing a fault in its normal operation.
Circuits depend on voltage and timing staying within expected conditions. Disrupt those conditions at the right moment, and an operation or decision may go wrong. A check or configuration step could consequently behave differently.
That is a different objective from overclocking, which aims to increase operating speed.
Still, saying that an experiment used voltage glitching does not identify the mechanism behind an apparent unlock. Did it affect a configuration that controls memory capacity? Did it interfere with some other check? Those are separate questions that need separate evidence.
And bypassing a restriction does not repair the hardware behind it. If a memory region is physically defective, getting access to it will not make it dependable.
An unlock cannot explain the product strategy
Chips made from the same design do not all emerge from manufacturing with identical capabilities. Some contain defective units. Others cannot operate reliably at the intended speed.
Manufacturers can disable problematic areas and sell the remaining hardware in a lower specification product. Sorting chips by their operating characteristics is known as binning. It helps recover value from silicon that might otherwise be discarded.
Commercial segmentation can produce a similar outward result. A manufacturer may restrict working resources to distinguish products at different prices. Both practices can exist within the same product family.
For the buyer, the visible symptom is the same: part of the hardware is unavailable.
Suppose one CMP 170HX were shown to run reliably with additional memory enabled. That would establish something meaningful about that particular card: usable resources existed behind its restrictions.
It would not establish that every CMP 170HX has the same resources available. Nor would it prove that every disabled region originally met the manufacturer’s qualification requirements.
Strong testing could make the segmentation question more pointed. It still would not let one successful modification explain the condition of an entire product line.
An 80GB unlock would deserve attention if the full capacity worked reliably under sustained load. Cheap memory can be valuable. Knowing what you can trust it to do is what makes it a bargain.
Comments
Loading comments...