What the hell is… SiS' HyperStreaming technology?

We asked the fellow who knows

SiS introduced its AMD-oriented SiS758 chipset last week and debuted its HyperStreaming performance-enhancement technology at the same time.

HyperStreaming will be incorporated into all future SiS chipsets, Nelson Lee, SiS' product chief, told us. The 748 is due to ship to mobo makers at the end of April. A Pentium 4 solution, the SiS640fs, is scheduled to go into production during May, he added.

But what exactly is HyperStreaming? A technobabble-filled press release failed to clarify just what the technology does and how. Fortunately, Mr Lee was on hand to spell it out to us in words of one syllable. We now understand it - we think...

Essentially, HS is a chipset architecture designed to enhance the handling of bulk data, whether that's information streamed from the Internet or between storage devices. It targets single streams of data and multiple, parallel streams through a mix of latency reduction, pipelining, real-time access, prioritisation and channel arbitration.

For single streams of data, such as Internet information transmission (or any network traffic, presumably) and the direct accessing of a single I/O device, e.g. a graphics card, or IDE drive, HS reduces the lag between the request for data and the arrival of that information.

Accessing a graphics card, for example, is between five per cent and 43 per cent better than unnamed - (sort of: they're "V Corp. and N Corp - guess who...) rival products, SiS claims. For CPU to IDE, the gain is between 13 per cent and 75 per cent. HS accelerates CPU to memory operations by 110 per cent to 162.5 per cent over rival solutions, says SiS.

The company tested HS using a Athlon XP 2400+ system with 256MB of 333MHz DDR, and a custom-made real-time operating system to cancel any latencies imposed by Windows.

Multiple streams of data are characteristic of applications which execute multiple threads and processes in parallel. That stretches from server apps right down to the operating system.

HS provides a parallel architecture in the North Bridge, the bus to the South Bridge and within device controllers embedded in the South Bridge. It also implements instruction pipelining to overlap commands sent between the CPU and I/O devices. This is possible because devices and the streams of data between them and the CPU are largely independent. Essentially, rather than process each stream sequentially, HS mixes these streams up to achieve greater parallelism and thus higher performance.

So, each CPU-to-device read operation is split into data request and device response phases. The upshot is that the bus can be relinquished once the request is sent and while the response is awaited. The chipset buffers all the transactions that have sent requests and are currently awaiting a response.

HS also pipelines groups of requests, some of them out of order, to increase the number of requests sent in a given time. Intelligence in the chipset stitches responses to requests in the correct order. HS also ties to exploit parallelism within streams from the same device.

The upshot, says SiS, is 1.2GBps of system bandwidth.

HS provides certain streams with priority access through their own channels. Priority accesses are tagged to ensure the components of the HS architecture speed them on their way as quickly as possible.

That, reckons SiS, makes a major improvement to playing back data streams such as audio and video, and downloaded data. For example, SiS claims a three to 8.5 per cent performance gain in FTP operations over chipsets from rivals.

Finally, HS implements what SiS calls smart streaming to help it interface with architectures based on different protocols, such as HyperTransport, AGP, the memory bus and the frontside bus. Requests aimed at the same destination - such as the same bank of memory - are combined and re-ordered, while requests that will delay the issuing of others that depend upon the responses they will get can be prioritised.

Each of these components combine to equip HS-based PCs with a better architecture for today's networked environments - whether that's the office, the server farm, or simply a home PC permanently connected to the Internet via a broadband connection.

Once the SiS748 ships next month, it will be interesting to see if independent tests show the same performance benefits SiS claims to have achieved. ®

Related Link

For more details on how HyperStreaming works, take a look at SiS' HyperStreaming white paper (PDF)