During the last Financial Analyst Day, AMD announced the CDNA (Compute DNA) architecture, finally putting in black and white what was already intuited at the presentation of the RDNA architecture of the most recent Radeon RX video cards: the solutions dedicated to machine learning, supercomputers and artificial intelligence will have an architecture different from the gaming proposals.
On the other hand, it does not make sense to create a complex architecture for the data center and to transpose it to the gaming sector, the risk is to create products with merits, but also many defects, such as the Radeon VII. It will therefore be very interesting to see what AMD will churn out in the HPC sector, an area in which the company will have to deal first with Nvidia and secondly with Intel, also ready to churn out high-performance GPUs based on Xe HPC architecture in the future.
Click to enlarge
The first AMD GPU based on CDNA architecture should be codenamed Arcturus (a code name that has been around for months) and according to Rogame hardware leaker would have really interesting specs, at least on paper. The future Radeon Instinct card should, as reported, count well 120 Compute Units for a total of 7680 stream processors (SP) – without prejudice to the 64 SP for CU. This would almost double the units of the current Radeon Instinct MI60 based on the Vega 20 GPU (64 CU, 4096 SP).
Compared to the proposals based on Vega architecture, the new Arcturus GPU should also have a greater amount of cache and a set of new features for the datacenter world, which will add to those already implemented in Vega 20, further improved. The leaker also talks about GPU frequency of 878 MHz, SoC frequency of 750 MHz and memory frequency of 1200 MHz. Clocks, being a test card, are the last thing to take into account, also because in the past rumors have circulated about much higher frequencies.
The only one that could be somewhat indicative is that of the memory, higher than that of the current cards, even if 1200 MHz still appear contained in the face of the capabilities of the latest HBM2E memories announced by the various manufacturers. As for the amount of memory, we are talking about 32 GB.
The current Radeon Instinct MI60 offers FP16, FP32 and FP64 computing performance up to 29.5, 14.7 and 7.4 TFLOPS respectively, therefore it cannot be excluded that Arcturus doubles these values. In this scenario, we must not forget the rumors about the possible Nvidia HPC solutions coming soon: the new proposals should count more than 8000 CUDA cores, for an even higher computing power than the estimates on Arcturus.
It is not clear when we will see the new generation of accelerators on the market, given the situation linked to the COVID-19 pandemic, with Nvidia having canceled the GTC 2020 and postponed any announcement to better times. The presentation of a new computing architecture for datacenters was expected at GTC 2020.
