We'll have to wait for benchmarks, but based on specs, this new Titan X looks like the best single GPU card you can buy for deep learning today. For comparison, the GTX 1080's memory bandwidth is quoted as "320GB/s." The new Titan X also has 3,584 CUDA cores (1.3x more than GTX 1080) and 12GB of RAM (1.5x more than GTX 1080). Nvidia says memory bandwidth is "460GB/s," which will probably have the most impact on deep learning applications (lots of giant matrices must be fetched from memory, repeatedly, for extended periods of time). As usual, it's only available with the Nvidia-designed cooler, although the usual suspects were quick to announce their third-party water blocks for anyone interested in water-cooling.Great news all-around for deep learning practitioners. In the UK and US markets, Nvidia is limiting availability of the standalone Titan X to its own web-store, although select etailers will have stock in those countries where the web-store isn't available. This is primarily a gaming card, then, although Nvidia has now dropped the GeForce GTX branding from the Titan X to indicate some sort of separation between it and its 'normal' gaming cards, while retaining the Titan X name from the last generation to suggest consistency in it being the ultimate card. Nvidia says the new Titan X was planned to run with GDDR5X all along rather than HBM2, and that with its compression technology and the raw speed here GDDR5X offers enough bandwidth to feed the massive GPU.Īs has been the case with Titan cards for some time now, FP64 performance is not accelerated and the Titan X is not marketed as a compute card – Nvidia will point you to its Tesla P100 for such things now. In raw terms, this is an uplift of more than 40 percent over the Maxwell Titan X, but as a Pascal part the new card also benefits from fourth generation delta colour compression, meaning that on average the gap should be even bigger. The 384-bit memory interface from the GTX Titan X (Maxwell) has been carried over, meaning we're looking at a memory bandwidth of 480GB/sec. Here, you get a whopping 12GB frame buffer and it's running at the same effective speed of 10Gbps. While the Tesla P100 uses HBM2 memory, for Titan X Nvidia is back with GDDR5X, first introduced with the GTX 1080. This is still a fair way behind the GTX 1080, however, and even further behind typical overclocked, third-party variants of said card. Nvidia has also been able to ramp up the clock speeds considerably thanks to Pascal's transistor-level improvements: the base clock is up by over 40 percent, now sitting at 1,417MHz, while the boost clock is rated at 1,531MHz. An astounding 12 billion transistors make up the GP102 GPU, which is 50 percent more than the Maxwell Titan X has yet thanks to the move to a 16nm FinFET process, the die size is actually down by over 20 percent to 471mm 2. With this GPU, Nvidia is able to deliver up to 11 TFLOPs of power, more than any seen before in the consumer market. Instead, two of the SMs are disabled, leaving us with 28, and thus texture unit and CUDA core counts of 224 and 3,584 respectively a sixth more than the previous Maxwell-based GTX Titan X and 40 percent more than a GTX 1080. However, even Titan X is not using a fully enabled GP100, presumably because fully enabled yields were too low. As such, we move from a layout of four Graphics Processing Clusters (GPCs) and 20 Streaming Multiprocessors (SMs) to six GPCs and 30 SMs. The Pascal variant of the Titan X is built around a new GPU, GP102, which is effectively the GP104 GPU used in the GTX 1080 scaled up by 50 percent. Click to enlarge - The full GP102 die in the new Titan X two of the SMs are disabled
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |