Nvidia Tesla V100: Has Nvidia hit the Volta again?

Keynote 2017 Nvidia Tesla V100 Volta Sk Hynix Gddr6 Amd Vega Hbm 1.gif
Keynote 2017 Nvidia Tesla V100 Volta Sk Hynix Gddr6 Amd Vega Hbm 1.gif

It seems that AMD does not have enough resources to do the right thing in the field of processors and in the field of graphics. And it is that although the launch of Ryzen has been the greatest success of the company in many years (despite the recent decline in its shares), it does not seem that its departments dedicated to the graphics market are able to do it as well. Do not get us wrong, AMD has done very well with its Polaris architecture, but the truth is that even the recent Radeon 580 and cia are still chips intended for the mid / high range, leaving the high range itself and the extreme range are a private preserve where Nvidia roams at ease. The proof of it is that Nvidia has carried out a new Keynote a couple of hours ago where Jensen huang, CEO of Nvidia, talked about the new Volta architecture, in addition to other diverse topics such as Virtual and Augmented Reality, Deep Learning, Deep Intelligence, loTs or autonomous vehicles among other things. If you are interested in seeing the entire presentation, we leave it linked here, but we are going to focus on the Nvidia Tesla V100, the latest chip presented by the company and which marks the coming-out of Volta, Pascal’s successor.

Nvidia Tesla V100: Using HBM2 memory wisely.

As the weeks go by and everything related to AMD’s VEGA does not stop being teasers and leaks, Nvidia already has a real product on the street using its new architecture. His name is Nvidia Tesla V100 and arrives as a successor to the Tesla P100, launching a new manufacturing process of 12nm FFN courtesy of TSMC.

The Tesla V100 GPU is sized 815mm2, so despite its renewed lithography, it is a larger chip than its predecessor (approximately 30%). The culprit for this size is a number of CUDA Cores rather beastly that amounts to 5120 (21 billion transistors), which represents an increase of almost 43% over the 3,584 CUDA cores present in the Tesla P100.

Keynote 2017-Nvidia-Tesla-V100-Volta-SK-Hynix-GDDR6-AMD-VEGA-HBM

With this generous number of cores, the Tesla V100 is capable of developing a performance of 15 FP32 TFLOPs (vs 10.6 TF) and 7.5 FP64 TFLOPs (vs 5.3 TF), all of this operating at a Base frequency of 1465 MHz, somewhat lower than the 1480 MHz of the P100.

Nvidia does decide here to pull the fastest memory on the market, and we find 16GB of HBM2 memory, which together with an interconnection NVLink 2.0 throw a monstrous bandwidth of 900 GB / s, 180 GB / s more than its predecessor.

For the rest, we have a 6144 KB Level 2 (L2) cache (vs 4096 KB) and a TDP that is kept in the 300W, the same present in the Nvidia P100 and denoting the good work done by the greens regarding the consumption of Volta.

And what about Vega ?: Well, possibly it has supply problems (again).

AMD seems not to be chastened by the supply and price problems that it already suffered in its day with the Radeon R9 Fury and Fury X, cards that, although not bad, were not as profitable as they could have been as a result of integrating a new memory HBM that despite its bandwidth, limited the maximum amount of memory to 4GB, caused stock limitations and also made the product more expensive. Now it seems that history could repeat itself with AMD Vega, because according to some sources such as TweakTown, AMD Radeon RX Vega would hit the market with a very limited stock of some 16,000 cards for the first months of life.

The reason for this lack of stock would again be the type of memory used, and that is that while Nvidia will only accompany Volta with memory HBM2 in professional cut products such as the Tesla V100, AMD would use this type of memory in its RX Vega series products, something that can be a new setback compared to the high-end Volta chips destined for the world of video games, which would make use of a new and much cheaper GDDR6 memory.

And it is that AMD does not seem to have learned from its mistakes, and continues to bet on expensive and scarce technologies for its mass consumer products, while a much more rational and logical Nvidia has managed to control the “High End” segment with memory GDDR5 and its GDDR5x revision. Now it will be GDDR6 memory that will allow greens to put their products on the street without stock limitations, something that could seriously harm AMD in the high-end, since Navi (the next AMD architecture) is expected to integrate also GDDR6 memory.

And speaking of GDDR6, Just yesterday SK Hynix showed the world a wafer with this type of memory, a memory that seems not to arrive until the beginning of 2018 from the hand of Volta and that will offer a speed of 16 Gbps compared to 10Gbps of the current GDDR5X. With this type of memory, a high-end chip with a 384-bit bus it would be able to offer 768 GB / s (548 GB7s with GDDR5X memory), all with a voltage of only 1.35v, compared to 1.55v of the current GDDR5X.

Hopefully AMD RX Vega hits the market early enough not to make the Voltas with GDDR6 available, because as they are released at the same time, things do not look good for the reds. Let’s see what the next few months hold for us.

via: Videocardz, TweakTown,