John Brady's Weblog

So Sun has finally announced servers based on the Niagara multi-threaded, multi-core CPU. Officially it is the
UltraSPARC T1 processor,
in the
T1000 and
T2000 servers.

At one level I think the whole Niagara/T1 chip multi-threaded (CMT) stuff that Sun has done is great, if not phenomenal. They have managed to design and build a processor that delivers more throughput in total, while not consuming any more resources to achieve it.
And they have done it in a way that deals with the memory latency that slows down the performance of normal single threaded processors.
The T1 should definitely deliver more overall performance per core than a single threaded
processor, and with 8 simpler cores packed together on one physical chip sharing related
infrastructure, it will deliver phenomenal amounts of processing power at very low cost.
And Sun should sell a lot of T1 based systems as a result.

But that leads to my other reaction - that Niagara/T1 could spell a major decline in Sun's revenues.
Why?
Because it is so cheap and yet has the equivalent power of
an E10000 system with 32 UltraSPARC-II processors in it.
So overnight a lot (but not all) of Sun's medium to high end server revenue,
based on its UltraSPARC-III and IV processors will disappear,
and will be replaced by sales of the T1000 and T2000.

Think about it. An 8 core T1 can have 32 executing threads.
Each core is equivalent to about an UltraSPARC-III (US-III) in terms of its technology and speed.
Even if a T1 does not achieve 32 times a single US-III CPU, it should achieve at least 16 times in terms of real throughput.
So EVERY 16 CPU US-III system out there could be replaced by a T1 based system!
That would have been a Sun Fire 6800 system when the US-III was first launched, costing several hundred thousand dollars.
And now it can be replaced by something costing between ten and thirty thousand dollars.
That is a major shift in the scales of revenues!

Furthermore, 16 US-III is equivalent to 8 US-IV dual core processors.
So every system from the V890 downwards can each be replaced by a single T1 based system instead. And most E2900 system sales too, as this can only hold up to 12 US-IV processors.
So, all of the revenue streams from all of the V890, V490, V440, V240 and V210 systems will just dry up. And very quickly.

The prices for these existing systems range from $118,995 for a 8 US-IV CPU, 32 GB V890, through
$80,995 for a 4 US-IV CPU, 32 GB V490,
and $40,995 for a 4 US-IIIi CPU, 32 GB V440,
to below $10,000 for the V240 and V210.

And the T1000 costs $11,995 for an 8 core, 16 GB configuration, and the T2000 costs
$26,995 for an 8 core, 32 GB configuration.

Do the math. Systems currently selling for between $50,000 and $100,000 will all be replaced by systems selling for between $12,000 and $27,000. The revenue Sun gets for this class of system will drop to roughly a quarter of its current revenue.

This means that at the high end, in practical terms, Sun will be left with the E6900, E20K and E25K only. And suddenly, compared to the T1 based systems, they look VERY expensive for the amount of processing power they deliver.

And if you are using application software that can be partitioned over a cluster, such as Oracle RAC, then you could replace an E6900 or E20K with a cluster of 4 * T2000 servers.
And Sun is saying that
Oracle will recognise a T1 processor as 2 CPUs for licensing its software. [Scroll down to the paragraph starting "Continuing to build on their longstanding collaboration"]

Let's say that 16 US-IV CPUs today equals 32 or more of the original US-III CPUs.
And one 8 core T1 processor = 32 threads = 16 US-IIIs at worst performance.
Then 4 * T1 = 64 US-IIIs presuming linear scaling when clustered.
But, given that Oracle RAC will not scale linearily, we might only get 3 times
the throughput of one T2000 system.
The net result is that the cluster of 4 * T2000 servers gives me around 48 US-IIIs worth of processing power, which is more than a server with 16 US-IV CPUs in it.

So a 16 US-IV CPU 64 GB E20K at
$905,142 could be replaced by 4 * T2000 of 8 core, 32 Gb each at 4 *
$26,995 = $107,980.
This represents a saving of $797,162, or 88% of the price of the E20K!

I imagine that Sun's figures for Q1 of calendar year 2006 will not look very good.
Great server volumes, but lousy revenues, and therefore lousy profit (if any).
Again, do the maths yourself. As a consumer, the T1 based T1000 and T2000 are great products.
Tremendous amounts of processing power for very little money,
and fully binary compatible for running Solaris 10 and all of the ISV applications already out there.
As a shareholder, just how does Sun expect to grow its revenue, and as a result
cover its costs and so make a profit?
I don't know. And I don't believe Scott or Jonathan know either.
Otherwise they would have fixed the revenue / profit problem a long time ago,
and Sun wouldn't be in the mess it is in now.

# posted by John Brady @ 8:30 pm