- Register
- Login
29°C
- Article
- . (0)
- Company info
Last week’s Nvidia GTC event was – from the hardware point of view, at least – all about the new Volta GPU generation, and its expected leadership in AI (Artificial intelligence) and HPC (High-performance computing). As an example, a single 19-inch rack sized moderately dense Volta cabinet, with 200 to 250 of them besides the (still mandatory for now) CPUs, would match or exceed the Linpack FP performance of Taiwan’s freshly ordered first Petaflop-class, multi-rack system from Fujitsu.
The HBM memory approach that Volta inherits from its predecessor, Pascal, could likely for the first time be implemented in Nvidia’s high-end consumer GeForce cards as Volta microarchitecture expands into the mainstream next year.
Why would Nvidia bother with HBM on the mainstream, when the GDDR5X approach seems to provide sufficient bandwidth to win just about any gaming benchmark out there on the Pascal generation? The likely answer is AMD. Already market-tested with Fiji for consumer-market HBM memory integration, the upcoming Vega will double that memory to 8GB, with US$500-class Radeon cards for the gamers out first. While probably behind Volta on raw performance, it will be "good enough" and, with extra bandwidth & integration, make a very neat solution for even a very compact multi-GPU enthusiast PC systems.
As seen with their Fiji-based Fury product, a consumer HBM-based card design was far smaller and more integrated. Nvidia’s dense datacenter MXM modules for their GPUs show the same – in both cases, those who can bring novel cooling approaches to maximize the benefits of that density will have a chance to gain market share.
But what about the board makers, right here in Taiwan? The lesson of AMD Fiji was that, the more integration, the less differentiation. Once the memory was inside the chip packaging, the vendor’s "uniqueness" space for extra capacity, speed, or even board tuning for that memory, was gone. Sooner or later, that will happen with the Nvidia consumer offerings as well.
What can the GPU card brands do then to keep the integration? As seen on the Fiji example, the cooling system provided on AMD reference cards left a lot to desire, and that could be the biggest trump card to play. When the board design becomes so compact, why not spread the space savings into the third dimension, depth, as well, and produce full-speed single slot high end cards? Top grade liquid cooling designs will enable that, even in all-in-one (AIO) designs.
These compact chips also occupy far less per-GPU board real estate, which beckons to the potential revival of old Asustek ROG and Gigabyte (among others) custom ultra high end dual-GPU (even triple-GPU with larger PLX PCIe switches) card designs, without having to wait on the GPU chip vendor to create one.
Why not be the first with dual Vega, before AMD RTG musters its busy resources to get one out? It’s so much easier to do one now due to the integration, again as seen with Fiji x2 last year. Besides innovative cooling, the usual board quality, OC tuning and other capabilities can then show each vendor’s skills on such higher end product for somewhat more margin, like in the good old times.
Finally, go beyond a single board: the very fast and ultra expensive US$69,000 (where does Jensen pick up these lovely pricing numbers?) quad-Volta Tesla DGX box, shown at the GTC, is a good guiding light for future uber-enthusiast consumer and 4K+ VR boxes with a similar approach.
While they obviously may not cost that much, these integrated multi-GPU box-level solutions will allow each vendor to show the most of their unique capabilities in all aspects. This year is a good time to start to be ready for the 2018 market. Who’s going to take the pole position?
Nebojsa Novakovic has been based in Singapore and China over the past 24 years and has over three decades of experience in high end computing, focusing on CPUs, system design and platform evaluation & developments – from high end desktops and workstations to servers and supercomputers. He wears many hats, including media member and analyst, consultant and vendor. He was worked on implementations of about every major Western and Chinese CPU & GPU platform under the sun, including board-level adventures and other projects in that realm for Taiwan vendors.
- Wireless broadband developments in Southeast Asia markets
As of 2013, the 10 ASEAN nations had a total of over 700 million mobile subscriptions, with the CAGR from 2003-2013 reaching 24%. This Digitimes Research Special Report analyzes the various mobile broadband markets in ASEAN and looks at the respective trends in 4G LTE development for those markets.
- 2015 global tablet demand forecast
This Digitimes Research Special Report provides a 2015 forecast for the global tablet market and analyzes the strategies of key market players such as Google, Apple, Intel, and Microsoft.
- 2015 China smartphone panel trend forecast
This Digitimes Research Special Report analyzes the strategies of key China-based major panel makers BOE, Tianma and IVO for attacking the different market segments through technology and pricing, and their relationship to local vendors Huawei, Lenovo, ZTE, Xiaomi and Coolpad.