Absent from the discrete GPU area for over 20 years, this yr Intel is ready to see the primary fruits from their labors to re-enter that market. The firm has been creating their new Xe household of GPUs for a number of years now, and the primary merchandise are lastly set to reach within the coming months with the Xe-LP-based DG1 discrete GPU, in addition to Tiger Lake’s built-in GPU, kicking off the Xe GPU period for Intel.
But these first Xe-LP merchandise are simply the tip of a a lot bigger iceberg. Intending to develop a complete top-to-bottom GPU product stack, Intel can also be engaged on GPUs optimized for the high-power discrete market (Xe-HP), in addition to the high-performance computing market (Xe-HPC).
That excessive finish of the market, in flip, is arguably crucial of the three segments for Intel, in addition to being the riskiest. The server-class GPUs will likely be chargeable for broadening Intel’s profitable server enterprise past CPUs, together with keeping off NVIDIA and different GPU/accelerator rivals, who in the previous couple of years have ridden the deep studying wave to booming earnings and market shares that more and more threaten Intel’s conventional market dominance. The server market can also be the riskiest market, as a result of high-stakes nature of the {hardware}: the one factor greater than the earnings are the chips, and thus the prices to enter the market. So beneath the watchful eye of Raja Koduri, Intel’s GPU guru, the corporate is gearing as much as stage a serious assault into the GPU area.
That brings us to the matter of this week’s teaser. One of the advantages of being a (comparatively) upstart rival within the GPU enterprise is that Intel doesn’t have any current-generation merchandise that they should defend; with out the chance of Osborning themselves, they’re free to speak about their upcoming merchandise even effectively earlier than they ship. So, as a little bit of a savvy social media ham, Koduri has been posting occasional photographs of Intel’s Xe GPUs, as Intel brings them up of their labs.
BFP – massive ‘fabulous’ package deal😀 pic.twitter.com/e0mwov1Ch1
— Raja Koduri (@Rajaontheedge) June 25, 2020
Today’s teaser from Koduri exhibits off a tray with three totally different Xe chips of various sizes. While detailed details about the Xe household continues to be restricted, Intel has beforehand commented that the Xe-HPC-based Ponte Vecchio could be taking a chiplet route for the GPU, utilizing a number of chiplets to construct bigger and extra highly effective designs. So whereas Koduri’s tweets do not make it clear what particular GPUs we’re – in the event that they’re all a part of the Xe-HP household or a mixture of totally different households – the photograph is an fascinating trace that Intel could also be a wider use of chiplets, because the bigger chip sizes roughly correlate to 1×2 and 2×2 configurations of the smallest chip.
And with presumably a number of chiplets beneath the hood, the ensuing chips are fairly sizable. With a useful AA battery within the photograph for reference, we will see that the smaller packages are round 50mm extensive, whereas the biggest package deal is definitely approaching 85mm on a aspect. (For refence, an Intel desktop CPU is round 37.5mm x 37.5mm).
Finally, in a separate tweet, Koduri rapidly talks about efficiency: “And..they let me hold peta ops in my palm(almost:)!” Koduri doesn’t go into any element concerning the numeric format concerned – an necessary qualifier when speaking about compute throughput on GPUs that may course of lower-precision codecs at increased charges – however we’ll be beneficiant and assume INT8 operations. INT8 has change into a reasonably fashionable format for deep studying inference, because the integer format provides nice efficiency for neural nets that don’t want excessive precision. NVIDIA’s A100 accelerator, for reference, tops out at 0.624 PetaOPs for normal tensor operations, or 1.248…
