Despite lack of news through official channels, news on Nvidia’s Pascal lineup of graphics cards continues to leak onto the internet. The latest whispers in the wind hint that the GTX 1080 & GTX 1070 fully embrace GDDR5X, but Nvidia have also ensured the GP104 supports options for GDDR5 variants for lower cost models.
As we’ve previously detailed, the benefits of GDDR5X over the older GDDR5 memory standard is simple – considerably higher memory bandwidths using the same bus widths. This helps cut down the complexity of the PCB (as you need fewer traces over the circuit), reduces power consumption while maintaining the memory throughput required to feed the next generation Pascal chips.
With this in mind, the rumors put the Pascal GTX 1080 to feature 8GB GDDR5X memory on a 256-bit memory bus, with the memory clocked to provide around the 12Gbps. This allows the total bandwidth across the board of between 320GB/s to 384Gb/s. The GTX 1070 cuts the clocks down to just 8Gbps GDDR5, meaning total bandwidth is 256GB/s.
To put these numbers into perspective, the GTX 980 offers 224Gb/s bandwidth, and the GTX 980 Ti offers 336GB/s. Remember, the 980 Ti uses the GM200 core (the Titan core) while the 980 uses the GM204 core, the predecessor to the GP104 which the 1080 and 1070 will utilize.
From what Nvidia revealed at the GTC conference when the company shown off the Tesla P100 architecture, there are a few major architectural changes to Pascal. One of the obvious differences is a fundamental shift in the number of Cuda Cores per SM (Streaming Multiprocessor). With Pascal, Nvidia opted to place in 64 CUDA cores per SM, compared to Maxwell’s 128.
Nvidia are also ramping up the clock speeds for the Pascal architecture, with figures touted to be up to 33% higher compared to Maxwell. In theory, this will allow the GTX 1080 to offer speeds greater than the GTX 980 Ti, while the GTX 1070 (likely to be one of the more popular Pascal models) will have performance between the 980 and 980 Ti. As usual though, take this speculation as just that – based upon the limited information Nvidia have officially revealed, and the multitude of leaks.
Speaking of leaks, a die shot of one of the GP104 cores supposedly pegs the core at measuring around 300mm2 in size, damn impressive considering 8 billion transistors are squeezed into this space.
The current rumors put Pascal being revealed at this year’s Computex, which takes place at the end of May.
And what of AMD? Well, they too are said to be announcing Polaris at Computex. We’ve already seen several demonstrations of the Polaris hardware, but lacking solid information on what exactly the specs are on this new hardware. Recent leaks hint to multiple SKU’s for Polaris 11 (the mid range GPUs) with one model offering at least 16 Compute Units.
Just like their rivals, AMD are also likely to use a mixture of GDDR5X and GDDR5 RAM for the higher end Polaris 10 SKU’s (for example, the R9 490), and on a similar 256-bit memory bus.
The official announcement from both companies doesn’t necessarily means a hard launch however, and it’s possible we’ll need to wait until some point in June for either Polaris or Pascal to be released and available in retail stores.
For those who’re hoping to jump on the “big” GPU’s – which would be the GP100 (the Pascal Titan) or the next generation Fury cards (Vega – AKA Greenland), you’re likely going to be waiting until Q1 2017. Vega 10 is scheduled for release at some point within 2017, and GP100 is likely the same. Most of this can be blamed on the production of HBM2 chips, which won’t be entering production until Q3 and Q4 (the forth quarter sees higher density RAM chips begin mass fabrication) which delays both team red and team greens plans.
It’s said that the Pascal architecture might still suffer from issues with Async. Compute, a problem plaguing the Maxwell architecture. Despite this, compute performance will certainly increase across the board, for both being a radical redesign over Maxwell.
Source – BitsnChips.it