Boom and bust —

Why this year’s insane graphics-card price surge might be over

You can finally buy a high-end graphics card at Best Buy again.

Timothy B. Lee - Apr 30, 2018 11:15 am UTC

Image of a line graph. — Enlarge / Radeon graphics card price data (left axis) from PC Part Picker. Ether price data (right axis) from Etherscan.io.
Timothy B. Lee / Ars Technica

The massive computing power of high-end consumer graphics cards make them ideal for mining Ethereum and a number of lesser-known cryptocurrencies. (Bitcoin's simpler mining algorithm long ago came to be dominated by custom ASICs, making GPU-based bitcoin mining unprofitable.) As a result, the cryptocurrency boom that began last year led to a boom in demand for graphics cards.

Prices soared last summer, then soared even more in early 2018. By February, the going rate for high-end graphics cards like the Radeon RX 570 and RX 580 was more than double what it had been nine months earlier.

But now the graphics-card boom seems to be coming to an end.

Two reasons the GPU price slump is likely to continue

The big question, of course, is whether recent price declines will continue in the coming weeks. As you can see in the chart above, ether's value has been on a bit of a tear recently, rising from less than $400 to more than $600 over the last month. But it's still way below the January peak of almost $1,400—and so mining still isn't nearly as profitable as it was earlier this year.

"Unless Ethereum price goes way up, you are better off buying ETH from exchange than buying a GPU to mine with it," an active Ethereum miner told me by email. "I expect prices of GPUs to continue dropping."

Promoted Comments

Rene Gollent Ars Tribunus Militum et Subscriptor

jump to post

fuzzyfuzzyfungus wrote:
Can anyone more knowledgeable of the matter than I tell me why "memory hard" excludes ASICs?

The main thing (most) memory hard algorithms achieve is effectively crippling your ability to parallelize, which has been the main strength of GPUs and ASICs. Yes, the algorithm itself isn't cheap computationally either, but if computing a single round of it essentially requires most or all of your available memory, then you can't just throw a bunch more compute cores at it without correspondingly adding the same amount of extra memory for each additional core, which makes it much less economical to abuse (i.e. imagine if, for an extreme case of a memory hard problem, for your GPU with ~500 compute units, you also needed several GB of mem to dedicate to each unit in order to do the equivalent of brute forcing hashes today. Such a card would obviously be impractical, which is what memory hard problems try to achieve). IOW, it's not so much about the bandwidth of the memory controller as it is about the amount of memory you'd need for the ASIC to be worthwhile, and no form of memory is getting cheaper at present.

1816 posts | registered 10/10/2003
adr_ Wise, Aged Ars Veteran

jump to post

Rene Gollent wrote:
show nested quotes
Can anyone more knowledgeable of the matter than I tell me why "memory hard" excludes ASICs?

The main thing (most) memory hard algorithms achieve is effectively crippling your ability to parallelize, which has been the main strength of GPUs and ASICs. Yes, the algorithm itself isn't cheap computationally either, but if computing a single round of it essentially requires most or all of your available memory, then you can't just throw a bunch more compute cores at it without correspondingly adding the same amount of extra memory for each additional core, which makes it much less economical to abuse (i.e. imagine if, for an extreme case of a memory hard problem, for your GPU with ~500 compute units, you also needed several GB of mem to dedicate to each unit in order to do the equivalent of brute forcing hashes today. Such a card would obviously be impractical, which is what memory hard problems try to achieve). IOW, it's not so much about the bandwidth of the memory controller as it is about the amount of memory you'd need for the ASIC to be worthwhile, and no form of memory is getting cheaper at present.

You wouldn't see as big of a speed up as you do with compute-bound algorithms, but you would probably see a speed up on both the performance/power and performance/cost fronts. The trick is it has to have wide enough perf improvement/market demand to justify making ASICs, which is part of why a lot of multi-coin ASICs you see now are a bunch of high performance DSP blocks wired together with a microcontroller core (previous antminer chips do this). For memory-hard problems like ethash, I'm guessing this means they designed a high performance, very wide memory controller to pair with their existing asic miner IP and a highly optimized ethash algorithm.

An important point where you're wrong is that ethash does not permute the data read, so it's shareable among any number of cores, so long as your memory bandwidth can keep up.

107 posts | registered 3/24/2016
Abhi Beckert Ars Tribunus Angusticlavius

jump to post

fuzzyfuzzyfungus wrote:
Can anyone more knowledgeable of the matter than I tell me why "memory hard" excludes ASICs?

They don't just require lots of memory, they require fast memory.

For example, the memory in a modern GPU operates at 224GB/s or so (some cards are faster). For comparison, the best Intel CPUs can only access RAM at around 60GB/s and an ASIC will be much slower.

To get good performance a CPU will avoid RAM altogether and stick to internal cache memory for "memory hard" hashes, which is fine if you just want to run one or two hashes (they only have a few megabytes of cache on the CPU - and it has to be shared with other processes, especially your OS kernel).

For actual mining you want to be doing hundreds or thousands of hashes simultaneously. That requires either thousands of high end Intel CPUs or several GPUs.

And since a cheap ASIC will not have any high speed memory they basically can't do memory hard hashes at all (or at least, they can only do it slowly and burning a ton of power).

