Discussion Techpowerup GPU database?

[Order 66]

I always enjoy looking at the Techpowerup GPU database, for interesting or unreleased GPUs. Recently, I've come across a few GPUs that have me a bit confused.

1. RTX 4500 ADA generation
https://www.techpowerup.com/gpu-specs/rtx-4500-ada-generation.c4172
I don't understand how this card can have a 130W TDP, but yet not require any external power connectors.

2. RTX 3050 4GB
https://www.techpowerup.com/gpu-specs/geforce-rtx-3050-4-gb.c3744
I understand there is a laptop RTX 3050 with 4GB of VRAM, but this is different. Also, I cannot find these for sale anywhere.

3. GTX 1080 ti 10 and 12GB
https://www.techpowerup.com/gpu-specs/geforce-gtx-1080-ti-10-gb.c3309

NVIDIA GeForce GTX 1080 Ti 12 GB Specs

NVIDIA GP102, 1670 MHz, 3200 Cores, 200 TMUs, 80 ROPs, 12288 MB GDDR5X, 1376 MHz, 384 bit

www.techpowerup.com

I realize these cards were never released, but why were they made in the first place?

4. RX 590 GME
https://www.techpowerup.com/gpu-specs/radeon-rx-590-gme.c3505
why does this card exist, and why was it launched as late as 2020?

 

[hotaru.hino]

1. RTX 4500 ADA generation
https://www.techpowerup.com/gpu-specs/rtx-4500-ada-generation.c4172
I don't understand how this card can have a 130W TDP, but yet not require any external power connectors.

Likely an error, because unless there's another version of this card, it's TBP is 210W (https://resources.nvidia.com/en-us-...idia-rtx-450?lx=CCKW39&contentType=data-sheet)

2. RTX 3050 4GB
https://www.techpowerup.com/gpu-specs/geforce-rtx-3050-4-gb.c3744
I understand there is a laptop RTX 3050 with 4GB of VRAM, but this is different. Also, I cannot find these for sale anywhere.

Either it was never released (despite what it says, because the release date corresponds to the 8GB variant) or it had a limited run in certain markets.

3. GTX 1080 ti 10 and 12GB
https://www.techpowerup.com/gpu-specs/geforce-gtx-1080-ti-10-gb.c3309

NVIDIA GeForce GTX 1080 Ti 12 GB Specs

NVIDIA GP102, 1670 MHz, 3200 Cores, 200 TMUs, 80 ROPs, 12288 MB GDDR5X, 1376 MHz, 384 bit

www.techpowerup.com

I realize these cards were never released, but why were they made in the first place?

Probably to recoup the cost of GPUs that didn't have enough memory channels. NVIDIA does this a lot. They probably got recycled elsewhere if not tossed.

4. RX 590 GME
https://www.techpowerup.com/gpu-specs/radeon-rx-590-gme.c3505
why does this card exist, and why was it launched as late as 2020?

Probably to throw the Chinese market a bone.

 

[Order 66]

Probably to recoup the cost of GPUs that didn't have enough memory channels. NVIDIA does this a lot. They probably got recycled elsewhere if not tossed.

Are there any of these out in the wild? I also know there is a GTX 1060 5GB that is out in the wild, which AFAIK it was never released either.

Also what exactly are these cards used for?
1. https://www.techpowerup.com/gpu-specs/rtx-1000-mobile-ada-generation.c4208
2. https://www.techpowerup.com/gpu-specs/rtx-5880-ada-generation.c4191
3. https://www.techpowerup.com/gpu-specs/rtx-2000-ada-generation.c4199
4. https://www.techpowerup.com/gpu-specs/rtx-500-mobile-ada-generation.c4207

 

[Order 66]

Likely an error, because unless there's another version of this card, it's TBP is 210W (https://resources.nvidia.com/en-us-...idia-rtx-450?lx=CCKW39&contentType=data-sheet)

maybe I'm missing something, but I don't see any reference to TBP in the link you provided.

 

[Eximo]

1060 5GB was released only through OEMs if I recall. I recall seeing some Dell's with it.

There are two RTX 4500.
RTX4500 Ada and RTX A4500 (A is for Ampere, Ada is for Ada) The older card is 200W, and the new 210W.

 

[Colif]

its not always accurate, it says my GPU needs 3 x 8 pin PCIe cables... I must have a mutant as it only needs two.
They also show the stats for the wrong model for my card. They have stats for the Hellhound when I have Red Devil. Easy enoughto check the Powercolor site to confirm.

I did try to correct it but I just gave up.

its probably because no one ever tested my card so there is no stats to check. If Powercolor had sent it out instead of Hellhound, it might be right

 

[hotaru.hino]

maybe I'm missing something, but I don't see any reference to TBP in the link you provided.

TBP = Total Board Power.

 

[Order 66]

TBP = Total Board Power.

I know. I don’t understand your point.

 

[hotaru.hino]

I know. I don’t understand your point.

That's from the product literature I linked.

 

[Order 66]

That's from the product literature I linked.

So sorry, I must’ve missed that. Wait, 24GB on 192 bit bus?! What a joke! Isn’t this supposed to be a professional card?

 

[hotaru.hino]

So sorry, I must’ve missed that. Wait, 24GB on 192 bit bus?! What a joke! Isn’t this supposed to be a professional card?

I wish people would stop seeing how wide the VRAM bus is and make an assessment on that alone. You need the transfer rate to get the bandwidth, which I'm sure is what people are actually trying to poke at. Its memory is clocked at 18GT/sec so total bandwidth is 432 GB/sec. It's less than the RTX 4070's, but it's likely clocked slower for reliability reasons. Also how much bandwidth you need really depends on the workload.

If anything, this just tells me the card has 4GB GDDR6 chips, which I still thought wasn't a thing yet.

 

[Order 66]

I wish people would stop seeing how wide the VRAM bus is and make an assessment on that alone.

Sure, but why? The RTX 4060ti 16GB has 128bit bus, and that (at least in part) holds it back.

If anything, this just tells me the card has 4GB GDDR6 chips, which I still thought wasn't a thing yet.

Neither did I.

 

[hotaru.hino]

Sure, but why? The RTX 4060ti 16GB has 128bit bus, and that (at least in part) holds it back.

This is a naive assessment of how wide the memory interface is means anything to performance:

• Again, you didn't mention transfers per second here.
  • While yes, there's a limit on how fast the chips can be clocked, every video card in a series doesn't clock the memory interface at the same rate.
  • AD107 is clocked at 16GT/s. AD106 at 17GT/s. AD104 at 21GT/s. AD103 at either 22.4 or 23GT/s. And AD102 is at 21GT/s
  • Considering that AD103 is clocked faster than the rest of the Ada Lovelace lineup, are we going to argue that the RTX 4090 is starved because it's clocked slower? No, because conversely, transfer rate means nothing without how wide the memory bus is. The RTX 4090 still has at least 25% more bandwidth than the RTX 4080.
• Ada Lovelace GPUs have at least 12 times the amount of L2 cache than Ampere. More cache can act as a buffer against slower memory bandwidth. That's mostly why cache was invented in the first place.
• What operations are held back by lower memory bandwidth? Like specificoperations and/or scenarios?
  • While I'm aware that overclocking memory improves performance, this seems to be a pattern among any GPU. The thing I'm looking for is actual concrete evidence from GPU profiling as to why this is. I can make plenty of guesses, but that only gets you so far.

Basically assessing performance based on memory bus width is the same as assessing CPU performance based on clock speed alone.