Friday, December 9th 2022
NVIDIA GeForce RTX 4070 and RTX 4070 Ti Detailed Specs Sheet Leaks
It turns out that NVIDIA has not one, but two new GeForce RTX 40-series "Ada" SKUs on the anvil this January. One of these is the RTX 4070 Ti, which we know well to be a rebranding of the RTX 4080 12 GB in the face of backlash that forced NVIDIA to "unlaunch" it. The other as it turns out, is the RTX 4070, with an interesting set of specifications. Based on the same 4 nm AD104 silicon as the RTX 4070 Ti, the new RTX 4070 is significantly cut down. NVIDIA enabled 46 out of 60 streaming multiprocessors (SM) physically present on the silicon, which yield 5,888 CUDA cores—the same count as the previous-gen RTX 3070, when compared to the 7,680 that the maxed-out RTX 4070 Ti enjoys.
The GeForce RTX 4070, besides 5,888 CUDA cores, gets 46 RT cores, 184 Tensor cores, 184 TMUs, and a reduced ROP count of 64, compared to 80 of the RTX 4070 Ti. The memory configuration remains the same as the RTX 4070 Ti, with 12 GB of 21 Gbps GDDR6X memory across the chip's 192-bit memory interface, working out to 504 GB/s of memory bandwidth. An interesting aspect of this SKU is its board power, rated at 250 W, compared to the 285 W of the RTX 4070 Ti, and the 220 W of its 8 nm predecessor, the RTX 3070.
Source:
harukaze5719 (Twitter)
The GeForce RTX 4070, besides 5,888 CUDA cores, gets 46 RT cores, 184 Tensor cores, 184 TMUs, and a reduced ROP count of 64, compared to 80 of the RTX 4070 Ti. The memory configuration remains the same as the RTX 4070 Ti, with 12 GB of 21 Gbps GDDR6X memory across the chip's 192-bit memory interface, working out to 504 GB/s of memory bandwidth. An interesting aspect of this SKU is its board power, rated at 250 W, compared to the 285 W of the RTX 4070 Ti, and the 220 W of its 8 nm predecessor, the RTX 3070.
93 Comments on NVIDIA GeForce RTX 4070 and RTX 4070 Ti Detailed Specs Sheet Leaks
For those not in the know, this isn't about artificial patterns, but about a trilinear optimization that translated to texture shimmering and, iirc, visible transitions between various LOD levels.
Fwiw, I would even be ok with the same or up to 10% less performance than current gen if that translates to a 25% or more reduction in price.
"The more you buy(pay), the more you save(get)".
Nvidia has also reported declining sales, so we know there is some pressure on them. We'll see how this plays out.
6700 XT 335 mm², 96 MB Infinity Cache $479
7900 XTX 306 mm², 96 MB Infinity Cache $999
But people still gonna buy it and claim they are voting with their wallets for better future.
A lot of people in Nvidia threads are furious about Nvidia pricing but they won't buy it anyway, actually they are praising Nvidia to increase prices so AMD respond with lower prices (Titan Z $3000 vs RX 295X2 $1500) so they could buy only AMD as always.
RTX 3080: 8704 shaders, 272 TMUs, 96 ROPs, 320-bit, 628 sq. mm.
RTX 4070: 5888 shaders, 184 TMUs, 64 ROPs, 192-bit, 295 sq. mm.
I don't see how RTX 4070 will get close to RTX 3080 even in the most cherry-picked cases. You need to calculate the total die area, which for Navi 31 is not 306 sq. mm but over 500 sq. mm.
And look at their theoretical compute performance. The 4070ti has almost 2x the compute power and the 4070 has about 1.5x the 3070s. Those aren’t numbers they can fudge. They don’t necessarily translate one-to-one in game performance, but the 4070s are going to be *significantly* faster than the 3070s. Roughly 1.5x-2x faster.
Because it would mean roughly RTX 3090 - RTX 4080 performance.
Still, at the end of the day internal details can be irrelevant. What matters is $$$/fps.
6700XT transistor count: 17.2 million
7900XTX transistor count: 58 million
6700XT memory: 12gb, 192bit bus, 16Gbps rate
7900XTX memory: 24gb and 384bit bus, 20Gbps rate
Sooooo the 7900XTX has 3.37x more transistors, 2x the amount of RAM, 2.5x more RAM bandwidth, about 4.5x more compute power, and costs about 2x more. Seems like a pretty decent deal to me, even if it *was* the successor to the 6700xt. My point was that the GDDR6/X bus width and data rate doesn’t really matter on the 4070s because the insane increase in L2 cache more than compensates for any loss in memory bus throughput.
They’re using the same strategy AMD started using with the 6000 series. The 6800xt/6900xt could get away with 256bit wide GDDR6 at 16Gbps because they had 128mb of L3 cache to compensate for it. Now we’re seeing Nvidia following suit with an increase to the size of their L2 cache by 1200%. Nvidia’s massive L2 cache solution is analogous to AMDs infinity cache solution. GDDR6/X is too slow for it keep up with these new GPUs and the solution is to implement an insane amount of cache.
I am inclined to believe that clock speeds don't scale performance linearly. It's more like 10% higher clock means 5% higher performance. I agree. I think 4070 is a badly designed chip with severe internal imbalances which would cause performance issues, especially at 2160p, 4320p and beyond.
nVidia can take these RTX 40 series and shove them up their @$$