Recent content by spworley

That impossible, no way, 24GB GDDR5X was demoed working at SIGGRAPH last month. Not shipping yet, though, probably because as you say that dense chip is not widely available yet.

I have long suspected the P100 doesn't even have the rasterizers needed for graphics and it's truly compute-only. The saved die space was used for those extra P100 features (fp16x2, flexible preemption, system wide memory addressing) and especially for the doubled number of registers per SM...

But remember NVidia dropped both the GeForce and GTX labels for the new Pascal Titan X. The annoucement and the product page both carefully label it "NVIDIA Titan X". The Maxwell card was carefully labelled "GeForce GTX Titan X". But of course, amusingly, the card itself has a lit up "GeForce...

A relevant comment from Tech Report gives an interesting contrarian data point to these Mercury and Peddie reports by looking at monthly Steam survey results. These aren't quarterly sales, these are user system reports, so these numbers tend to have a much smoother trend, essentially low-pass...

That's an excellent point! Tom's hardware shows GTX1080 load power between 180 and 210 watts. It's probably not a doubling because the ROP ratio is only 1.33x, only shaders are 2x But as you point out, that still gives extra headroom for Vega and should make it much easier to match GP104. In...

Current indications are Vega's performance will compete with GTX 1070 and 1080, not with GP102 Titan X. You can conclude this from power efficiency analysis alone. A GTX 1060 and Polaris 10 are effectively tied in performance. A GTX 1080 has double the compute of a GTX 1060, so to match it AMD...

Does Pascal's hardware texture compression reduce actual storage size of textures in RAM? The compression's main utility is to boost effective bandwidth, but if it also reduces RAM use then NVidia's superior compression rate may make its 3GB card perform better than expected in RAM limited...

You must have inside info on this fast Vega rollout plan! Polaris's events started with the first proclamation in November 2015. Unveiled in December 2015. Demoed January 2016. Specced in February. Announced in March 2016. Revealed in May 26 2016. Heralded on May 31, 2016. Released June...

There's an amusing admission buried inside. It's no surprise to any even casual Imagination observer, but the company's complete dependence on just one key customer is smoothly downplayed in the report to sound like a positive. "There are no parties with whom the Group has contractual or other...

On the CPU side, PMADDWD does four parallel 16 bit weighted pair summations, so it's like 4 DP2As. It goes all the way back to MMX, so it's been in your CPUs for 20 years.

The vmad instruction in CUDA performs bytewise multiply-and-accumulate. Describing throughput rate ratios isn't as straightforward as fp32 since int32 mads are multiple instructions and not a single clock unlike fp32, hence giving vmad the throughput rate bonus relative to that baseline. And...

That chart would be more useful if you added a DP4A row, which is the simple but useful new integer instruction in non-P100 Pascal. Ryan did a great writeup of the fp16x2 compute in his GTX 1080 review. He shows a similar chart but also usefully extending back to Maxwell. All Kepler, Maxwell...

It's really a good match for the GTX 950 in both price and speed... a tie, not an AMD loss. The interesting observation from TR's article is that the 2012 Pitcairn based R7 370 gives slightly better performance than the RX 460 at slightly less cost.

Not peanuts at all. Not a tiny market at all. NVidia's current HPC revenue is about 20% of the gaming segment and growing. Add in the fact that profit margins are not unreasonably many multiples higher for $4000 HPC parts versus $600 gaming parts, we can conclude the profits for NVidia's HPC...

I keep seeing this misconception from multiple posters that fast INT8 and INT16 support is new to Pascal. It's not. All Kepler, Maxwell, and Pascal parts have the key 4x rate INT8 MAD and 2x rate INT16 MAD, plus other functions like min/max and shift. These are commonly used in CUDA and are...