Nvidia Pascal GP104 Spotted?

So far all of the rumours around the GP104 and GTX 1000 series have mostly been about release date and specifications. The closest we’ve gotten to physical evidence have been the shrouds for the GTX 1080 and GTX 1070. For the first time, we’re getting a picture of the physical die and parts of the GPU board around it. According to ChipHell, the die shot you see below belongs to the GP104, the mainstream Pascal GPU.

From the die shot picture, GP104 appears to be about 15.35mm x 19.18mm for a total of about 290 to 300mm². This is the same as GK104 which was also a die shrink and came in at 294mm² and much smaller than GM204 which was a relatively massive 398mm². This shows that Nvidia is starting out with small dies first with the GTX 1070 and 1080 and releasing a GP100/102 Titan and 980Ti later on.

For now, we still don’t know what GP104 will look like, but it seems that most of the FP64 units in GP100 will likely be stripped out and replaced by the more ‘useful’ FP32 ones. The leak also suggests that total FP32 CUDA core count will be around the same as the Titan X but the TMU and ROP count seems closer to the GTX 980. I expect that clock for clock, GP104 won’t be much faster than the Titan X but it will be ahead and much more efficient.

Finally, we can see what appears to be Samsung 2Ghz 1GB GDDR5 DRAM modules for 8GB total. This suggests that that GDDR5X isn’t ready in time or will be reserved for the GP100/102 consumer release. This follows the same trend set by the GTX 680 which was more powerful than the 580 but featured lower memory bus width but faster VRAM and more memory overall.

While the leak is promising, it is a leak after all and I would make sure to take all of this with a shipload of salt. Given the information we know though, this leak may very well reflect reality.

AMD R9 Fury Series Fiji GPU And HBM Die Shots Revealed

We all have a lot of chips inside our systems, but companies such as AMD and Nvidia rarely show us what that looks like and we have to rely on third-parties such as ChipWorks to show us what the insides look like. They have now done that with the Fiji chip from the Radeon R9 Fury series and we finally get to see a die shot of it.

A lot of the time this isn’t that interesting as they’re using the same old technology, but this time it is particularly interesting because the Fiji chip is the first to use the HBM JEDEC standard for stacked memory. While the GDDR5 standard has been great for many years, it is getting old and has more or less reached its limit. Granted, a GDDR5x standard has been rumoured for a release next year, but HBM is the future.

HBM offers extremely high memory bandwidth at a considerably lower thermal and power cost than GDDR5, and it was also said to require fewer transistors than GDDR5. That was a thing that we hadn’t been able to verify up until now, but it looks like it was correct. Below we see the Fiji chip next to the Tonga chip and in size relation. While the Fiji HBM bus takes up a little more overall space, it also provides a lot more with its 1024bit over the 64bit GDDR5 bus.

AMD and SK Hynix have spent seven years on the development of the HBM standard and it looks like it’s going to pay off, especially with what we’ve already heard of the HBM2 generation that will take the whole thing to a completely new level.