Igor's Lab, who's well-connected within the graphics card industry, just got the scoop on the quality of silicon that we can expect from Nvidia's GeForce RTX 30-series (codename Ampere) graphics cards.
Unlike the GA100 that's built with TSMC's 7nm FinFET process node, Nvidia's Ampere-powered GeForce gaming graphics cards employ Samsung's 8nm process node. The South Korean giant's 8nm manufacturing process isn't new by any means, as it has been around for a couple of years now. Nevertheless, it's still a big step forward considering that Turing was on TSMC's 12nm process node. Furthermore, Samsung has improved and tailored the 8nm fabrication process specifically to Nvidia's needs. The custom process, which has been dubbed "8N Nvidia", produces chips that deliver up to 10% higher performance than other chips that are made with Samsung's 8nm process node.
Both the GeForce RTX 3090 and GeForce RTX 3080 employ the GA102 silicon. Nvidia hasn't revealed dimensions of the GA102, but current speculation puts the die over 500mm square. There's no doubt that it'll be a big chip, though it should still be smaller than Nvidia's elephantine Turing TU102 die that checks in at 754mm square. The die shrink helps push the GA102's number of transistors up to 28 million as well, a 50.5% increase over the TU102.
GeForce RTX 3090 & GeForce RTX 3080 Specifications
| Header Cell - Column 0 | GeForce RTX 3090 | GeForce RTX 3080 |
|---|---|---|
| Architecture (GPU) | Ampere (GA102) | Ampere (GA102) |
| CUDA Cores | 10,496 | 8,704 |
| RT Cores | 82 | 68 |
| Tensor Cores | 328 | 272 |
| Texture Units | 328 | 272 |
| Base Clock Rate | 1,400 MHz | 1,440 MHz |
| Boost Clock Rate | 1,700 MHz | 1,710 MHz |
| Memory Capacity | 24GB GDDR6X | 10GB GDDR6X |
| Memory Speed | 19.5 Gbps | 19 Gbps |
| Memory Bus | 384-bit | 320-bit |
| Memory Bandwidth | 935.8 GBps | 760 GBps |
| ROPs | 96 | 80 |
| L2 Cache | 6MB | 5MB |
| TDP | 350W | 320W |
| Transistor Count | 28 billion | 28 billion |
| Die Size | ? | ? |
| MSRP | $1,499 | $699 |
As you would expect, initial yields and silicon quality in early stages of production won't be phenomenal. As production matures, both start to gradually improve. The early figures for Ampere aren't that bad actually. According to Igor, vendors divide the quality of the silicon into three main categories: Bin 0 for average quality, Bin 1 for good quality and Bin 2 for very good quality.
Ideally, you would want to land a chip from Bin 2. Dies from this category reportedly operate at a significantly lower temperature under full load with a higher boost clock speeds at the same voltage. Additionally, there is more overclocking headroom.
The German publication's sources claim that 30%, 60% and 10% of the GeForce RTX 3080's dies belong to the Bin 0, Bin 1 and Bin 2, respectively. The data is reassuring since it suggests that the majority of consumers should get a good sample, unless you're one of the unlucky ones to fall inside the 30%.
This should be the same for the GeForce RTX 3090 as well, except that the percentage of Bin 2 chips should be higher, maybe. This is due to the fact that manufacturers are only receiving a small number of GA102 dies from the initial production. However, RTX 3090 will necessitate using the higher binned GA102 chips in the first place since it has 82 SMs instead of only 68 SMs.
That last part will likely impact the overall quality of RTX 3080 chips. If most of the Bin 2 chips end up going into RTX 3090, and there's a reasonable chance that most of the Bin 0 chips will end up in RTX 3080, your chance of getting a 'very good' RTX 3080 drop. Beyond that, the AIBs will further bin whatever chips they receive and offer varying models of each GPU, with the best chips being reserved for higher factory overclocks ... and higher prices.
If you want the best performing RTX 3080 GPU, you're likely better off just buying a heavily factory overclocked card. That will guarantee you a minimum level of performance. How much faster will those cards end up being relative to 'reference' clocks? We don't know for sure, but in the past we've seen typical factory overclocks improve performance by around 5%, with the top overclocked models being 7-10% faster than the base models. After overclocking, most cards end up within a narrow performance range.
Our general advice is to buy whatever card has the best price, unless you're particular about aesthetics. All that RGB bling adds to the price, but at least it usually comes with a modest boost in clocks. Check back next week for the full RTX 3080 launch review.
