GPUs are headed in the wrong direction
theverge.com
theverge.com
This is a strange comparison because the 3080 was essentially never (and still is not) available at the MSRP of 699 $ (719 €). Even now the cheapest 3080 I can find is almost 800 €, and that's with the new generation already out two years later. Before the crypto collapse you'd never see these under four digits.
However the overall trend is still correct - nVidia has kept the price for 1 TFLOPS very constant since 2016ish
2004: ~5000 $/TFLOPS for a 6800 Ultra (~0.1 TFLOPS, ~500 $)
2006: 1200 $/TFLOPS for a 8800 GTS (~0.4 TFLOPS, ~500 $)
2008: 500 $/TFLOPS for a 280 (1 TFLOPS, 500 $)
2010: 300 $/TFLOPS for a 580 (1.6 TFLOPS, 500 $)
2012: 160 $/TFLOPS for a 680 (3 TFLOPS, 500 $)
2014: 115 $/TFLOPS for a 980 Ti (5.6 TFLOPS, 649 $)
2016: 60 $/TFLOPS for a 1080 Ti (11.4 TFLOPS, 699 $)
2018: 51 $/TFLOPS for a 2080 Ti (23.5 TFLOPS, 1199 $)
2020: 48-76 $/TFLOPS for a 3080 (25 TLFOPS, 1200-1900 $) or around 50-60 $/TFLOP for a 3080 Ti (28 TFLOPS, ~1600 $)
2022: 40 $/TFLOPS for a 4080 (43 TFLOPS, seems to be in the 1600 $ region now)
If the price per Tflop and the power consumption per Tflop do not decrease in a new generation of GPUs, then there is no reason for a new generation to exist.
The power consumption per Tflop has continued to decrease reasonably well, but the price has almost stagnated, for FP32. On the other hand for FP64 computations, the price per Tflop has greatly increased during the recent years, due to the fact that both NVIDIA and AMD have restricted this feature to their "datacenter" GPUs, which have prices that are acceptable only for those who buy them in very large numbers and for whom the costs for the expensive GPUs are compensated by savings in rack space, power consumption and cooling.
It seems that AMD has removed completely the FP64 execution units in RDNA 3, while NVIDIA provides only a 1:64 throughput ratio for FP64:FP32 in RTX 4000, so they should have better also removed completely the FP64 hardware, because a software implementation would be faster than that.
Now a CPU like Ryzen 9 7950X has a lower price per FP64 Tflop than any GPU, which is opposite from a few years ago, when e.g. a Radeon VII had a much lower price per FP64 Tflop than any CPU.
So the RTX 4080 TFLOPs are higher, but require you to be more cache friendly then the previous gen.
Hot fucking take. Another quality article here from The Verge's Sean Hollister.
Do they expect the same performance for half the price? And measuring it by volume in Cubic inches?
While The verge hasn't earned a spot in my RSS feeds, I have always thought their reporting were of high quality. This Editorial, however shows a lack of basic understanding how GPU performance works, sales, TDP, production cost, R&D, Fabs, and AIB works.
If you're of an age where you remember Moore's law being in effect, that expectation wasn't always unrealistic.
Instead Nvidia is offering more flops at much higher prices, with much lower usability. So sure a 4080 has more flops, but less memory bandwidth than a 3080 at 1.7x the price, and that's after the lower end 4080 was cancelled.
(RTX 4080 vs. RTX 4090)
(Moreover, the cost of big chips grows non-linearly, so the expected price ratio is actually less than the size ratio, so even if one would accept $1600 as a price for RTX 4090, RTX 4080 should be at most $800).
"I predict that within 10 years, computers will be twice as powerful, ten thousand times larger, and so expensive that only the 5 richest kings of Europe will own them"
(The joke being that it was a crazy prediction to think computers would become larger and more expensive)
Personally I'm going to stay a couple generations back until prices calm down, even if that means playing games from my backlog instead of the newest cutting-edge ones.