Tesla GPU Accelerator Bang for the Buck, Kepler to Volta | Hacker News Reader

Tesla GPU Accelerator Bang for the Buck, Kepler to Volta

nextplatform.com

78 pointsjonbaer27 comments

> thanks in part to the growing demand of Nvidia’s Tesla cards in those markets but also because cryptocurrency miners who can’t afford to etch their own ASICs are creating a huge demand for the company’s top-end GPUs.

This part is incorrect. Unless you need functionality specific to workstation & server GPUs, it can be an order of magnitude cheaper to buy a high end gaming GPU. Cryptominers normally use motherboards with 12+ PCI-E 1x-16x adapters and consumer cards: https://www.youtube.com/watch?v=B6gB7GYkkiA

The same performance/dollar math applies to anyone else who needs to do GPU processing and doesn't need:

* double-precision math

* high-bandwidth connection between the GPU and motherboard

* a setup that doesn't violate EULAs

* a warranty

In my testing with raytracers, there is around a 5-15x improvement in perf/$ with consumer cards.

In data centers, nvidia seems to want to limit usage of consumer products [1].

[1] https://news.ycombinator.com/item?id=15983587

>* a setup that doesn't violate EULAs

This was a reference to what you're talking about.

Yes. but I am not sure if Nvidia is pushing it yet. But at least in Linux, you have to apply some patches to enable passthrough for consumer GPUs.

current gaming GPUs are vastly over priced from normal because many are buying them for mining. it is one of the few times where buying a prebuilt PC is actually a better deal.

I do believe Nvidia does offer to sell at MSRP with limits by household yet even then those are not the cards many home builders prefer

This has got me thinking recently, even the second hand market is really strong for 5-year-old mid-level GPU's. Not the sort that are useful for mining either, just the knock on impact. Also it seems to me that a 5-year-old mid-range card is more than up to playing popular games like WOW or Overwatch. I wonder if there is an opportunity for either

a) AMD to offer something mid-market for a low price and steal a load of market share back from Nvidia, while they concentrate on the $900 plus market. I suppose AMD's CPU's with onboard graphics are sort of like this.

b) A new entrant?

There is also less risk associated with purchasing gaming GPUs for mining, as some cost can generally be recovered in the event of a crash through resale. Especially now, where cards like the 1080TI are going for, what, double price, if you can find them.

Meanwhile an ASIC is useless outside of mining.

This is repeated a lot, but I don't know many that are interested in buying a used card these days due to mining and wear on them. I don't think the resale value is nearly as great as is implied. I could be wrong though. If we get to the point where resale starts to happen in mass, it'll be an interesting item to watch.

A returned 1080TI was the only option for me. Everything else was taken. So far so good. I was the ONLY one at Fry's NOT buying a graphics card for mining (deep learning for me). 4 other people were there to get cards for mining. Apparently they didn't want the 1080TI though because it was not as power efficient for mining compared to the 1070.

> Apparently they didn't want the 1080TI though because it was not as power efficient for mining compared to the 1070

I'm sure there's still a demand for them, because it's still possible to turn a profit over the lifetime of the card; it's just not as high as that of the 1070.

Are the GPU makers ramping up production to meet demand? I feel like this shortage had been going on for, what, at least 6 months?

One thing I don't fully understand is why these accelerators all have relatively small memory sizes? Is it purely because of cost?

The most expensive Tesla accelerator I see in this list has 16 GB, which sounds like a lot, but it's 'only' twice as much as the consumer-level GPU I have in my PC at home, and it seems like a pretty paltry amount for large-scale simulations. I'm thinking about volumetric modeling applications, for example, where storing a dense voxel grid as a volume texture would easily run over 30 GB even for quite modest grid size/resolution (e.g. 2K x 2K x 2K grid). It's exactly these kinds of things where a GPU accelerator could shine, without the need to implement spare sampling solutions or accelerator data structures that would be complex and severely impact performance on GPU architectures.

If I understand it correctly, GPUs use specialized high-bandwidth memory that is more expensive instead of commodity memory. See https://en.wikipedia.org/wiki/GDDR5_SDRAM.

