updated 18 October 2017
-------------Recommended low profile card: ZOTAC GeForce GTX 1050 Ti ZT-P10510E-10L 4GB 128-Bit GDDR5 (Newegg link) ($165)
)Second option: MSI GeForce GTX 1050 2GT LP 2GB (Newegg link) ($120)Cheap option: MSI Radeon RX 550 2GT LP OC 2GB (Newegg link) ($85)Recommended low profile, single slot card: MSI GeForce GT 1030 DirectX 12 GT 1030 2G LP OC 2GB (Newegg link) ($72)
------Low profile, double slot:
The GeForce GTX 1050ti is the obvious choice here, and I've listed the regular 1050 version as a second option. You may want to look over some reviews to decide if you think the extra $45 is worth it to jump to the 1050ti. Personally, I would do it, but that's because I'd expect a card in this form factor to be in use for a long time, and I'd want the extra 2GB memory for longevity, and the extra performance is an added bonus. Gigabyte and MSI also have versions of both 1050 and 1050ti in low profile, double slot, and I've typically linked the MSI cards, but those have been showing out of stock for a while so I've got the Zotac one linked here.
I've added a third option, the cheapie. I wouldn't have bothered with this but MSI also has a low profile, double slot RX 550. It isn't in the same performance class as the GTX 1050 but it's also $85, which is nearly $50 less than the 1050. Tom's has a good writeup
comparing it to an RX 460, R7 260X and Intel HD 530; if you look at an RX 460 vs. GTX 1050 review you'll have an idea where it falls for performance. But if the goal here is getting a better setup than integrated graphics, and you're on a really tight budget, $85 isn't bad at all.Low profile, single slot:
NVIDIA's GeForce GT 1030 finally brings a new architecture (Pascal) to the low end, which previously was filled with various Fermi and Kepler parts, dating back to tech launched in 2010 and 2012, respectively. I haven't seen any comparative reviews yet but I'd expect it to be roughly twice as fast as the old GT 730 and to probably trade back and forth with the RX 550. The really nice thing about it is the price, though: $70. Sure it's sitting on a 64-bit GDDR5 memory setup but that will be fast enough to feed the tiny GPU.
I've listed the MSI card here for two reasons: 1) it seems to be the cheapest and 2) it has both a DisplayPort and HDMI output, instead of wasting space on an old DVI port. EVGA's got the highest clocked 1030 if you're interested, and some Newegg user reviews seem to indicate it overclocks really well (which you might not think is all that useful, but if you look at some of the old data I've got below with doing an iGPU overclock on Haswell, you might be surprised how much of a difference it can make on lower end parts).
There are currently no good alternatives in this form factor.Low profile, single slot, crazy expensive workstation stuff:
Before I get into these, let me stress that I really don't recommend buying a workstation card for gaming. They are far more expensive than the performance you will get out of them. That said, some have asked about it, so here is a very brief overview:Super expensive AMD Radeon Pro workstation cards (low profile, single slot):
Radeon Pro WX 4100: an RX 560 with 4GB memory, a 50MHz core clock decrease and a 250MHz decrease on the memory
Radeon Pro WX 3100: an RX 550 with 4GB memory, a 175MHz core clock decrease and 250MHz mem clock drop
Radeon Pro WX 2100: an RX 550 with 2GB memory, 175MHz core clock decrease, memory speed unchanged but bus cut in half to 64 bitSuper expensive NVIDIA Quadro workstation cards (low profile, single slot):
Quadro P1000: a GTX 1050 with 4GB RAM, 100MHz core clock decrease and 500MHz mem clock drop
Quadro P600: no real consumer equivalent; a GT 1030 with 2GB RAM, a 120MHz core increase, 500MHz mem decrease (but double bus width to 128 bit, so should be a net gain) and an extra 8 ROPs, so I guess if there were ever a "GT 1040" it would be similar
Quadro P400: no real consumer equivalent; a GT 1030 with 2GB RAM, core clock speed unchanged, 500MHz memory drop, and a 256/16/16 shader/TMU/ROP config vs. the GT 1030's 384/24/8, I guess if they released a "GT 1020" it would look something like this
this is out of date by now, originally written in 2014, but I think the general principle, which is "look at some benchmarks because your IGP may be more capable than you think" or "look at some benchmarks because that video card might not be much of an upgrade after all", still applies
The discussion thread about obsolete GPUs
got me wondering about the kind of performance improvement you'd really get out of these cards, especially the single slot ones, compared to Intel's HD 4600 (Haswell), 4000 (Ivy Bridge) and 2000 (Sandy Bridge) graphics.
