NVIDIA GeForce RTX 3080 Offers Up Incredible Linux GPU Compute Performance

[mdedetrich]

They didn't get capacity because Apple and AMD were given first priority due to both having long standing status and the fact Apple uses AMD's cards in their systems, it made more sense for them not to lose the Apple contract which is the largest in the industry and will balloon even more so with Apple Silicon fab orders.

There were leaks (and there is no way to confirm this otherwise) that it wasn't just a simple case of TSMC "not having capacity" (NVidia can be a client of TSMC just like AMD or Apple is if they pay enough). The leaks implied that Nvidia was trying to push NVidia by being aggressive into selling their chips at a lower price than they were happy with by blackmailing them into losing a deal that would give a lot of capacity to TSMC; plus the mentioning the less than fruitful history between TSMC and NVidia.

This of course backfired greatly since TSMC already has plenty of other customers happy to pay higher for their premium die, so TSMC just said f'off to NVidia. Its true that TSMC has limited capacity but just like with any standard business, NVidia would have been talking with TSMC for years before manufacturing their 3000 series.

If this is true than NVidia shot themselves on the foot anyways, they may have saved money with Samsung's more inferior die (which apparently is sold with razer thin profit margins to NVidia because no one outside of Samsung really wants the die) but because the Samsung die use a lot more power NVidia had to spend a lot more money on cooling solutions for the card.

The profit margins on the FE editions are much smaller than what NVidia likes and I doubt that AIB's are happy with having to also spend more money on cooling the 3000 series (I doubt AIB's are happy in general with the crappy launch, i.e. they didn't get drivers till launch day which contributed to the crashing situation with capacitors, don't get started about the stock issues).

Also note that GDDR6X is also using more power than standard GDDR6 (and also runs quite hot, ~100 degrees even though its in spec) so this is also contributing however the situation now is that you have a double whammy.

[mdedetrich]

Not buying another Nvidia hardware in future. Very disappointed since they blocked Nouveau development with signed firmaware.

This was unintended, NVidia didn't sign the drivers because of Linux but because a lot of chinese fake cards was being sold with modified firmware during crypto boom.

Unfortunately Linux was kind of collateral damage here.

[GruenSein]

TFlops are TFlops. They're a bit like potential energy: useful if you put it to use, useless otherwise. But it's still there in both cases.

Unfortunately, not every algorithm has linear time complexity.

Hardly any if these results and Nvidia's numbers are to be believed. What I was actually getting at is that the miraculous doubling of CUDA cores Nvidia has been bragging about is not in fact a doubling at all. Rather, their already existing integer data path can be used for floating point operations under certain conditions. Many other things within the core are not duplicated at all. It's like claiming Bulldozer had eight cores. Nonetheless, this might be a clever and efficient use of silicon. The peak throughput numbers that are floating around are even more theoretical than usual TFlop claims. This time around, they are just bogus.

[WalterCool]

This was unintended, NVidia didn't sign the drivers because of Linux but because a lot of chinese fake cards was being sold with modified firmware during crypto boom.

Unfortunately Linux was kind of collateral damage here.

Ohhh, didn't knew about reason, but yes, agree on piracy do severe damage tech industry.

If that's the case, they really extinguished the fire with gasoline, because Nvidia videocards have now a limited lifespan depending on Nvidia... or stay on very basic 2D/3D rendering by nouveau (even slower than Intel Graphics)

[skeevy420]

I've also considered undervolting, but not really sure if it can be done with Linux AMD/Linux drivers. Possibly with AMDGPU? There are these rather recent 5300 / 5500 cards.

It can. I run my RX 580 with a slight undervolt set by a systemd unit because it allows me to lower the temps while not losing any performance.

[leipero]

Much more likely is that NVidia was trying to get TSMC to manufacture the 7nm die for their 3000 series GPU's but because they were so bullish/aggressive they didn't get any capacity from TSMC (also to note that NVidia doesn't have the best history with TSMC, they used them in the past and didn't get the best results). So instead they had to settle with inferior 8nm node from Samsung (which tbh is probably closer to a 10nm node).

So the only way that NVidia could get such performance in their cards was by feeding them more power, its the same reason the cards are so hard to overclock. NVidia is basically doing the same thing as Intel, only difference is that NVidia unlike Intel doesn't have their own fab.

To be fair, none of those fabrication processes are even close to the 14nm, let alone 8/7 or whatever, it's all marketing BS. It depends what you measure, how you measure, and what type of transistors you use.
Long story short = there's a lot of space for imporvement actually for years to come, and their numbers are sort of "pulled out of their sittng places" (in a way at least).

[mdedetrich]

To be fair, none of those fabrication processes are even close to the 14nm, let alone 8/7 or whatever, it's all marketing BS. It depends what you measure, how you measure, and what type of transistors you use.
Long story short = there's a lot of space for imporvement actually for years to come, and their numbers are sort of "pulled out of their sittng places" (in a way at least).

There are objective ways to measure it, its just that each fab company has their own metric of measurement (Linus did a great video on this).

In any case, right now TSMC has the best silicon you can get and Samsungs 8nm is no way close to TSMC's 7nm

[leipero]

mdedetrich Not saying there's no objective way to measure, I'm just saying those fabrication processes are nowhere near what they claim. But I agree with you about TSMC.

[freedonuts]

You'd say with this performance it should be no problem running GRUB @1080p. I guess not...