HBM2 to be specific: https://en.wikipedia.org/wiki/High_Bandwidth_Memory#HBM_2

> that would be complex and severely impact performance on GPU architectures

Agree about the complexity. However, adaptive grids can be a huge win in terms of performance. The GPU doesn’t need to simulate the whole 8G grid nodes, it only simulates much fewer nodes, only in the areas of large gradients / higher frequencies of the simulated fields.

Here’s an old SpaceX presentation about their GPU simulations of rockets: https://www.youtube.com/watch?v=vYA0f6R5KAI

That was very interesting and helpful, thanks!

It's not just the cost, at least not just the cost of the RAM. GPU ram is designed to be wider than desktop memory, meaning that you need more pins on the die for the same amount of memory. The IO in top end accelerators uses all the real estate available for pins, and the dies are already at the reticle limit, so to add memory you have to either:

1) Use desktop memory that allows more BW/pin but is also much slower per pin

2) Develop completely new memory just for your application

3) Put some kind of buffering between the chip and the ram, greatly increasing latency.

None of these is currently that appealing for the manufacturers. HMC was originally going to do 3 by doing 2, but HMC appears stillborn, as all the momentum went to HBM instead, and HBM cannot increase the memory count per pin nearly as much as HMC could.

NVIDIA makes a chip called the P6000 that offers 24GB memory. It isn't high-throughput (HBM2) memory but regardless, the chip will benchmark relatively close to the Volta V100 and costs about half. Here's a deep learning benchmark for comparison: https://github.com/rejunity/cnn-benchmarks

I was shocked how difficult it was to get a 1080TI at the local store.

Pardon my ignorance but with all the demand for graphics cards for cryptocurrency mining, why has a viable ASIC not become commonplace so miners aren't grabbing up every available graphics card they can find making it hard for non-miners to have anything and driving up prices.

I'm hoping that mining moves more to ASIC and deep learning gets more desktop based "tensor" options. This is a very strange market right now with companies trying to appeal to everyone instead of specializing more and not really serving anyone well.

If the coin you're mining goes bust you can still sell the GPU and recoup some of the investment. With ASICs there's not much of a market for that.

Also, most coins try to design them to be ASIC resistant because they do not want them to become centralized.

I've heard this elsewhere but it wasn't clear to me what that being ASIC resistant actually entails

The algorithms used for Ether and some other altcoins require more RAM than is economically available for the current ASIC implementations. This was specifically done to deter use of ASICs.

From what I have seen on HN, the algorithms are very memory heavy. Lots of relevant discussion in this thread https://news.ycombinator.com/item?id=16378417

Pro tip: google "hacker news" + whatever your interested in to find lots of high quality comments about a topic on old articles. That is how I found the previous link and what I do when looking into new technologies.

I think ASICs are still winning for bitcoin itself, but they currently have two big problems for a lot of buyers.

First, absolutely zero resale value. Once either the difficulty raises, or the value of the coin falls, to the point where you're spending more on power than you're generating in currency, your ASIC goes in the trash.

Second, the algorithm they use is baked in. So an ASIC built for bitcoin can only be used for other sha256-based currencies. With a GPU, when you realise you're not cost-effective for BTC, you can try your hand at Eth instead.

Many cryptocurrencies are designed to be efficiently mined on GPUs, with the idea that mining would be more decentralized. This may be effective to an extent, but in the case of rapidly rising prices like we're seeing now / over the past months, there are some side effects.

Furthermore, in the 'alt' space (i.e. not-Bitcoin), there is a lot of competition and you don't want to tie up large capital expenses to inflexible hardware. GPUs can mine whatever trendy coin is currently pumping. Miners switch to mining different coins in order to maximize return.

Can the demand for bitcoin ever be sated?

Given the CAPEX cost of the GPUs, and the OPEX of the electricity someone has probably worked out the rates of return for each of the crypto currencies.

If there’s an expectation the virtual currencies will soar in value why wouldn’t people buy all they can, perhaps to two or three times the cost to mine.

Reply on news.ycombinator.com