NVIDIA is currently embroiled in legal proceedings to be able to register the Hopper architecture as its own, because another company uses this name for its products. If that’s already a stumbling block in the road, what we’re about to see next is no less so, as it suggests that many of the issues we understand have been secretly resolved.
Here again, we must not play with what is said, but with what is not said. The first Leaker reported that the GH100 chip (high-end Hopper architecture) would have more or less an approximate area of 1,000 mm2. That in itself is an impressive number considering that GA100 is just over 800mm2 (826mm2 to be exact) and is manufactured at 7 nm.
These are 858mm² (26*33) and 980mm² (35*28), yields ~<50% without counting the redundant structures seem a bit too much to me.
Unless it’s a flagship of flagships sold for $20k each to HPC customers. It almost seems more cost effective to just take a whole wafer to take away. pic.twitter.com/2Fh1twbaqr
– Rule110 (@Rule1103) January 29, 2022
The GH100, on the other hand, will reach 5 nm and it is speculated with a size greater than 858 mm2, since the size of the slices is always 300 mm and the number of chips would be more optimal in terms of profitability for each registered , since of those 858 mm2 (chip por 33 x 26 mm) 26 good days would come out, whereas if we increase the area to 35 x 28mm (980mm2) fewer useful chips are wasted, but that translates to 5 good dies, with NVIDIA losing a lot of money per wafer, or else making the chips at a much higher price.
GH100: monolithic architecture or single monolithic chip
The GH100 has no MCM, it is still monolithic.
— Greymon55 (@greymon55) January 29, 2022
Here the differences lie in the concept and in the information provided. The leaker’s response Greymon55 a Kopite7kimi This is quite interesting, as it doesn’t refute the approximate die measurements, but rather supports the idea of a single die in the so-called GH100, therefore, it’s not a dual MCM package as speculated , it would still be a traditional monolithic chip.
Why is this important? Leaks so far told us that Ada Lovelace was NVIDIA’s monolithic architecture for gaming, while Hopper would go for data centers, AI and DL as an AMD-style MCM chip with Ryzen (economy of distances). With this confirmation, the idea is ruled out and the changes could be minor in terms of chip design, leaving AMD alone with RDNA 3 in this section.
DO NOT overreact. We know that the actual die size will be significantly smaller than the package size.
But GH100 > GA100 is confirmed. 😁– kopite7kimi (@kopite7kimi) January 29, 2022
It’s also possible that NVIDIA will launch a new series of MCM chips, so we’re not approaching all of this in the right way. It should be noted that we are talking about a specific chip (GH100) which so far and according to the nomenclatures should be the best of the brand for architecture, but nothing tells us that it does not launch a new series separate chipset with MCM designs per package, where HBM3 and newer substrates are used SoIC 3D TSMC.
Little by little, the data is confirmed, but we are very far from being able to see the complete picture, so be patient, the information will come.
