MSI GeForce RTX 5070 Ti Ventus 3X OC

[N/A]

Each SMX has significant amounts of SRAM that your calculation is neglecting: 128 KB of L1 and 256 KB of registers in each Ada SM. Therefore, the SMX won't scale as well as you think.

Great so the GB203 has 64MB L2 and 31,5MB L1 plus registers for a total of 78 mm² cache.

 

[AnotherReader]

Great so the GB203 has 64MB L2 and 31,5MB L1 plus registers for a total of 78 mm² cache.

Each SMX also has an instruction cache of unspecified size. Then there is the register cache and the RT cores probably have caches of their own.

 

[oxrufiioxo]

Each SMX also has an instruction cache of unspecified size. Then there is the register cache and the RT cores probably have caches of their own.

That's why you need so much cash just to buy one....

 

[wolf]

Some of the arrogance on display in this thread is astounding, I had no idea these people were so much smarter and better informed than W1zzard and the 95% of GPU markets customers and even GPU designers and game devs. Just be careful on your moral high horses, falling from such make-believe heights usually bruises ego's pretty hard.

Some launch stock here in Aus was available for MSRP, but I suspect like has happened to the 5080 and 5090 both here and abroad, it was pricing for the launch only, at least till the market is saturated with card availability. If you wanted any of these cards getting in right at launch seems to be the best play. I eagerly await AMD's response, what they get right and invariably, what they get wrong.

 

[N/A]

CBse reported 44.5 dBA, complete madness. But some reviews aren't that bad. Essentially most everybody got sent a ventus.

 

[tussinman]

I remember back in December a handful of users were hyping this card and how it's going to be a beast. Even comparing the 3070ti and 4070ti to the previous gens high-end/flagship.

I didn't see the hype even at the time. The on-paper spec leaks weren't that impressive, and reality is this is the 3rd generation on the same node so you can only push it so far.

 

[Scattergrunt]

I remember back in December a handful of users were hyping this card and how it's going to be a beast. Even comparing the 3070ti and 4070ti to the previous gens high-end/flagship.

I didn't see the hype even at the time. The on-paper spec leaks weren't that impressive, and reality is this is the 3rd generation on the same node so you can only push it so far.

If you could actually find it at MSRP then I wouldn't be bothered by it because at least its cheaper than a 4080 and 4080S on paper but like.. no. It just isnt and probably wont be for awhile. Its only runs ideally 20w better vs a 4080/S as is, its just so.. mild. Thats just my opinion on 50 series so far. Incredibly mid. 5090 is the only really cool thing imo. Don't care about MFG so it doesnt matter to me.
I'm not looking forward to the 5070 or the 5060 either.

 

[192kbps]

Not doing any CPU overclocking?

 

[sLowEnd]

MSI didn't even try with this cooler. lol
They can easily make something that's not so loud.

 

[Bwaze]

MSI didn't even try with this cooler. lol
They can easily make something that's not so loud.

Several AIBs went on the record to say for instance to Der8auer that there was almost no time to do the design, evaluation, final testing - all the buyers now are basically beta testers.

 

[BorntoPlayGJFF]

Fake MSRP, it wasn't in stock on release day.

 

[AtiAmd]

If new lower gen cant surpass upper old gen - where is the progress?

 

[Knight47]

If new lower gen cant surpass upper old gen - where is the progress?

But nvidia said 5070ti is up to 30% faster compared to 4070ti in pure raster.

 

[192kbps]

Is your CPU running at default settings?

 

[THU31]

Bought this card a few days ago. The value of the US dollar is quite low these days (3.60 PLN instead of 4.15 PLN back in January), which has dropped prices of most electronics in Poland. I paid about $780 + VAT, but in Polish currency it was 3449 PLN. You couldn't really buy even the cheapest models below 4100 PLN around launch (MSRP was and still is 3899 PLN).

Started playing around with undervolting right away. Currently I'm at 2700 MHz @ 0.85 V. In Expedition 33 I'm getting 65-85% GPU usage with 130-180 W power usage (capped 60 FPS), so I'm happy with that.
In less demanding games like Motorfest or Valhalla I can do native 4K60 now instead of 1620p, with similar headroom left.

Efficiency is similar to my 4070, so ~50% more performance results in roughly ~50% more power. Uncapped benchmarks can peak at 230-250 W (4070 would peak around 160 W), but actual gaming at 60 FPS is mostly under 200 W. I'll definitely use more conservative settings during the summer, will probably push the card more between October and May.
Idle wattage is rather high (over 20 W), and very high during video playback - can go over 40 W watching YouTube/Twitch, and even over 50 W when upscaling SD videos to 4K with madVR.

I noticed different boosting behavior compared to previous generations. You can't really lock it to a single frequency even if the curve is flat, the clock will always jump around 10-50 MHz lower (so 2650-2690 in my case). And it used to jump in increments of 15 MHz, now it's 7-8 MHz.
The card also doesn't seem to like voltages lower than 0.85 V under heavy load. Even if I set the curve to 2500 @ 0.8 V, the voltage will still often jump to 0.83-0.85 V, so it's better to just lock it to 0.85 V with a higher clock for the strongest undervolt.

The fan profile is rather aggressive, trying to keep the card around 60 degrees. Setting a custom curve drastically reduces fan speed while keeping temps under 70 degrees (around just 200 W, though).
Fan stop behavior is also different. Fans seem to start up around 51-52 degrees instead of 60 on my 4070 Ventus. They also won't stop spinning during load even if the temperature drops to ~40. A bit weird. Custom fan curve doesn't seem to affect the fan stop threshold.

I have a GT 1030 to use for 32-bit PhysX games and it works nicely. Not sure what will happen when they remove driver support for Pascal, though. In theory you can install an older driver for the second GPU via device manager, but I don't know if PhysX will work that way. I guess we'll see.

Overall I'm definitely happy with the card so far. I paid 3199 PLN for the 4070 in April 2023 (which had a much lower MSRP of $600, but the US dollar was at 4.60 PLN back then), so taking inflation into account, this card is a much better product at 3449 PLN.


On a side note, this is only my second ever "Ti" GPU. I had the GF3 Ti200 for a while like 20 years ago, but since they brought it back with the Fermi refresh, I always had a regular model (with the 2070 Super being the only exception).