RTX 4090/4080 spec /price

[firejuggler]

Hi.
https://www.tomshardware.com/news/nv...specs-revealed


basically 4090/3090 TI
Transistors 76b/28.3b
Streaming multiprocessor 128/84
GPU core 16384/10752
TFlops FP32 82.6/40

edit, found https://www.techpowerup.com/gpu-spec...rtx-4090.c3889
FP64 is 1:64 FP32 speed, so that mean 1.290 TFlops against 0.556 TFlops fpr the 3090.

[Xyzzy]

Note that the 12GB "4080" is really just a 4070.

[James Heinrich]

I have added entries for the 40[89]0 but of course no benchmarks yet.
https://www.mersenne.ca/mfaktc.php?filter=rtx+4

[M344587487]

I know it's just a number but not so long ago breaking the billion transistor barrier was impressive. Doubt we're ever going to hit a trillion transistors on a monolithic die, 100 billion may happen but not much beyond that as everyone transitions to MCM.



Rumour has it that the 40 series is going to cost a lot for a long time as there is a backlog of stock of 30 series. And mining should be crashing as Ethereum is now PoS so 2nd hand cards will keep them in stock for a while. And epic recession on the horizon etc etc.


4nm TSMC is tasty, it's a shame it's an nvidia card with questionable design choices like 1:64 and extreme power consumption. Hopefully low end 40 series underclocked and undervolted at reasonable prices end up existing and perform well.

[LordJulius]

Originally Posted by M344587487

Rumour has it that the 40 series is going to cost a lot for a long time as there is a backlog of stock of 30 series. And mining should be crashing as Ethereum is now PoS so 2nd hand cards will keep them in stock for a while. And epic recession on the horizon etc etc.

https://arstechnica.com/gaming/2022/...of-moores-law/

"When asked about those price increases, Nvidia CEO Jensen Huang told the gathered press to, in effect, get used to it. "Moore's law is dead," Huang said during a Q&A, as reported by Digital Trends. "A 12-inch wafer is a lot more expensive today. The idea that the chip is going to go down in price is a story of the past."
...
"Generational price comparisons aside, Huang's blanket assertion that "Moore's law is dead" is a bit shocking for a company whose bread and butter has been releasing graphics cards that roughly double in comparable processing power every year. But the prediction is far from a new one, either for Huang—who said the same thing in 2019 and 2017—or for the wider industry—the International Technology Roadmap for Semiconductors formally announced it would stop chasing the benchmark in its 2016 roadmap for chip development."


I have no problem switching to the red side for cards...

[Prime95]

Originally Posted by LordJulius

I have no problem switching to the red side for cards...

I've been watching Radeon VII prices on eBay since the Etherum PoS change. I've seen Buy-it-now prices around $400. One miner was offering 4 cards for $1000 today.

I should have my head examined for even considering another card. Though I do wonder if one could increase throughput noticeably running 5 cards at lower power vs. 4 cards at higher power (same total power consumed).

[kriesel]

Originally Posted by Prime95

I do wonder if one could increase throughput noticeably running 5 cards at lower power vs. 4 cards at higher power (same total power consumed).

I run mine at -20% from nominal power on Windows which only costs ~-7% throughput. I've seen an occasional card go into an odd mode sometimes in Windows on a multi-RadeonVII system, where 1 card will produce ~half the speed, at ~one third the power and gpu clock rate, often preceded by gpu->host repeated error, and the AMD software can't change the GPU clock until the card is disabled and enabled again in device manager. So if one could force them to ~700MHz GPU clock, and leave ram clock at ~11xx MHz, more could be run on the same power budget and produce more aggregate throughput. That gets more appealing as the unit cost drops, and utility rates go up.

[kriesel]

For more on that, see https://mersenneforum.org/showpost.p...82&postcount=1
https://mersenneforum.org/showpost.p...postcount=2783

[M344587487]

Originally Posted by LordJulius

https://arstechnica.com/gaming/2022/...of-moores-law/
"When asked about those price increases, Nvidia CEO Jensen Huang told the gathered press to, in effect, get used to it. "Moore's law is dead," Huang said during a Q&A, as reported by Digital Trends. "A 12-inch wafer is a lot more expensive today. The idea that the chip is going to go down in price is a story of the past."
...

Nice political wording from Jensen, the chip is a small piece of the product and most people don't need the 76 billion transistor version (which I assume has horrible yields which also doesn't help the chip price). MCM is making more sense every day.

[henryzz]

MCM is inevitable and has been for some time. They just needed the cost/reward to hit a critical point. We appear to be reaching that point.
It was dabbled with many years ago with Core 2 Quads which were dual duos in reality.

[M344587487]

rdna3 high end should use at least two compute dies, seems similar to Zen 1's toe-dipping. In the next few years intel is going nuts splitting up their CPU designs into modular components which might make its way to GPU, it'll at least make it to iGPU and server farms. Nvidia might not go MCM until the 60 series, they tend not to use unproven designs and given their position in the market they can probably get away with it.