Nvidia’s Pascal GP100 GPU: huge bandwidth, huge double-precision performance

For the beyond yr, fans had been chomping on the bit looking forward to the next era of photographs playing cards to reach. The 28nm node has endured for a long way longer than any preceding era, and while both AMD and Nvidia have added a couple of merchandise on that node, clients have simply wanted the power efficiency and overall performance improvements that the 14/16nm node ought to offer. these days, Nvidia showcased the entire HPC version of Pascal and unique what the card might offer as compared with its preceding Maxwell and Kepler merchandise.

Pascal’s renewed awareness on high-pace compute

whilst Nvidia designed Maxwell, it made the design to dispose of a lot of the double-precision floating factor capabilities that had been baked into its previous Kepler structure. The old Tesla K40, primarily based at the GK110 GPU, changed into capable of up to 1.68 TFLOPS/s, whilst the Tesla M40, which used the Maxwell GM200, may want to simplest attain 213 GFLOPs. The M40 still had an advantage over the K40 in single-precision floating point, but double-precision floating point overall performance changed into sharply curtailed. As we discussed last week, when AMD released its FirePro S9300 x2, this limited the varieties of workloads wherein the M40 may want to excel.

Pascal’s present day GP100 variant adds again all of the double-precision floating factor that Maxwell was lacking — then stuffs some more in, only for correct degree. The chart below compares Kepler, Maxwell, and Pascal. word that the dev blog post states that Pascal can include up to 60 SMs, whilst the variant defined beneath has simply 56.

One exciting aspect of Pascal’s design is that Nvidia has once more decreased the variety of streaming cores in every processing block, or SM and adopted the equal ratio that AMD uses, with each compute block containing 64 processors. the full range of streaming processors has increased 17%, as has the range of texture processors. There’s no phrase yet on ROP counts, but assuming Nvidia observed its historical pattern, the GP100 ought to have at least ninety six ROPS and possibly 128. Base clock is also up forty% over Maxwell, and at the same time as Tesla clocks are normally more conservative than their laptop opposite numbers, the reality that Nvidia squeezed a 40% clock soar out of this silicon shows we can look forward to similar profits when Pascal involves the consumer market.

The memory interface is the most important generational upgrade. HBM2 offers a 4096-bit bus and 720 GB/s of memory bandwidth, compared with 336GB/s of bandwidth available on the best-end Titan X.

Pascal also makes use of a simpler datapath enterprise, advanced scheduling with better power efficiency, overlapped load/save commands, aid for Nvidia’s NVLink interface, help for sixteen-bit floating factor (1/2 precision), and progressed atomic functions. GP100 also supports ECC memory natively, meaning there’s no overall performance or storage penalty for activating the characteristic.

One note on NVLink: There’s been confusion over where and how this bus is used. For the most element, NVLink is a way of connecting multiple GPUs to every different, in particular move-connections in a multi-socket device, wherein forcing GPUs attached to 2 special CPUs to talk to each other could significantly degrade overall performance.

NVLink can be used to attach the GPU to the CPU immediately, but Nvidia’s blog post specifies that that is simplest applicable to electricity processors.

The diagram above is described as follows: “The [above] determine highlights an instance of a 4-GPU gadget with dual NVLink-capable CPUs related with NVLink. on this configuration, each GPU has a hundred and twenty mixed GB/s bidirectional bandwidth to the alternative three GPUs within the system, and forty GB/s bidirectional bandwidth to a CPU.”

Nvidia is likewise claiming that Pascal will provide “Compute Preemption” with a extensively progressed computing model. that is one region wherein team green has extensively lagged AMD, whose asynchronous compute performance has been plenty stronger than some thing NV has added to bear. Asynchronous compute and compute pre-emption are not the same thing — we’ll need to watch for transport hardware to look how this compares with AMD’s implementation and what the variations are.

an excellent bounce ahead for HPC, but no patron launch date but

It’s apparent that Pascal will appreciably improve Nvidia’s HPC function, and that’s essential because the business enterprise has large plans for deep learning, self-using cars, and other HPC workloads. Pascal looks like it’ll be a mighty suit for Xeon Phi, Nvidia’s primary competitor in this area.

Nvidia has remained mum on purchaser launch dates, but, so we’ll need to be patient when this tech makes it to the mass marketplace. Rumors we’ve heard in different contexts recommend that HBM2 hardware gained’t hit the patron market until later this 12 months due to excessive preliminary costs for first run system. It’s entirely feasible that Nvidia is using GP100 to fill out its initial high-give up merchandise, but will simplest pass to the HBM2 fashionable for upper-cease patron ranges within the back 1/2 of 2016.

when the ones cards do arrive, they need to be a full-size upgrade over Maxwell. The core counts on Pascal aren’t a whole lot higher than Maxwell, but the progressed clock speeds will pressure overall performance higher as nicely, and that’s before any development from efficiency profits. if you’re inside the marketplace for a new GPU this year, I strongly advise waiting to peer what NV and AMD ship inside the purchaser space if that’s viable.