Nvidia is keeping the RTX 4070 under wraps but it looks like absolutely no one else is. We’ve got what’s claimed to be an entire specification list, and proof of third-party models from MSI and Palit, to ponder over ahead of the graphics card’s expected release.
Videocardz has got hold of a slide that is claimed to be from Nvidia. We’ve no guarantee of that, but it looks like it holds water.
The slide shows Nvidia will be using the AD104 GPU for the RTX 4070—the same GPU used in the RTX 4070 Ti. That checks out with what we’re expecting, and the rest of the spec is no surprise, either. There’s reportedly 12GB of GDDR6X memory capable of pumping out 504GB/s bandwidth across the card, working in tandem with a large L2 cache size of 36MB to push performance. That’s 9x the L2 cache of the equivalent RTX 30-series cards.
In terms of performance, there’s a graph comparing the RTX 4070 to the RTX 3070 Ti and RTX 3070. It looks good, but what we’re looking at is an RTX 4070’s lead with DLSS Frame Generation enabled. That’s what that big chunk cited as ‘300 OFA’ is. Ada’s Optical Flow Accelerators, alongside those 4th Gen Tensor Cores, can run frame generation to massively improve performance, whereas the ones in Ampere cannot.
For a more apples-to-apples comparison, you can see the shader TFLOPs noted as ’29 shader’ for the RTX 4070 in that graph. An improvement over the 22 TFLOPs for the RTX 3070 Ti, which means we should see a reasonable step-up in non-DLSS game performance; that’s your traditional rasterized stuff.
That 29 TFLOPs number actually puts the RTX 4070 at around the same as the RTX 3080 10GB at 29.7 TFLOPs, and just a touch under the RTX 3080 12GB’s 30.6.
Raw compute performance measured in TFLOPs isn’t an entirely accurate measure of game performance. Memory is also very important.
Nvidia has seemingly rolled out the same memory configuration on the RTX 4070 as on the RTX 4070 Ti. That’s a memory system that does seem a bit weak next to last gen’s RTX 3080, with a slimmer memory bus and lower overall bandwidth, but the RTX 4070 would be roundly topping the RTX 3070 for bandwidth overall. Also where the RTX 3070 used GDDR6 memory, the RTX 4070 is sticking to the GDDR6X memory favoured by the higher-end RTX 40-series cards.
And then there’s that extra pool of L2 cache on the new Ada cards that has seen them deliver so much extra performance when the overall core architecture has remained largely familiar.
If the RTX 4070’s price tag is the $599 sum so far reported by rumours, this specification looks to be one of the better value RTX 40-series cards for the money. Still, a serious sum for a fourth-tier graphics card—will we ever see a price correction for graphics cards?