To try and get an idea of this, I browsed through Futuremark's database
, looking for "Cloud Gate" scores. I focused on these processors:
Ivy Bridge: i3-3225
Sandy Bridge: i3-2100.
Since it's easy enough to spec an ITX system with 8GB of system RAM, I've looked for results from systems with that amount, and a few different video card options for each proc. I've "normalized" the results to the IGP score for each proc.
I chose i3 processors for the results comparison for two reasons: First, I have an i3 in my system, and second, I figure smaller systems with these space/power constraints will not likely be crammed full of unlocked Core i7s.
Cloud Gate is not a graphics-only test; it also does a physics test and the overall score is averaged somehow from all of the results. I've listed all three scores below, but I think the "overall" score is going to be most indicative of the kind of performance increase you can expect from installing one of these cards.i3-4360 results (HD 4600 graphics):
Quick notes: systems specced with 8GB of RAM. Memory type was determined by reported clock speed and known video card configurations.
- HD 4600, DDR3-1600: 5973
(1.00) - Graphics 7132 - Physics 3809
- GT 730, 2GB DDR3: 5522
(0.92) - Graphics 6278 - Physics 3886
- R7 250, 2GB DDR3: 6925
(1.16) - Graphics 8953 - Physics 3864
- GTX 750, 1GB GDDR5: 11387
(1.91) - Graphics 28650 - Physics 3663
- GTX 750Ti, 2GB GDDR5: 12739
(2.07) - Graphics 34469 - Physics 3973
Well. Some interesting results in their database here, and it confirms my suspicions (and those of many others) that a card with DDR3 is absolutely not worth the cost of entry at this point. The GT 730 with DDR3 is actually slower
than the HD 4600 on the i3-4360. The R7 250 with DDR3 is a marginal improvement over the HD 4600 - but given the prices of these cards, you'd be paying near $100 for a 15% improvement in performance. Figure that in actual games, not synthetics, this difference might be even smaller, and I can easily say it's just not worth the cost.
The 750 and 750ti are really the best option if your case has room for a double-slot low profile card. If it doesn't, well, I couldn't find numbers for a GDDR5 730 or 250 with the i3-4360, but I did find some on the Ivy Bridge proc. Read on!i3-3225 results (HD 4000 graphics):
Quick notes: systems specced with 8GB of RAM. Memory type was determined by reported clock speed and known video card configurations. R7-250 results come from a system with an i3-3240 (100MHz faster) so may be slightly askew. GT 740 results from a system using i3-3220; same clock speed as 3225 but has onboard HD 2500 instead of HD 4000.
- Baseline HD 4000, DDR3-1600: 4533
(1.00) - Graphics 5133 - Physics 3218
- GT 730, 1GB GDDR5: 7578
(1.67) - Graphics 11949 - Physics 3324
- GT 740, 2GB of GDDR5: 8447
(1.86) - Graphics 15117 - Physics 3320
- R7 250, 1GB of GDDR5: 8824
(1.95) - Graphics 16021 - Physics 3431
- GTX 750, 2GB GDDR5: 10676
(2.36) - Graphics 28903 - Physics 3329
- GTX 750ti, 2GB GDDR5: 11407
(2.52) - Graphics 35552 - Physics 3378
I was able to find results for a GT 730 with GDDR5 and it paints a pretty compelling picture: even though those cards have a 64-bit memory interface, it manages to punch a full 66% faster, overall, than the HD 4000 graphics. A 730 might be worthwhile for an Ivy Bridge system, but not a big bump for a Haswell proc with HD 4600.
