vacancies advertise contact news tip The Vault
facebook rss twitter

Intel Nehalem CPU running at 3.20GHz, right now

by Tarinder Sandhu on 2 April 2008, 07:41

Tags: Intel (NASDAQ:INTC)

Quick Link: HEXUS.net/qamh2

Add to My Vault: x

Please log in to view Printer Friendly Layout

Intel put a dampener on prospective Spring IDF 2008 revelations by first announcing scant details of its upcoming Nehalem CPU architecture over a year ago, adding liberally at Fall IDF 2008, and then providing some real meat on the specifications' bones two weeks ago.

Combining what we know already, Nehalem will be core-scalable from two to eight execution cores on a single socket that, along with the architecture's simultaneous multi-threading, lead to concurrent four-to-16-thread computing. Initial desktop-oriented models, dubbed Bloomfield, will feature four execution cores with, obviously, eight-thread capability.

Nehalem, like the Core architecture, will provide platform-wide support, from notebooks to servers. Unlike Core (2), however, Nehalem will utilise a building-block approach that can potentially add-in non-CPU cores on to the die - integrated graphics being a prime example.

Thinking of the 1,366-pin, 731M-transistor quad-core model, each core will have access to its own 256KiB of L2 cache, and share a pool of 8MiB L3 cache The CPU's memory controller is integrated, supporting tri-channel DDR3-1333, and the processor will interface with an all-new I/O hub, Tylersburg - and each other in a two-socket environment - via the QuickPath point-to-point Interconnect that replaces the ageing, incumbent Front-Side Bus (FSB).

Tylersburg will support a single processor, via QuickPath, obviously, that'll be aimed at high-end desktop PCs, and a two-socket arrangement primarily aimed at the high-performance computing crowd. Hooking up to the southbridge - ICH10, most likely - via the present DMI bus. Expect to see the first iterations of Nehalem hit the shelves in Q1 2009

The long and short of it is that Nehalem, we reckon, will offer around 30 per cent better performance, on a clock-for-clock basis, when compared to Core 2, in a heavily-multithreaded environment - HPC and low-end servers, mainly. These gains diminish with lighter workloads that are the mainstay of the desktop environment, so don't be surprised to see some not-quite-as-impressive-as-expected numbers coming your way soon.

Naturally, Intel was keen to point out that early silicon, A1, was in good working order. That's why a number of Nehalem-based systems were on show at IDF 2008.




Here's a Bloomfield-based (quad-core) Nehalem system that was up and running a dynamic-airflow simulation.



Now, with the SMT present on the cores, the operating system sees eight processors in device manager. Remember, however, that it's a single-socket system.

Intel isn't talking about launch frequencies at the moment, but we expect to see Bloomfield ship at up-to 3.6GHz in Q4 2008.

The sample was running at 3.20GHz and, as you would expect, interfacing is tri-channel DDR3-1333 memory.

We weren't able to gain any performance numbers, no matter how much misdirection was aimed at the Intel representative.



HEXUS Forums :: 18 Comments

Login with Forum Account

Don't have an account? Register today!
Just can't see it offering much performance. The memory controller is replacing, not adding to, the existing on die logic that reduces latency and most of that prefetching will go as there isn't cache to use it. Unlike the jump from low cache, highish latency Athlon XP's to a low latency ATh 64 with onboard mem controller directly responsible.

High latency P4 to massively reduced latency Core 2 Duo has already happened. Infact, DDR3's higher latency with massive bandwidth is perfect for a latency hidden accurate predicting prefetch large cache setup. Low speed high latency DDR3, or high speed very high latency DDR3 with large bandwidth strikes me as worse for a chip that will lose a lot of its latency hiding ability and rely directy on quick memory access. Especially with triple channel, theres going to be massively more expensive to buy bandwidth than we remotely need. DDR2 would actually work better, lower latency, triple channel, new bus, you'd be talking about a fairly easy 30gb/s anyway and the system will never ever use that. But it would be quite a bit lower latency than a DDR3 setup.

THe only place Nehalem will be gaining in performance will be in the all 4 cores 100% loaded area. But even now the Kentsfield/Yorkfield, despite the lack of bandwidth is actually very good, scaling from 1-3 cores is almost perfect, the 4th core in many applications scales very well and only in a few area's is there a significant drop, and thats where the 30% performance will come in, a very few select apps.

Its a necessary move, but nothing thats going to provide useful benefits for now. Its the stepping stone to on die intergrated gpu, and intergrated PCI-E control and possibly some better sli/crossfire scaling in the future on a whole better faster system architechture. But none of thats now, this is just whats needed to enable that in the future.
I disagree - we never thought we'd talk about low latency at the same time as DDR2 once upon a time, but now you are throwing it about. Why won't the same thing happen with DDR3 eventually?

Prefetching and the other improvements in Core 2 only mask the higher latency of off-die memory controllers, not eliminate it, and they don't do so in every scenario. The onboard controller done right should still enable a further reduction in latency, and have much better behaviour where the Core 2 was previously weak.

Lets not forget that QPI and the onboard controller have another important effect - reduction in complexity of the motherboard. How much that directly affects performance I'm not sure.
looks impressive but will cost an arm and a leg when it first releases and like someone already said, is it needed? todays quad core are already an overkill for many of us
j.o.s.h.1408;1380896
and like someone already said, is it needed?
Hehe the technology world wouldn't exist if we were concerned about that question.

Some people will always want the latest. Others will be upgrading from an older generation and so will see a significant performance increase. I'll be getting a nahalem system I'm sure. Do I need it? Of course not, but it'll still be a noticeable upgrade from my 939 based system.
We must have progress for the sake of progress in the tech world. Even if few people use (or have the need for) Nehalem, at least Intel can learn from it and it will keep AMD on its toes. The same principle applies to the graphics card market. Nvidia is currently playing marketing games, making small advances in technology compared to Core2 -> Nehalem, and slapping new model numbers on them, but even small steps are better than stagnation. Even if a handful of techies buy the 9800GX2, at least it drives the market forward a little.