That's hilarious I got downrated for stating facts. (AMD fanboys are upset that GK110 is a real chip?)

http://www.anandtech.com/...k10-gk110-based-tesla-k20

Total of 2880 cores in 15 SMX modules.

Sorry, I thought there were 256 stream processors per SMX. Corrected, thanks!

Your sub-heading still states:

Nvidia Tesla K20: GK110 Chip, 3840 stream processors

Anton, but they were 256 CUDA Cores per SMX. 192 for Single precission and 64 extra cores for the double precission.

So, Can they run simultaneously?, I don´t know, but it´s probably (independent schedulers and many ports to CUDA cores arrays for them).

Yes, the math are correct:

15 SMX x 192 CUDA Cores (SP) = 2,888 CUDA Cores.
15 SMX x 64 CUDA Cores (DP) = 960 CUDA Cores.

Then:

2,880 + 960 = 3840 CUDA Cores (SP and DP Cores).

That's not how it works.

The CUDA cores are able to perform Single Precision OR Double precision calculations. Double precision calculations in GK110 are limited to 1/3rd of single precision performance. That's why 2,880 CUDA cores are "equivalent" to having 960 Double Precision capable Cuda cores. (2,880 x 1/3 = 960).

GK104 works exactly the same. There are 1536 CUDA cores, which are capable of 1/24th Double precision performance, or stated otherwise "equivalent" to 64 Double Precision CUDA cores. GK104 does not have 1536 + 64 = 1600 CUDA cores....

You should read up on GK110 white paper. There are no "extra" 960 CUDA cores in GK110 or 64 extra in GK104.
http://www.anandtech.com/...0-gk110-based-tesla-k20/2

It's 2,880 in total.

1 SMX Cluster has 192 CUDA cores.

Where are you getting 2:1?

The CUDA Cores of the Fermi arquitecture runs at double speed (frecuency) than the kepler ones. So they can do the same work that two Kepler CUDA cores in the same cycle.

GK104 does not have as much L2 cache in CUDA cores and doesn't have dynamic scheduling for compute work as Fermi did. Therefore, for computation work, the CUDA cores in Kepler GK104 are much weaker than 2x.

Also, you cannot assume a linear 2:1 relationship for games either since you don't know what architectural changes were applied to each CUDA core to make games run faster on them without diving deeper into the architecture. This is even more complex since Kepler now has 128 TMUs so it's impossible to isolate CUDA cores on their own outside of a lab environment.

If your theory was true, this wouldn't happen:
http://images.anandtech.com/graphs/graph5818/46450.png

1344 SPs using your 2:1 ratio => 672 CUDA cores and yet 512 CUDA Cores GTX580 is almost as fast.

It is incorrect to say that Fermi CUDA core is twice as fast as a Kepler CUDA core at the same frequency since it's an enhanced architecture in some ways for games but in GK104's case a worse architecture for double precision compute.

You can't just make blanket statements like that. You have to assess each specific GPU based on its specs and how that GPU fits specific market segments. Kepler is a flexible architecture that can be good for certain sectors even in compute since not all compute work needs to have high double precision throughput:
http://images.anandtech.c...0/TeslaMarketSegments.jpg

GK110 has also been beefed up from GK104. GK110 includes 64KB of on-chip memory for each SMX that can be allocated with a bit more flexibility when compared to the previous Fermi architecture, enabling 32/32KB split between shared memory and L1 cache. Also, GK110 packs 1536KB of L2 cache, double the amount of L2 cache found in the Fermi architecture and offers up to twice the bandwidth per clock. GK110 also has ECC memory protection support for all register files, shared memories, L1 and L2 cache and DRAM memory. Unlike GK104's (GTX670/680) double precision limit to 1/24 of single precision, GK110's double precision is now 1/3 of FP32 single precision.

If you followed the press release and what people who use compute in the professional markets have demanded are 3 things:

1) Single precision performance - K10 delivers that. These customers said they don't care for double precision performance. In the oil and gas exploration industries this is the perfect chip for them.

2) Double precision performance - K20 GK110 delivers 3x the double precision/watt of the previous M2090 / Fermi based board. With over 1 Tflop of double precision, the overall double precision throughput is actually faster than HD7970.

3) Multi-trasking and multi-threading - Dynamic Parallelism & Hyper-Q.

You have to be careful about concluding outright that Kepler is awful at compute. It is for some tasks but there are plenty of compute tasks such as Folding@Home where it is very fast, and faster than GTX580/7970 series.

Compute performance isn't just 1 thing. It is OpenCL, Single and Double precision performance (also compute shaders, etc.). People tend to clump all of this together as "Compute" which is incorrect.

This is why K10 makes a lot of sense for customers who want raw compute performance (i.e., use single precision compute programs) where Kepler is actually very competitive. It is true that for double precision, GCN is the superior architecture but so where HD4800/5800/6900 series relative to G80/GT200B and GF100/110! Most people don't know this because they actually don't use compute programs but post comments on the Internet because a review said it.

There were a ton of consumer compute programs where HD5800/HD6900 series mopped the floor with Fermi but hardly many professional programs where this was the case. As a result almost no one seemed to care that for 3 generations AMD cards have superior double precision performance, that is until reviews magically pointed it out with GCN.

In fact AMD has had the lead in double precision compute for over 5 years. However, what makes NV's Tesla cards so popular for professional computing applications that include data analytics, weather modeling, computational chemistry and physics is the software support and NV's technical support. AMD's GPU are very good from a technical standpoint but their professional software and technical support is lacking in this area, which doesn't allow them to show off their full potential.

A note:

The strenght of nvidia´s arquitectures is something more than "good support or professional software".

Its arquitectures are more flexible and have more efficiency (the practical performance in many types of applications are nearer to the theorical peak performance than AMD´s arquitectures).

Ya, I know. That's a good point. I was saying that while AMD has had superior double precision in HD4870/4890/5850/5870/6950/6970 series and now in HD7900 series, what has been holding them back was software and developer support. Their hardware was good.