Also, you'll notice the GT 740 is marginally faster than the GT 730 when both are using GDDR5, but the R7 250 still has a slight advantage - seems the 64-bit GDDR5 interface is a pretty good match for the 384SP config in the 730, and the 740, with the same amount of SPs, isn't really able to make use of the extra bandwidth from the 128-bit config. Looks like the "sweet spot" of bandwidth will be somewhere in the 60GB/s range.i3-2100 results (HD 2000 graphics):
For the curious, before I jump into the numbers from the i3-2100, I checked the DB to see if there were comparable results for the HD 2000 vs HD 3000. There were, although they weren't "validated". The i3-2120 and i3-2125 are same clock speeds, but one has HD 2000 graphics and the other has HD 3000. I don't think the 3000 was as common in the low-end processors for Sandy Bridge, but honestly can't recall - I skipped from Clarkdale to Haswell on the desktop, with a big gap of "not paying attention to this stuff" in between.
- i3-2120, DDR3-1333, HD 2000: 2164
- Graphics 1992 - Physics 3107
- i3-2125, DDR3-1333, HD 3000: 3155
- Graphics 3150 - Physics 3173
Quick notes: the R7 250 results were found paired with an i3-2105 CPU, which has the same clock speed as the 2100 but HD 3000 graphics onboard. I figure it is comparable. The 750ti results came from a system with only 4GB of system RAM. Memory type was determined by reported clock speed and known video card configurations.
- HD 2000 with DDR3-1333: 2082
(1.00) - Graphics 1929 - Physics 2889
- GT 730, 2GB of GDDR3: 4729
(2.27) - Graphics 5733 - Physics 2933
- GT 730, 2GB of GDDR5: 7048
(3.39) - Graphics 11685 - Physics 2951
- R7 250, 1GB of GDDR5: 8013
(3.85) - Graphics 15859 - Physics 2934
- GTX 750, 1GB of GDDR5: 10179
(4.89) - Graphics 31616 - Physics 3018
- GTX 750 Ti, 2GB of GDDR5: 10724
(5.15) - Graphics 36271 - Physics 3095overall
So the R7 250 is the definite winner for single-slot low-profile. If you're already on Haswell HD 4600 graphics, however, The roughly $100 one costs would be getting you about 50% more GPU performance. That's a tough call, but in that form factor, it's the best option. If these cards were priced a little better it'd be an easier choice, but even regular sized R7 250s are $80-$90, so we're not really looking at much of a premium for low profile format. The pricing makes less sense when compared to other equivalently priced cards. AMD needs to get these into the sub-$75 range and make them the new "bottom line" - much the same way NVIDIA has done with the GT 730 cards sitting in the $60-$70 range. Maybe they'll do that with the Rx 300 series.
The R7 250 looks like a much nicer proposition if you're using Ivy Bridge HD4000 (and obviously would be an even bigger upgrade for Sandy Bridge HD2000/3000, as well).
If you can do double-width low profile, though, the 750 or 750ti are the definite winners. The 750 isn't much slower than the 750ti, so unless you can get one for an especially good price, it seems the 750ti makes the most sense.
-----Some results from my own testing:
As kuririkura pointed out
, the higher-end processors can push their onboard graphics a little harder than these i3s. I've done a bit of testing on one of my own systems about this, out of curiousiry, including some memory tweaking and overclocking results.
My two system configs:
i3-4370 (3.8GHz), 8GB DDR3-1600, HD 4600 graphics & Radeon HD 7570 1GB GDDR5 & Radeon R7 250E (HD 7750) 2GB GDDR5
i5-4670K (3.4GHz, 3.8GHz turbo), 8GB DDR3-1600, HD 4600 graphics
Conveniently the i5's max turbo clock matches the i3-4370's clock.
