Looks like RWC is just a minor tweak of GLC. "New core architecture" my butt, and this is coming straight off Intel's accusations of snake oil...
View attachment 90478
Good find! I think I may be correct when I wrote that Intel wanted to downplay any improvements in RWC as that would highlight the latency penalty in moving to tiles. Increasing L1 to offset the tile penalty makes sense. There would have been an IPC (throughput) uplift if Intel had remained monolithic. But that ship has sailed. The future is tiles.
I have been following this release closely, trying not to make knee jerk judgements as it generally takes a while to fully understand the short and long term implications of a new architecture and this one more so than normal because there are so many monumental changes and we haven't really had a lot of "trusted" reviews.
First, we have Intel 4 finally going "live" in silicon. That's great. But how great is it? Intel claims a significant transistor density uplift but since we don't know the number of transistors and die size we can't verify that. In addition, we don't have good numbers on density for a few generations now. So there's one question.
Another is the frequency Intel 4 can hit with this first release of the node. 5GHz for the current top of the stack part is pretty good. But the fact that a 5.1GHz part is coming next year suggests at least for mobile they are currently topped out at 5GHz and are binning like mad for 5.1GHz parts. Raptor mobile parts are in the same range of clocks as Meteor Lake so clocks don't seem like the issue (at least for mobile). This does seem like a good indicator for future Intel 4 releases that the will be able to probably be able to hit somewhere between Alder and Raptor (1st gen) frequency for ARL.
Speaking of ARL, since they have given up some throughput moving to tiles and already made a tick with RWC Intel is going to have to make up a bit more throughput for the tiles to equal RPC IPC, then make up another few percent for the clock speed regression from Intel 4 to Intel 7 at it's latest incarnation, and this is just to equal Raptor Cove performance much less improve on it. Lion Cove has a lot on it's shoulders.
As far as power consumption, initial testing shows it's better than Raptor. Much better at higher mobile Watts (like 40 or 50) but not so much better where it really matters for mobile in the 15 to 30 watt range. But as I wrote above it's early and we need more testing. Better than the last Intel offering and perhaps equal to AMD. That can be argued as a win or a loss. Like the hybrid approach it looks like there are use cases to work out. Seems like the GPU and CPU tiles need to stay sleeping during web browsing, watching video, and general light office work otherwise it seems as though firing up those other tiles incurs a big electrical penalty. We have seen 21 hour battery life when only the SoC is active, but how realistic is that. As is usually the case AMD and Intel are very competitive in the mobile space and Intel has the advantage of being available in a lot more systems.
The GPU performance again catches up to AMD. I think this is a win for Intel. No, it's not groundbreaking but it's significant and since many applications are using the GPU, such as the video my ARC 750 is encoding on Magix Vegas Pro right now, this is important.
I think we can safely say that with Meteor Lake Intel caught up to AMD in the mobile space. They got a lot right here. Intel 4 is out the door and hitting 5GHz. That's big. The tiled architecture is working. That's huge.
But is this the end of the beginning of Intel's resurgence or it is the beginning of the end for them? Meaning are they going to forever be playing catch up and slowly but surely losing market share or is their new tiled approach going to have "legs?"
One positive is that since they are in the discrete GPU arena they can trickle down that tech into the iGPU's with great economic and intellectual efficiency since both platforms share hardware and software, there should not be too much duplication of effort.
As I wrote above a lot is riding on Lion Cove. It's going to need to show 20% uplift in IPC to show perhaps 10% uplift over Raptor Refresh when you consider the tiled latency and clock regression. The problem when you run a node to extreme clocks at the expense of power is you set a high bar for the next release. Not only is Intel having to deal with the clocks of Raptor but the inherently lower latency of its monolithic design as compared to future products.
It's fascinating to stand back and watch this play out. I am personally awed by what both AMD and Intel have accomplished over the past 4 or 5 years in particular. Intel with their hybrid approach and insane clocks and AMD being able to produce parts with such enormous compute and efficiency. It's easy to sit here and armchair quarterback this and that and another thing entirely to design, fabricate, test, and build these things into working microprocessors!
Facebook
Twitter