Intel Coffee Lake Coming October 5, Here Are The Details

Page 3 - Seeking answers? Join the Tom's Hardware community: where nearly two million members share solutions and discuss the latest tech.
Status
Not open for further replies.

MichaelElfial

Commendable
Oct 22, 2016
9
0
1,510
MUFFINDELL, I wonder too - is it just lack of information or the memory bandwidth is limiting the number of cores. The cache is more, but not that much. I am going to upgrade and can't choose between i5 and i7 or waiting for the next gen (assuming it will really bring 8 cores - Intel will have to do something for the memory interface there if it is a problem). Watching Ryzens it is quite obvious that beyond 6 cores the efficiency drops drastically, so I wonder where exactly is this "golden" border in the Intel's case.
 

TJ Hooker

Titan
Ambassador

Even the most well-threaded of applications won't scale perfectly with core count, and games generally aren't particularly well threaded (the 25% number is from a comparison of gaming performance in Gears of War).

L3 cache is 50% higher as well, so cache per core is the same as it was for Kaby Lake.
 

InvalidError

Titan
Moderator

HPC software running on supercomputers with over 100k cores come pretty darned close - you have to waste less than 0.001% of compute time on thread synchronization and communications to achieve any scaling gains at that scale. Software optimized to a comparable extent for single-socket systems for tasks that lend themselves similarly well to multi-threading would achieve practically perfect scaling.

Of course, this sort of computing task is highly uncommon on desktops and since the scope of similar problems that can run on a typical PC is so much smaller, optimizing every last bit of it isn't as critical either..
 

InvalidError

Titan
Moderator

There are no official JEDEC specs for DDR4 speeds faster than that, which means nobody can officially claim support for anything faster. Same thing with AMD which officially only supports DDR4-2666 when using single-sided DIMMs with only one DIMM per channel on Ryzen.
 

faizanali1997

Reputable
Sep 25, 2017
3
0
4,510


Yea I'm also keeping an eye on the next APU releases by AMD (Raven Ridge I think), I heard they'll be using Zen so there's probably a significant jump in performance on the CPU side. I don't know much about AMD CPUs, I've only recently been following them, since Ryzen released. I don't know for how many generations (of CPU) the AM4 will hold me off, but this recent decision by Intel hasn't helped me take their side in the slightest. I'm looking for anything that'll hold me off till I get a proper build. Looking for a budget CPU with the best value. For now, I have my eyes on the Coffee-Lake Pentiums and Raven Ridge. If the AM4 lasts long, I might not even consider CL Pentiums.

@MichaelElfial This is definitely why I'm more interested in them. I'm basically buying an i3 (2C4T) for the price of a Pentium, except for some CPU feautures like AVX2 and the slight performance difference in the iGPU. Interesting how the CL revisions will turn out.
 

InvalidError

Titan
Moderator

It only kills the i5-7640X and i7-7740X which never made sense to anyone but Intel anyway. It doesn't affect X299's relevance to anyone who wants more than 16 PCIe lanes, more RAM or memory bandwidth than can be crammed on a dual-channel platform or need 8+ CPU cores.
 

samer.forums

Notable
BANNED
Sep 30, 2017
662
0
1,160


I know but they should have moved to four channels memory for the 8th generation already , it is about time ... I am 100% positive had AMD released the Ryzen with four channels intel would have followed ... but anyways its INTEL .
 

InvalidError

Titan
Moderator

Having four channels means needing close to twice as many PCB layers on the motherboard and motherboards costing $50-100 extra. Most people who buy $50-200 CPUs don't want to pay $200+ for a motherboard with quad-channel memory which they don't need, same goes for the extra premium on quad-channel memory kits. It makes no sense to significantly increase the budget-oriented platform's cost until the extra bandwidth becomes absolutely necessary.

If you are one of the few who absolutely need quad-channel memory on an eight-core CPU, get an i7-7800X or ThreadRipper 1920. The rest of the market is much happier with 4-8 cores platforms that cost half as much while giving up little to no performance in most mainstream and prosumer workloads.
 

samer.forums

Notable
BANNED
Sep 30, 2017
662
0
1,160


we have been using dual channel for consumer grade since the time of Pentium 4 .. it is time already to move up. we moved once from single channel to dual. and it is time to move up already. more over , the xeons now are six channels , maybe they should release B299 chipset to make things more affordable , that is only 16 lanes motherboard instead of connecting 44 lanes to the slots as in the expensive one.

Actually the X299 chipset does not offer anything more than the Z chipset at all. I have the feeling it is the same chipset with different name .. not even ECC is added to the X299 which is stupid for people who need more cores for engineering . People who want an i9 with 10+ cores will look at the xeons for the ECC support. there is no way a gamer wastes his money on more cores.
 

InvalidError

Titan
Moderator

There is no point in "moving up" until mainstream CPUs actually hit a substantial memory bandwidth bottleneck that meaningfully impacts more than SiSoft memory benchmarks. Also, since the P4 which used DDR-266 memory, we're at DDR4-2666 as the highest current standard, that's a 10X increase without significantly impacting motherboard design and manufacturing cost.

By the time we have mainstream CPUs that genuinely require significantly faster memory, not to be confused with the need of some CPUs to crank memory clocks arbitrarily high to overcome architecture quirks as is the case with Ryzen, we'll have DDR5 and dual-channel will be good for another couple of years assuming on-package HBM/HMC doesn't obviate the need for faster off-chip memory.
 

samer.forums

Notable
BANNED
Sep 30, 2017
662
0
1,160


How about the CPU internal Graphics performance ? they will take advantage of the more bandwidth because they share System memory ?

Also Dedicated GPU using system memory when it is needed ?
 

TJ Hooker

Titan
Ambassador

The only reason you're using integrated graphics is if you have low GPU performance requirements and/or you're on a tight budget. If it's the former then who cares, and if it's the latter then you won't be buying a CPU/mobo with quad channel memory. For the price premium you could probably buy a discrete GPU that would perform better anyway.

Even with quad channel memory, max total system memory bandwidth would be a fraction of the VRAM bandwidth available on even mid-tier graphics cards, never mind the added latency. Running out of VRAM will always result in a performance hit and should therefore be avoided if at all possible, and quad channel memory wouldn't change that.
 

InvalidError

Titan
Moderator

As Hooker already wrote, if you are so cost-sensitive that you are going to forgo a discrete GPU, then you don't have the budget to afford a quad-channel CPU and motherboard. On the discrete GPU side, PCIe 3.0 x16 is only 16GB/s vs 42GB/s for dual-channel DDR4-2666 and 100-1000GB/s for the GPU's VRAM. System memory is utterly uselsss to even entry-level discrete GPUs due to the PCIe bottleneck and even PCIe 4.0 will hardly make a dent in that.
 

samer.forums

Notable
BANNED
Sep 30, 2017
662
0
1,160


This is assuming it will be more expensive when intel chooses to upgrade and leave the dual channel behind forever.

They can release 16 lanes only x299 motherboards priced cheaper and aimed at consumer level CPU ... that is only 16 lanes are connected from the socket and not 44 lanes which will reduce the Motherboard Prices .
 

InvalidError

Titan
Moderator

Removing PCIe slots to support fewer lanes wouldn't reduce motherboard prices by a meaningful amount since the extra PCB layers would still be required to route the 400+ densely packed and tightly timed signals required for quad-channel memory. Since PCIe lacks the strict bus-wide timing requirements that DDR4 has thanks to each PCIe lane operating independently and getting re-aligned with the other lanes on a word-wise basis after the deserialization process, laying out the 176 traces for 44 PCIe lane is much easier than doing the same for the ~100 traces of a DDR4 interface.
 
Status
Not open for further replies.