Creating RAM that runs at hundreds of gigabytes per second requires a state of the art fabrication and even then there is a high failure rate of chips fabricated that end up being unusable. A chip fabricated at 8GB will be tested and maybe only 2GB of it is actually functional (it can often be sold to customers as a 2GB chip).

As technology improves, the hashing algorithm can be modified to require even more memory.

6230 posts | registered 9/27/2005

Promoted Comments

Rene Gollent Ars Tribunus Militum et Subscriptor

jump to post

fuzzyfuzzyfungus wrote:
Can anyone more knowledgeable of the matter than I tell me why "memory hard" excludes ASICs?

The main thing (most) memory hard algorithms achieve is effectively crippling your ability to parallelize, which has been the main strength of GPUs and ASICs. Yes, the algorithm itself isn't cheap computationally either, but if computing a single round of it essentially requires most or all of your available memory, then you can't just throw a bunch more compute cores at it without correspondingly adding the same amount of extra memory for each additional core, which makes it much less economical to abuse (i.e. imagine if, for an extreme case of a memory hard problem, for your GPU with ~500 compute units, you also needed several GB of mem to dedicate to each unit in order to do the equivalent of brute forcing hashes today. Such a card would obviously be impractical, which is what memory hard problems try to achieve). IOW, it's not so much about the bandwidth of the memory controller as it is about the amount of memory you'd need for the ASIC to be worthwhile, and no form of memory is getting cheaper at present.

1816 posts | registered 10/10/2003
adr_ Wise, Aged Ars Veteran

jump to post

Rene Gollent wrote:
show nested quotes
Can anyone more knowledgeable of the matter than I tell me why "memory hard" excludes ASICs?

The main thing (most) memory hard algorithms achieve is effectively crippling your ability to parallelize, which has been the main strength of GPUs and ASICs. Yes, the algorithm itself isn't cheap computationally either, but if computing a single round of it essentially requires most or all of your available memory, then you can't just throw a bunch more compute cores at it without correspondingly adding the same amount of extra memory for each additional core, which makes it much less economical to abuse (i.e. imagine if, for an extreme case of a memory hard problem, for your GPU with ~500 compute units, you also needed several GB of mem to dedicate to each unit in order to do the equivalent of brute forcing hashes today. Such a card would obviously be impractical, which is what memory hard problems try to achieve). IOW, it's not so much about the bandwidth of the memory controller as it is about the amount of memory you'd need for the ASIC to be worthwhile, and no form of memory is getting cheaper at present.

You wouldn't see as big of a speed up as you do with compute-bound algorithms, but you would probably see a speed up on both the performance/power and performance/cost fronts. The trick is it has to have wide enough perf improvement/market demand to justify making ASICs, which is part of why a lot of multi-coin ASICs you see now are a bunch of high performance DSP blocks wired together with a microcontroller core (previous antminer chips do this). For memory-hard problems like ethash, I'm guessing this means they designed a high performance, very wide memory controller to pair with their existing asic miner IP and a highly optimized ethash algorithm.

An important point where you're wrong is that ethash does not permute the data read, so it's shareable among any number of cores, so long as your memory bandwidth can keep up.

107 posts | registered 3/24/2016
Abhi Beckert Ars Tribunus Angusticlavius

jump to post

fuzzyfuzzyfungus wrote:
Can anyone more knowledgeable of the matter than I tell me why "memory hard" excludes ASICs?

They don't just require lots of memory, they require fast memory.

For example, the memory in a modern GPU operates at 224GB/s or so (some cards are faster). For comparison, the best Intel CPUs can only access RAM at around 60GB/s and an ASIC will be much slower.

To get good performance a CPU will avoid RAM altogether and stick to internal cache memory for "memory hard" hashes, which is fine if you just want to run one or two hashes (they only have a few megabytes of cache on the CPU - and it has to be shared with other processes, especially your OS kernel).

For actual mining you want to be doing hundreds or thousands of hashes simultaneously. That requires either thousands of high end Intel CPUs or several GPUs.

And since a cheap ASIC will not have any high speed memory they basically can't do memory hard hashes at all (or at least, they can only do it slowly and burning a ton of power).

Creating RAM that runs at hundreds of gigabytes per second requires a state of the art fabrication and even then there is a high failure rate of chips fabricated that end up being unusable. A chip fabricated at 8GB will be tested and maybe only 2GB of it is actually functional (it can often be sold to customers as a 2GB chip).

As technology improves, the hashing algorithm can be modified to require even more memory.

6230 posts | registered 9/27/2005

Timothy B. Lee Timothy is a senior reporter covering tech policy and the future of transportation. He lives in Washington DC.

Channel Ars Technica

← Previous story Next story →

Boom and bust —

Why this year’s insane graphics-card price surge might be over

You can finally buy a high-end graphics card at Best Buy again.

Further Reading

Two reasons the GPU price slump is likely to continue

Further Reading

Promoted Comments

Promoted Comments

Channel Ars Technica

Further Reading

Two reasons the GPU price slump is likely to continue

Further Reading

Promoted Comments

Promoted Comments

reader comments

Channel Ars Technica