For comparisons, I got my hands on an OEM low profile, single slot Radeon HD 7570 with GDDR5. It's an earlier Terascale 2 (VLIW5) chip, near as I can tell a rebranded Radeon HD 6670, rated at 65W.
I also got ahold of an R7 250E - Powercolor's 4-displayport R7 250 (see the list of cards below for more detail). These are rebranded HD 7750s, GCN 1.0 parts, and they're rated at 55W.
i3-4370 / 8GB @ DDR3-1600 / HD 4600: 6168
(1.00) - graphics 7399 - physics 3899
i3-4370 / 8GB / Radeon HD 7570 1GB GDDR5: 7011
(1.14) - graphics 8951 - physics 3987
i3-4370 / 8GB / Radeon R7 250E 2GB GDDR5: 9836
(1.59) - graphics 17024 - physics 3970
So the 7570 is faster, but not by much. That puts HD 4600 graphics a little slower than a mid-range part from April of 2011... that sounds about right to me. And the HD 7750 / R7 250E muscles ahead of the numbers I've got listed above for other R7 250 cards, by a pretty decent amount.
Moving on, let's see the i3 vs the i5, to explore whether or not there's big gains to be had from having four physical processors:
i3-4370 / 8GB @ DDR3-1600 / HD 4600: 6168
(1.00) - graphics 7399 (1.00) - physics 3899 (1.00)
i5-4670K / 8GB @ DDR3-1600 / HD 4600: 7077
(1.15) - graphics 7968 (1.08) - physics 5087 (1.30)
So the i5 is a little faster than the i3, but not by much - seems a good portion of that increase comes from the higher physics score. I also have a system with an i7-4790k in it, and if I get a chance to connect another drive to that box for a quick Windows install and benchmark run I'll include those numbers here, but I expect what you'll see is largely the same as the i3 vs the i5: higher physics score, slightly higher graphics score, overall HD 4600 is still bumping into a ceiling due to available memory bandwidth and GPU clock.
I think the takeaway here is that a beefier processor may well make the IGP a little faster, but not by a massive amount.
But hey, if you've got an unlocked processor and can do a little overclocking, why not try that and see what it does for performance? I was running the i5-4670k on an ASRock Z87E-ITX which has a few "one-click" type overclocking settings. I've normalized these results to the DDR3-1600 one. I changed memory speeds but couldn't get my DIMMs to go past 1866MHz, also tweaked the max turbo (default 3.8GHz) and max GPU speed (default 1.2GHz). I didn't push these too far - small systems aren't likely to be used for overclocking, and OEM boxes in this form factor with i5s/i7s may not even have an overclock-capable motherboard, so this was really just for fun:
i5-4670K / 8GB @ DDR3-1333 / HD 4600: 6817
(0.96) - graphics 7670 - physics 4907
i5-4670K / 8GB @ DDR3-1600 / HD 4600: 7077
(1.00) - graphics 7968 - physics 5087
i5-4670K / 8GB @ DDR3-1866 / HD 4600: 7403
(1.05) - graphics 8435 - physics 5184
i5-4670K @ 4.0GHz Turbo / 8GB @ DDR3-1600 / HD 4600 @ 1500MHz max: 8244
(1.03) - graphics 9488 - physics 5651
i5-4670K @ 4.0GHz Turbo / 8GB @ DDR3-1866 / HD 4600 @ 1500MHz max: 8564
(1.07) - graphics 9986 - physics 5717
That 25% increase in GPU clock speed, paired with the extra 266MHz on the RAM, made for a 25% increase in the graphics score, although the overall score didn't jump by much - just 7%. That might make a difference in gaming performance, but I don't know that it would be enough to keep me from wanting to install a discrete card. It might be enough to hold me off until all the new cards have launched, though!